Strategy & Tactics

From PRIMO.ai
Jump to navigation Jump to search

YouTube search... ... Quora search ...Google search ...Google News ...Bing News


An AI strategy is a set of choices about where artificial intelligence should change how an organization makes decisions, serves customers and gets work done, and about what it will deliberately not do. AI tactics are the methods that turn those choices into working systems: data pipelines, models, experiments, workflow changes, governance routines and the habits that keep them improving. Strategy answers where and why. Tactics answer how, in what order and how we will know it worked.

The problem this addresses is now well documented. Organizations buy tools and launch pilots, but value at the level of the whole business stays elusive. In McKinsey's 2026 global survey, nearly nine in ten respondents said their organizations regularly use AI in at least one business function, yet only 37% attributed any EBIT impact to it, and only about 6% qualified as "high performers" that credit AI with 5% or more of EBIT. Closing the gap between using AI and benefiting from it is the job of strategy and tactics.

The ideas on this page come from several traditions:

  • Classic strategic planning: environmental scanning, trend analysis, scenario planning and portfolio management.
  • The economics of AI: Ajay Agrawal, Joshua Gans and Avi Goldfarb (University of Toronto) describe AI as a sharp drop in the cost of prediction, which raises the value of its complements: data, judgment and the ability to act.
  • General-purpose technology research: Erik Brynjolfsson (Stanford) and colleagues show why technologies like AI need large complementary investments before productivity rises (the "productivity J-curve").
  • Practitioner playbooks: Andrew Ng's 2018 AI Transformation Playbook, and adoption frameworks from Microsoft and AWS.
  • Field experiments: since 2023, randomized studies at companies such as Boston Consulting Group and Procter & Gamble have measured what generative AI actually does to productivity and quality.

The central mechanism is simple to state and hard to execute. AI makes some predictions, drafts and actions much cheaper. Value appears only when an organization redesigns a decision or workflow around that cheaper input, invests in what the change needs (data, skills, process changes, controls), and measures the result honestly. That is why the organizations reporting the most value redesign workflows and set growth goals, rather than layering AI onto unchanged processes.

How this differs from earlier "big data" and machine learning strategy:

  • From projects to workflows. A churn model was a project. Generative AI and agents touch almost every knowledge task, so strategy is about redesigning work, not adding one model.
  • From building models to choosing and evaluating them. Most organizations now start from pretrained foundation models and spend their effort on data, integration, evaluation and change management.
  • From predicting to acting. Agentic systems call tools and take actions, which raises both the upside and the cost of mistakes.
  • From annual plans to fast cycles. Model capability and price change every few months, so plans need regular review points.


Strategy decides where AI should change decisions and work, and which risks are acceptable; tactics turn those choices into data, models, experiments and routines that prove the value. Weighing opportunity and risk at every step raises the odds of the best outcomes.


The Mechanism

How an AI strategy becomes measurable results:

  1. Sense. Scan signals and trends inside and outside the organization: model capability and price changes, competitor moves, customer behavior, regulation, and your own usage and cost data. (Detailed in Strategy below.)
  2. Frame. Turn the main uncertainties into a few plausible scenarios, and set goals that hold up across them.
  3. Choose. Build a portfolio of use cases scored on value, feasibility, data readiness and risk. Decide whether to buy, build or partner for each one.
  4. Build foundations. Data architecture, platforms, security, evaluation tooling, governance and skills. Many use cases share these, which is why they belong to strategy rather than to any single project.
  5. Execute. Run pilots with success and stop criteria set in advance, then move the winners into production through a disciplined lifecycle (see Tactics).
  6. Redesign the work. Change roles, steps, handoffs and incentives around the new capability. Without this step, gains stay at the level of individual tasks.
  7. Measure and reallocate. Track business outcomes with experiments and monitoring, stop what does not work, and move people and budget to what does. Then return to step 1.
Stage Key question Typical outputs Common failure
Sense What is changing that matters to us? Signal log, trend briefs, usage and cost telemetry Chasing headlines; missing weak signals
Frame Which futures should we be ready for? Two to four scenarios, goals, guardrails Betting everything on one scenario
Choose Where will AI change a decision or workflow enough to matter? Scored use-case portfolio; build, buy or partner decisions Picking what demos well instead of what is valuable
Build foundations What do many use cases need in common? Data products, platform, evaluation harness, policies Every team rebuilding the same plumbing
Execute Does it work on real data with real users? Pilot results against preset criteria; production system "Pilot purgatory": no path to production
Redesign the work Who does what differently now? New process maps, roles, training Bolting AI onto an unchanged process
Measure and reallocate Did the business result change? Experiment readouts, dashboards, portfolio decisions Counting usage instead of outcomes

Major components.

  • People: an accountable executive owner, product owners for each workflow, data and ML engineers, domain experts, a governance function and change leaders.
  • Process: stage gates, experiment design, incident response and a regular portfolio review.
  • Technology: data platforms, access to models, evaluation and monitoring tools, and security controls.
  • Governance: policies, risk classification, documentation and an inventory of AI systems. See AI Governance.

Dependencies and conditions. The chain only works when the data is available and trustworthy, someone owns the workflow being changed, the cost per completed task beats the alternative, and users trust the system enough to use it. Each condition can fail on its own, which is why so many promising pilots never reach production.

Where uncertainty remains. How much of the measured task-level gain shows up in company results; how reliable agents will become on long, multi-step tasks; how fast prices will keep falling; and how regulation will settle. Good strategies treat these as open variables and revisit them on a schedule.

Economics, Statistics and Scaling

AI strategy has its own "physics": a handful of quantitative relationships that explain why some plans pay off and others stall.

Prediction gets cheaper, so its complements gain value

Agrawal, Gans and Goldfarb argue that AI's core economic effect is to lower the cost of prediction, meaning the filling in of missing information. When an input gets cheaper, people use more of it, and the things used alongside it become more valuable. For AI those complements are proprietary data, human judgment about which outcomes matter, and the ability to act on a prediction. Generative AI extends the same logic to drafting text, code and images. The practical reading: look for decisions where a better prediction would change the action taken, and expect value to shift toward whoever owns the data, the judgment and the workflow.

Exposure is measured in tasks, not jobs

Researchers from OpenAI and the University of Pennsylvania estimated that about 80% of U.S. workers have at least 10% of their tasks exposed to large language models, and about 19% have at least half of their tasks exposed. Strategy therefore works at the level of tasks: list the tasks in a process, mark which ones AI can do, assist with or should not touch, and redesign the flow around that map.

The frontier is jagged

AI capability is uneven. It can excel at tasks that look hard and fail at tasks that look easy. In a preregistered experiment with 758 Boston Consulting Group consultants, those using GPT-4 on tasks inside the model's capability completed 12.2% more tasks, 25.1% faster, with more than 40% higher quality. On a task designed to fall outside that frontier, consultants using AI were 19 percentage points less likely to reach the correct answer than those working without it. Strategy has to map where each task sits relative to the frontier and re-map it as models change.

The productivity J-curve

General-purpose technologies such as electricity, computers and AI need large intangible investments: redesigned processes, new skills, data and organizational change. Brynjolfsson, Rock and Syverson show that measured productivity can dip before it rises, because these investments are costly and poorly captured in the statistics. AI programs should expect this pattern and judge early progress with leading indicators (adoption inside redesigned workflows, cycle time, quality) before lagging ones (profit).

Price curves and cost per task

Epoch AI reported in March 2025 that the price of reaching a fixed level of LLM performance has been falling by roughly 9x to 900x per year, depending on the task; for GPT-4-level performance on PhD-level science questions it fell about 40x per year. At the same time, reasoning models and agents consume far more tokens per task. The useful metric is therefore cost per completed task, including retries, tool calls and human review. In McKinsey's 2026 survey, one in five respondents said their organization limits AI use because of operating costs.

Expected value of a use case

A simple model keeps a portfolio honest:

Expected value = P(success) × (annual benefit × years in use) - (build cost + run cost + change cost) - expected risk cost

Here the expected risk cost is the probability of a harmful failure multiplied by its impact. Two consequences follow. A modest use case with a high chance of success and low risk can beat an ambitious one. And shared foundations that raise P(success) for many use cases at once (data quality, evaluation tooling, governance) are worth more than any single business case shows.

Reliability compounds

If each step in an automated chain succeeds with probability p, a chain of n independent steps succeeds with probability pn. At 95% per step, a 10-step agent finishes correctly only about 60% of the time (0.9510 ≈ 0.60). This is why good tactics favor short, checkable workflows, validation between steps and human review at high-stakes points before granting agents long autonomous runs.

Amdahl's law for workflows

If AI speeds up a fraction f of a process by a factor s, the whole process speeds up by 1 / ((1 - f) + f/s). Making 30% of a process ten times faster yields only about a 1.37x overall speedup. Large gains require redesigning the whole workflow, not just accelerating one step.

Experiment math

To detect a difference δ in a metric with standard deviation σ (5% significance, 80% power), each group in an A/B test needs roughly n ≈ 16σ²/δ² observations. Halving the effect you want to detect quadruples the sample you need. Choose metrics that can move, run tests long enough, and do not stop early because the results look good.

Goodhart's law

When a measure becomes a target, it stops being a good measure. Usage counts, tokens processed and the number of pilots are easy to inflate. Pair efficiency metrics with quality, risk and customer measures.

Principle What it predicts What to do
Cheaper prediction Value shifts to data, judgment and action Invest in proprietary data and decision ownership
Task exposure Work changes task by task Map tasks before redesigning roles
Jagged frontier Big gains on some tasks, errors on others Test each task; keep human review where AI is weak
J-curve Results lag investment Use leading indicators; fund complements
Price curves Capability per dollar keeps rising Re-evaluate models often; track cost per task
Reliability compounding Long chains fail more often Short workflows, checks, human approval
Amdahl's law Speeding one step barely helps Redesign the end-to-end process
Experiment math Small effects need large samples Size tests before running them

Why This Matters

Value is real but concentrated. Most organizations use AI somewhere, yet few see enterprise-level financial impact. McKinsey's 2025 and 2026 surveys both found that only about 6% of respondents qualify as high performers. Those organizations are about three times as likely to have fundamentally redesigned workflows, they set growth and innovation goals alongside efficiency, and their senior leaders visibly own AI initiatives and measure impact. The difference is mostly strategy and execution, not access to models.

Task gains do not automatically become business gains. Field experiments show large improvements on well-matched tasks (see Recent Developments), while studies of whole labor markets find small effects so far. The J-curve and Amdahl's law explain much of the gap: gains need complementary investment and workflow redesign.

Engineering implications. Strategy shapes architecture: shared data products and evaluation tooling, model-neutral integration so you can switch providers, cost monitoring per task, and safety controls proportional to the risk of each use case.

Risk and accountability. Expect to be held responsible for what your AI says and does. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline liable for its chatbot's wrong answer about bereavement fares. Autonomous agents raise the stakes: in July 2025, a coding agent on the Replit platform deleted a production database during an explicit code freeze. Governance belongs inside the business case, not after it. See Risk, Compliance and Regulation and AI Verification and Validation.

People and skills. Some AI tools help less-experienced workers most, while early-career employment in highly exposed occupations has fallen relative to other groups. Workforce planning, training and job redesign belong in the strategy from the start.

What would change if the optimistic case proves right. If agents become reliable on long tasks and costs keep falling, the bottleneck moves from technology to organizational redesign, data rights and governance capacity. What remains uncertain: whether today's task-level gains will appear in productivity statistics, which advantages will last as models become commodities (see Moat), and how quickly regulation will settle.

Recent Developments

Developments from roughly mid-2024 to October 2026 that matter for AI strategy and tactics. Each topic notes its evidence status so you can tell what is established: demonstrated findings come from randomized or quasi-experimental studies of specific settings; self-reported figures come from executive surveys; forecasts are analysts' informed hypotheses; preliminary results are early, small or contested; and the remaining open questions are listed at the end.

Enterprise adoption and returns

  • What was measured. McKinsey's global executive surveys: 1,993 respondents in June and July 2025, and 1,719 respondents from 97 countries in May and June 2026.
  • What they found. In 2025, 88% of organizations regularly used AI in at least one business function, up from 78% a year earlier, and 39% of respondents attributed some EBIT impact to AI, mostly under 5% of EBIT. In 2026, nearly nine in ten used AI, 44% reported scaling it across the enterprise (up from 38%), and 37% attributed any EBIT impact to it. High performers stayed at about 6%. Among organizations with more than $1 billion in revenue, 40% were scaling AI agents in at least one function, up from 27%, while smaller organizations held flat at 22%.
  • New signals in 2026. One in five respondents said their organization limits AI use because of operating costs, yet 60% expected to increase AI investment over the next year. And 32% said their organization had skipped buying at least one software product because agentic coding tools let them build it in-house, a shift in the build-versus-buy line.
  • Evidence status: self-reported. Good for tracking trends over time, but it does not establish causality, and companies measure EBIT effects in different ways.

The "GenAI Divide" claim

  • What was claimed. A report from MIT's NANDA initiative (July and August 2025) said about 95% of organizations saw no measurable profit-and-loss return from generative AI, mainly because tools did not learn from feedback or fit daily workflows.
  • Method. A review of more than 300 public AI initiatives, 52 structured interviews and 153 survey responses collected at industry conferences.
  • Other findings. Workers at over 90% of companies reported regular use of personal AI tools for work, while only about 40% of companies had bought an official subscription: widespread "shadow AI".
  • Evidence status: preliminary. The headline number traveled widely, but it rests on a small, conference-based sample and loose definitions of "return". Treat it as a sign of a real pilot-to-production problem, not as a precise failure rate.

Agentic AI: enthusiasm, standards and a reality check

  • Cancellation risk. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. It also estimated that only about 130 of the thousands of vendors marketing "agentic AI" offer genuine agent capabilities, calling the rest "agent washing". Its advice: pursue agents where the value is clear, rethink workflows rather than bolting agents onto legacy systems, and measure cost, quality, speed and scale.
  • Adoption. A Gartner webinar poll in January 2025 found 19% of organizations had made significant agentic AI investments and 42% conservative ones. By mid-2026, McKinsey found 40% of large enterprises scaling agents somewhere.
  • Standards. The Linux Foundation formed the Agentic AI Foundation in December 2025 around Anthropic's Model Context Protocol (MCP), Block's goose framework and OpenAI's AGENTS.md convention; Google's Agent2Agent (A2A) protocol joined in August 2026. Shared standards lower integration costs and lock-in, which matters for build, buy and partner decisions.
  • Practitioner consensus. OpenAI's and Anthropic's agent-building guides converge on the same tactics: start with the simplest workflow that works, prototype with the most capable model, then move routine steps to cheaper models where evaluations still pass, and keep humans in the loop for high-risk actions.
  • Evidence status: forecast plus self-reported adoption. Gartner's cancellation rate is a prediction, in effect an informed hypothesis to test against your own portfolio, and the adoption figures come from polls and surveys. The standards developments are established facts.

Field experiments on productivity

  • Customer support. Brynjolfsson, Li and Raymond studied the staggered rollout of a generative AI assistant to 5,179 customer-support agents at a software company (NBER working paper, 2023; published in the Quarterly Journal of Economics in 2025). Issues resolved per hour rose 14% on average and 34% for novice and low-skilled agents, with minimal impact on experienced, highly skilled agents. Customer sentiment improved and attrition fell. Limit: one company and one task family.
  • Consulting. The "jagged frontier" study of 758 BCG consultants (Dell'Acqua and colleagues at Harvard Business School, Wharton, MIT and BCG; now published in Organization Science) showed large gains inside the frontier and a real accuracy penalty outside it. Limit: short, designed tasks rather than months of client work.
  • Teamwork. The "Cybernetic Teammate" experiment with 776 Procter & Gamble professionals (NBER, 2025) found that individuals using AI performed about as well as two-person teams without it, that AI blurred the line between technical and commercial thinking, and that users reported more positive emotions. Limit: one-day innovation challenges.
  • Software engineering. METR's 2025 randomized trial found experienced open-source developers took about 19% longer on real tasks when allowed to use early-2025 AI tools, even though they believed the tools had sped them up. In a February 2026 update, METR said it believes newer tools are raising productivity but can no longer measure the effect reliably: many developers would not work without AI, which undermined the control group, so METR is reworking its study design. Lesson: measure in your own setting rather than trusting perceived speed.
  • Whole labor markets. Humlum and Vestergaard (University of Chicago Becker Friedman Institute, 2025) linked Danish surveys on AI chatbot use to administrative records and found no detectable effect on earnings or recorded hours two years after adoption, ruling out effects larger than about 2%.
  • Evidence status: demonstrated for the settings studied. These results come from randomized or quasi-experimental designs, and the customer-support and consulting studies are now peer reviewed. Together they show large gains on well-matched tasks, smaller or negative effects on poorly matched ones, and little effect so far in labor-market statistics. Reading that gap as the J-curve at work is a proposed interpretation, not a demonstrated fact.

Workforce effects

  • Using payroll data from the largest U.S. payroll provider, Brynjolfsson, Chandar and Chen (Stanford, August 2025) found that employment for 22 to 25 year olds in the most AI-exposed occupations fell about 13% relative to others, concentrated where AI tends to automate rather than augment work. More experienced workers in the same occupations held steady or grew.
  • In McKinsey's 2026 survey, 39% of respondents expected AI to reduce their organization's headcount over the next year. Yet only 14% of respondents from AI-using organizations said AI had contributed to a headcount decline over the past year, less than half the 32% who had predicted reductions a year earlier.
  • Evidence status: preliminary. The payroll study covers one data source over a short period, and the survey answers are self-reported. The direction matters for workforce planning even while the size and causes of the effect are debated.

Costs, investment and transparency

  • Prices for fixed capability kept falling quickly (Epoch AI, 2025), while capability rose.
  • Stanford's 2026 AI Index reported global corporate AI investment of roughly $580 billion in 2025, and a fall in the average Foundation Model Transparency Index score from 58 to 40, meaning major developers now disclose less about their models. Lower transparency makes vendor due diligence harder.
  • Evidence status: measured. These are compiled market data, though investment totals depend on how each source defines "AI investment".

Governance and regulation

  • European Union. The AI Act's bans on prohibited practices (since February 2025) and its obligations for general-purpose AI models (since August 2025) already apply. The Digital Omnibus, in force since 27 July 2026, moved high-risk obligations to 2 December 2027 for stand-alone systems (such as hiring, credit scoring and education) and 2 August 2028 for AI built into regulated products.
  • United States federal government. OMB memorandum M-25-21 (April 2025) requires agencies to name a Chief AI Officer within 60 days, requires the larger agencies to convene AI governance boards within 90 days and publish AI strategies within 180 days, and sets risk-management practices for "high-impact" AI. It is a useful public template for corporate governance.
  • Standards. The NIST AI Risk Management Framework and ISO/IEC 42001, a certifiable AI management-system standard, give organizations a shared vocabulary for risks and controls. See Ethics, Privacy and Law.
  • Evidence status: established. These are adopted laws, policies and standards; how they will be enforced and audited is still taking shape.

Open questions

How to measure AI's contribution to company-level productivity; how reliable long-horizon agents can become; whether advantages from proprietary data and workflow redesign persist as models become commodities; and how regulators will treat autonomous agents. Several research efforts aim to narrow these gaps: METR is reworking its developer-productivity study, payroll-based studies like the Stanford team's continue to track employment, and the AI Index updates its investment, adoption and transparency data every year.

Strategy

Determine the Best Strategy

Determine Best Strategy: There are several steps an enterprise can take to use signals, analyze trends, itemize uncertainties, and develop future scenarios to better understand opportunities and threats within the organization:

  1. Identify and analyze signals, including associated risks. Signals are important to identify, but there is always a risk of overlooking or misinterpreting relevant signals. The enterprise should conduct a comprehensive environmental scan and consider the risk of missing critical signals or interpreting them incorrectly.
    • Analyze trends. While trend analysis can provide valuable insights, there is a risk of assuming that past trends will continue in the future. The enterprise should consider the risks associated with potential changes in the market, customer behavior, or other factors that may impact the trends.
    • Itemize uncertainties and associated risks. Itemizing uncertainties is crucial to developing effective contingency plans, but there is a risk of overlooking important uncertainties or underestimating their potential impact.
    • The enterprise should consider the risks associated with potential uncertainties and develop contingency plans that are flexible enough to adapt to changing circumstances.
    • AI signals worth tracking today: capability and price per task of leading models; your own usage, cost and quality telemetry, including unofficial "shadow AI" use; competitor and customer adoption; regulatory milestones such as EU AI Act deadlines; and the talent market.
    • AI uncertainties worth naming: the pace of capability gains, agent reliability, cost curves, vendor stability and lock-in, data rights and copyright outcomes, and regulatory timing.
  2. Develop future scenarios; Goal. Scenario planning can help the enterprise anticipate potential outcomes, but there is a risk of becoming too focused on a specific scenario and failing to consider other possibilities. The enterprise should consider the risks associated with each scenario and develop strategies that are adaptable to a range of potential outcomes.
    • A practical AI scenario grid crosses two uncertainties, for example how fast AI capability per dollar improves (steady or rapid) and how fast customers and regulators accept automation (cautious or permissive). Moves that pay off in all four quadrants, such as data foundations, evaluation tooling and workforce skills, are "no-regret" moves. Others become options to exercise when a signpost appears.
    • Generative AI can help draft scenario narratives and stress-test assumptions, but deciding which futures matter stays with leaders.
  3. Plan and perform actions. Developing a plan of action is critical, but there is a risk of implementing actions that do not address the underlying issues or that have unintended consequences. The enterprise should consider the risks associated with each action and develop contingency plans in case the actions do not have the desired impact.
    • Turn the plan into a use-case portfolio with stage gates and preset success and stop criteria (see the table below and Tactics).

By considering the risks associated with each step, the enterprise can develop a more comprehensive understanding of the opportunities and threats within the organization and increase the probability of best outcomes while minimizing the risks of negative consequences. This helps the enterprise to stay ahead of the competition, adapt to changes in the market, and achieve its strategic objectives while effectively managing risks.

You Exec

Building the use-case portfolio

Score every candidate on the same criteria, then balance the portfolio across time horizons.

Criterion Questions to ask Signs of a strong candidate
Value Which decision or workflow changes? How large is the benefit in revenue, cost, risk, speed or quality? A measured baseline; frequent, high-volume decisions
Feasibility Is the task inside today's AI frontier? Does a quick test on real data work? Pilot results on real cases, not demos
Data readiness Is the data available, clean, permitted for this use and kept current? Named data owner; documented access rights
Risk What happens when it is wrong, and who is affected? What regulatory class applies? Errors are visible and reversible; human review is possible
Strategic fit Does it build a lasting advantage such as proprietary data, a better workflow or customer trust? Hard for competitors to copy (see Moat)
Time to value Can a pilot show results in weeks rather than quarters? Small, testable first release

Three horizons.

  • Horizon 1: productivity in existing work. Assistants for drafting, search over internal knowledge, coding help. Fast to deploy, low differentiation.
  • Horizon 2: redesigned core workflows. Claims, support, underwriting, procurement, software delivery. This is where high performers report most of their value.
  • Horizon 3: new products and business models. AI-native services and agents that act for customers. Highest uncertainty, highest potential.

Build, buy or partner

  • Buy when the use case is common and not a differentiator.
  • Build or customize when proprietary data or a distinctive workflow is the source of advantage.
  • Partner for scarce skills or to share risk on early bets.
  • Agentic coding tools are moving this line: in McKinsey's 2026 survey, 32% of respondents had skipped buying at least one software product because they could build it in-house.
  • Keep options open with model-neutral architecture and open standards such as MCP and A2A. For choosing approaches and models per problem, see AI Solver and Model Search.

Frameworks you can borrow

  • Andrew Ng's AI Transformation Playbook (2018). Five steps: run pilot projects to gain momentum, build an in-house AI team, provide broad AI training, develop an AI strategy, and develop internal and external communications. Note the order: early wins come before the grand plan.
  • AWS Cloud Adoption Framework for AI. Four iterative stages (Envision, Align, Launch, Scale). Work backward from business outcomes, and treat the data "flywheel" (better data, better predictions, better outcomes, more data) as the engine.
  • Microsoft Cloud Adoption Framework, AI adoption. Leaders define outcomes, decision rights and how value will be measured; a central function such as an AI Center of Excellence owns organization-wide policy while business groups innovate within it; employees get approved tools plus training; and teams choose the option that meets requirements with the least operational responsibility.
  • Gartner's agent guidance (2025). Use agents for decisions, automation for routine workflows and assistants for simple retrieval; measure cost, quality, speed and scale.

Videos: developing an AI strategy

Load video
YouTube
How do you develop an Artificial Intelligence Strategy for a business?
In this video I explain in very simple terms how a business can develop an AI or machine learning. Here is the link to the AI strategy template If you would like more information on this topic, please feel free to visit my website and sign up for content updates! I write articles every week on various different topics such as Big Data, Artificial Intelligence & Machine Learning.

Load video
YouTube
Amy Webb Launches 2023 Emerging Tech Trend Report | SXSW 2023
Join Amy Webb for the launch of the Future Today Institute's 16th edition of its Tech Trends Report and a deep dive into all the tech trends you'll need to follow in 2023. In this provocative session, futurist Amy Webb, CEO of the Future Today Institute and professor at NYU Stern School of Business, will provide a data-driven analysis for the emerging tech trends that need to be on your radar this year, and she’ll show you scenarios that will change your perspective on the futures. For those who have attended this session at previous SXSW festivals and already know what to expect from this session, you'll be in for an extra-special treat this year. At the end of the session, you will receive a special SXSW-edition of the Tech Trend Report, which is downloaded 1 million times every year. (A new edition of the report comes out each year; Webb's scenario approach is a worked example of step 2 above.)

Load video
YouTube
How to Realize Your AI Strategy?
Raj Ramesh Knowing about AI is one thing, but integrating it into the business could be a lot more challenging because there are many factors to consider. Here I explore some of the major things to consider. www.topsigma.com www.linkedin.com/in/rajramesh

Load video
YouTube
Does Your Company Need an AI Strategy?
Raj Ramesh Everyone in the business world is talking about AI. Companies are starting to implement AI without understanding where it’d give the business the biggest bang for the buck. How do you prioritize where to implement AI?

Load video
YouTube
Why do you need an AI Framework and an AI Strategy?
Raj Ramesh The terms framework and strategy are often confusing. If your organization is ready to use AI in its digital transformation (and it better be because the future of the digital organization will have lots of AI), then it needs a great strategy. Here is discuss how you can think about AI strategy and AI frameworks.

Load video
YouTube
So you want AI for your business. Where do you start?
Raj Ramesh AI is an important technology that will fundamentally shift how your business operates. But many companies are scrambling with AI without a focus. What’s a good way to think about it?

Strategy Example: Writing a Joke

Why a joke? Humor is a compact test of strategy versus tactics. The strategy is the choice of topic and angle that will surprise a particular audience; the tactics are repeatable steps that produce candidate punch lines. Comedy writer Joe Toplyn, a former head writer for late-night television, turned his method into Witscript, a system that has a language model follow those steps.

The Three-Part Joke Algorithm:

  1. Topic: This is the statement that the joke is based on. It usually involves a common or familiar situation, observation, or fact that the audience can relate to or understand. For example, a possible topic could be “I hate going to the dentist.”
  2. Angle: These are words that smoothly connect the topic to the punch line. They usually involve adding some details or context to the topic that make it more interesting or specific, but also hint at the direction of the punch line. For example, an angle using the topic above could be “I hate going to the dentist. Especially when he asks me questions while he’s working on my teeth.”
  3. Punch Line: These are words that result in a laugh. They usually involve revealing something unexpected or absurd about the topic or the angle that contradicts or exaggerates them. For example, a punch line using the angle above could be “I hate going to the dentist. Especially when he asks me questions while he’s working on my teeth. Like, ‘Do you floss regularly?’ or ‘How do you feel about polygamy?’”



Take these steps to write a three-part joke. The structure above describes the finished joke; this procedure is the order in which a writer, or a system like Witscript, builds it:

  1. Get a topic: This is the general subject or theme of the joke, such as politics, sports, animals, etc. It helps to narrow down the scope and context of the joke.
    • Select two topic handles: These are two specific words or phrases related to the topic that can be used to create a contrast or a connection. For example, if the topic is politics, the topic handles could be "president" and "election".
    • Generate associations of the two topic handles: These are words or ideas that come to mind when thinking of the topic handles. They can be positive, negative, neutral, or absurd. For example, some associations of "president" could be "power", "speech", "scandal", etc. Some associations of "election" could be "vote", "campaign", "fraud", etc.
  2. Create a punch line: This is the final sentence or phrase of the joke that delivers the surprise or twist. It usually involves combining or contrasting two associations of the topic handles in an unexpected or humorous way. For example, a possible punch line using the associations above could be "The president gave a speech about how he won the election by a landslide...of fraud allegations."
  3. Generate an angle between the topic and punch line: This is the intermediate sentence or phrase that sets up the punch line and connects it to the topic. It usually involves introducing one of the topic handles and one of its associations in a way that seems normal or logical at first, but leads to the punch line. For example, an angle using the punch line above could be "It's been a tough week for the president." This introduces the topic handle "president" and its association "scandal", but also creates a contrast with the punch line.

What the research shows

  • Structure helps, but does not reach human level. Witscript 2 (presented at the International Conference on Computational Creativity, 2022) used GPT-3 to follow exactly these steps: pick two topic handles, list associations, combine them into a punch line, then write the angle. Crowd workers judged its responses to be jokes 46.2% of the time, versus 25.1% for plain GPT-3 and 70.3% for a professional human comedy writer.
  • Professionals find LLMs useful for drafts, not for comedy. In a Google DeepMind study presented at ACM FAccT 2024, 20 professional comedians wrote material with ChatGPT and Google's Bard (running Gemini Pro). They valued help with first drafts and structure, but most felt the tools failed as creative partners: the output was bland and leaned on stale tropes, which one participant called "cruise ship comedy material from the 1950s, but a bit less racist." The average creativity-support score was 54.6 out of 100, which the authors rate as mediocre, and participants reported that safety filters blocked dark or edgy material and sometimes erased minority perspectives. The authors note that participants were recruited because they already used AI, which likely made their ratings more favorable than those of comedians overall.
  • The strategy lesson. Breaking a creative task into explicit steps makes AI much more useful, but selection, taste and delivery (the judgment complement from the economics above) still sit with people. The same pattern applies to marketing copy, product ideas and strategy memos.

Tactics

  1. Think backwards
  2. Build a data pipeline before a model
  3. Deliver actions over accuracy
  4. Modularise and abstract
  5. Brand the solution


These methods can be employed to plan and execute actions identified in the last step of your AI strategy. Organizations can benefit greatly from implementing these methods and from using generative AI to plan and perform actions. Each method carries its own risk, so each is paired below with a way to manage it. Managed this way, AI initiatives are more likely to succeed and to deliver the strategic objectives they were chosen for.

Core tactical methods

  • One example is predictive analytics, which can help organizations identify potential risks and outcomes associated with various actions. However, it's important to note that there is a risk of inaccurate predictions, especially when conditions shift away from the historical data (see Out-of-Distribution (OOD) Generalization). To mitigate this risk, track forecast error, report ranges rather than single numbers, and use generative AI to explore different outcomes and develop contingency plans in case predictions do not have the desired impact. See Predictive Analytics and Forecasting.
  • Another method is simulation modeling, which can help organizations identify potential risks and opportunities. However, there is a risk of relying too heavily on simulated data and making decisions that do not reflect the real world. Generative AI can assist in simulation modeling by generating a wide range of inputs and scenarios, enabling organizations to explore different potential outcomes. Organizations should validate simulations with real-world data and develop contingency plans in case simulated results do not have the desired impact.
  • Agile methodology is also a valuable tool for organizations, allowing for continuous monitoring and adjustment of actions based on feedback. However, there is a risk of losing sight of overall strategic objectives. Generative AI can assist in agile methodology by generating insights that inform the plan of action and help identify potential risks and opportunities. Tie each sprint goal to a portfolio objective so that fast iteration stays aligned with the strategy.
  • A/B testing is another method that can help organizations identify which version of a product or service is more effective. However, there is a risk of making decisions based on incomplete or biased data, or on tests too small to detect real effects (see the experiment math in The Physics). Generative AI can produce more variants to test, but every variant still needs a fair, adequately sized comparison.
  • Finally, continuous monitoring and evaluation can help organizations identify potential risks and opportunities in real time. However, there is a risk of relying too heavily on short-term trends and making decisions based on incomplete data. Generative AI can assist by summarizing signals and generating scenarios from historical and real-time data; pair it with statistical controls and clear escalation thresholds so that noise does not trigger action.

Tactics for generative AI and agents

  • Evaluation-driven development. Before choosing a model or building an agent, write a test set of real cases with expected results. Use it to compare models, catch regressions and decide when to ship. See Evaluation, Measures and AI Verification and Validation.
  • Start simple, add autonomy gradually. Use a single model call or a fixed workflow before an autonomous agent. Require human approval for irreversible actions, and give agents least-privilege credentials. The Replit incident above shows why instructions in a prompt are not the same as permissions.
  • Route by difficulty. Prototype with the most capable model, then move routine steps to smaller, cheaper models wherever evaluations still pass.
  • Manage cost per task. Monitor tokens, retries and human review time per completed task; set budgets; use caching and batch processing for work that can wait.
  • Redesign the workflow, not just the step. Map the end-to-end process, then change roles, handoffs, training and incentives. Amdahl's law explains why speeding up one step rarely moves the business result.
  • Channel shadow AI. Widespread unofficial use signals demand. Provide approved tools with data protections and training instead of relying on bans.
  • Red-team and monitor. Test for prompt injection, data leakage and harmful outputs before launch, and log every model and tool call in production.
  • Ground answers and allow "I don't know". For customer-facing assistants, answer from current, cited sources and route uncertain cases to people. The Air Canada ruling shows the cost of confident wrong answers.

From Kaggle to production

The table below contrasts a real business application with a Kaggle competition. Most of the work in production happens outside model building.

Application Kaggle Competition
Translate a business question into a data question  
Think about how the model is going to be consumed  
Validate assumptions and methods  
Consider how a machine learning model can connect to a (existing) tech stack  
Determine data sources  
Extract Data; SQL queries (sometimes across multiple databases), third-party systems, web scraping, API’s or data from partners Download some data (probably one or several CSV files)
Transforming and cleaning data; removing erroneous data, outliers and handling missing values Perhaps do a little cleaning, or chances are the data set may already be clean enough
Exploratory data analysis & feature extraction  
Feature engineering (vast range of variables) Feature engineering (a finite number of variables)
Perform preprocessing such as converting categorical data into numerical data Perform preprocessing such as converting categorical data into numerical data
Version models, Feature selection, Hyperparameter Tuning, and Web Service Endpoints  
Model selection; best model that can be integrated into the existing tech stack with the least amount of engineering Model selection
Build the model Build the model
Set up training and deployment pipelines  
Train the model Train the model
Validate the model Validate the model
Test the model Test the model
Optimise the model until it is ‘good enough’ considering the business value; perform hyperparameter tuning and compare results Run the data through a variety of suitable models until you find the best one; perform hyperparameter tuning and compare results
Make experiments reproducible  
If new tech stack…  
- Review prior art and available libraries, services, scafolding tools  
- Implement and test tech stack  
Deploy the model  
Monitor the model performance in production  
Set up monitoring alert systems  
Retrain when/if necessary  
Mature to automatically retrain and redeploy models  

What changes for generative AI and agents

The same lifecycle applies to applications built on foundation models, with new work at almost every step. See AIOps/MLOps for the operational side.

Lifecycle step Classic machine learning Added for generative AI and agents
Translate the business question Define the prediction target and metric Define the task, acceptable outputs, failure modes and who reviews them
Data Collect, clean and label training data Curate knowledge sources, access permissions and evaluation cases; version prompts and context like code
Model selection Compare algorithms on a validation set Compare foundation models on your own evaluation set for quality, latency and cost per task; plan for model updates
Build Feature engineering, training, tuning Prompting, retrieval, tool integration (for example through MCP); fine-tune only when evaluations show a gap
Validate and test Holdout and cross-validation Code-based checks, LLM-as-a-judge calibrated against human ratings, red-teaming, safety and security tests
Deploy Batch or real-time endpoints Guardrails, human approval steps, least-privilege credentials for agents, fallbacks
Monitor Data drift and accuracy Quality sampling, incident tracking, token and tool costs, user feedback
Retrain or update Retrain on new data Re-run evaluations on every model, prompt or data-source change; keep versions so you can roll back

Videos: building and choosing models

Several of these talks predate today's tooling. Theano development ended in 2017; Caffe's successor, Caffe2, merged into PyTorch in 2018; Chainer's developer, Preferred Networks, moved its research to PyTorch in December 2019; and Apache MXNet was retired to the Apache Attic in 2023. Today most work uses PyTorch, JAX or TensorFlow with Keras, usually starting from pretrained models (see Hugging Face). The fundamentals these talks teach (data quality, evaluation, bias and variance) still apply.

Load video
YouTube
The 7 steps of machine learning
How can we tell if a drink is beer or wine? Machine learning, of course! In this episode of Cloud AI Adventures, Yufeng walks through the 7 steps involved in applied machine learning. The 7 Steps of Machine Learning article: https://goo.gl/XEo6i2

Load video
YouTube
Nuts and Bolts of Applying Deep Learning (Andrew Ng)
The talks at the Deep Learning School on September 24/25, 2016 were amazing. I clipped out individual talks from the full live streams and provided links to each below in case that's useful for people who want to watch specific talks several times (like I do).

Load video
YouTube
Strategies for Implementing AI and Machine Learning Techniques
Speaker: Daniel Protz, Founder and CEO, FlavorWiki. How can traditional F&B companies adopt new approaches while maintaining organisational support, comparability of data and continuity of product understanding? In the session Daniel outlined three change management approaches to stimulate teams, win organisational buy-in and secure budget for non-traditional tools using Big Data, Machine Learning and AI. What you will learn about: Use AI and Machine Learning to enhance existing systems and practices, Gain support for the use of such methods within their teams, Explain to management how investment in emerging tools can drive project ROI and increase competitiveness. Organised by IFST's Sensory Science Special Interest Group.

Load video
YouTube
Your choice of Deep Net - Ep. 4 (Deep Learning SIMPLIFIED)
Deep Nets come in a large variety of structures and sizes, so how do you decide which kind to use? The answer depends on whether you are classifying objects or extracting features. Let’s take a look at your choices. Deep Learning TV on Facebook: https://www.facebook.com/DeepLearningTV/

Load video
YouTube
Lessons learned from 100 deep learning models
Summary of tips for getting higher accuracy on your CNN deep nets.

Load video
YouTube
CS231n Lecture 12 - Deep Learning libraries
Overview of Caffe/Torch/Theano/TensorFlow

Load video
YouTube
What is a Deep Learning Library? - Ep. 16 (Deep Learning SIMPLIFIED)
Deep Learning libraries provide pre-written, professional-quality code that you can use for your own projects. Given the complexity of deep net applications, reusing code is a wise choice for a developer. A library is a set of functions and modules that you can call through your own programs. Library code is typically created by highly-qualified software teams, and many libraries bring together large communities that support and extend the codebase. If you’re a developer, you’ve almost certainly used a library at one point in time. For a commercial-grade deep learning application, the best libraries are deeplearning4j, Torch, and Caffe. The library Theano is suited for educational, research, and scientific projects. Other available libraries include Deepmat and Neon.

Load video
YouTube
Comparing machine learning models in scikit-learn
We've learned how to train different machine learning models and make predictions, but how do we actually choose which model is "best"? We'll cover the train/test split process for model evaluation, which allows you to avoid "overfitting" by estimating how well a model is likely to perform on new data. We'll use that same process to locate optimal tuning parameters for a K-Nearest Neighbors (KNN) model, and then we'll re-train our model so that it's ready to make real predictions.

Load video
YouTube
Approaching (almost) Any Machine Learning Problem | by Abhishek Thakur | Kaggle Days Dubai | Kaggle
Abhishek Thakur is the chief data scientist at https://boost.ai building state-of-the-art chatbots primarily for banking and insurance industries. His passion lies in solving difficult world problems through data science. He is the co-organizer of the Berlin Machine Learning Meetup and not long ago was ranked no. 3 worldwide on the Data Science Platform Kaggle. He is a Competitions and Discussions Grandmaster. Abhishek did his Bachelors in Electronics Engineering from India and moved to Germany for pursuing MSc from University of Bonn, Germany with a focus on image processing and computer vision. He dropped out of PhD in 2015 and since then has been working in industries. Kaggle Days Dubai was held May 30 - April 1 2019 as a part of AI Everything. Participants came to meet, learn and code with Kaggle Grandmasters, and compete in a full-day offline competition. This edition is presented by LogicAI with sponsorship from Kaggle and Google Cloud. Kaggle Days are a global series of offline events for seasoned data scientists and Kagglers created by LogicAI and Kaggle.

Load video
YouTube
Building Robust Machine Learning Models
Modern machine learning libraries make model building look deceptively easy. An unnecessary emphasis (admittedly, annoying to the speaker) on tools like R, Python, SparkML, and techniques like deep learning is prevalent. Relying on tools and techniques while ignoring the fundamentals is the wrong approach to model building. Real-world machine learning requires hard work, discipline and rigor. Development of robust models requires due diligence during data acquisition phase and an obsession with data quality. Feature engineering, choice of evaluation metrics and an understanding of the model bias/variance trade-off is often more important than the choice of tools. Experienced machine learning engineers spend most of their time dealing with data-related issues, model evaluation and parameter tuning while spending only a fraction of their time in actual model building. This is the 80/20 rule. Unlike most talks these days, this talk is not about deep learning. We will ignore the hype and strictly focus on fundamentals of building robust machine learning models.

Load video
YouTube
Guide to choose right deep Learning framework for your AI project
Rishikesh As world rolling around Artificial Intelligence (AI), demand for the AI-based product seen exponential growth, so the AI research. Deep learning algorithms and techniques are widely used for research and development of these products. Good news is that year by year Deep Learning has seen its glory in the release of many open source frameworks which ease the pain to develop and implement these algorithms. As there are many deep learning frameworks out there and it can lead to confusion as to which one is better for your task. And choosing a deep learning framework for an AI project is as important as choosing a programming language to code product, Data science project coupled with the right deep learning framework has truly amplified the overall productivity. In this talk, I will discuss the common points which help developers to understand which framework will be the perfect fit for solving given business challenges. Also, we will look into some of the most widely used frameworks and comparing with standard benchmarks. Following deep learning/machine learning frameworks will be discussed: 1. Tensorflow, 2. PyTorch, 3. Chainer and/or MXNET Key Highlight of this talk: define key points to judge any deep learning framework. hardware dependencies. anatomy of widely used open source frameworks. comparison of above-mentioned frameworks as per defined key points. Who is the audience? Anyone who inspired to code deep learning algorithms. pyconza2018 python

Artificial Intelligence Center of Excellence (AI CoE)

Establishing an AI Center of Excellence is a strategic move that can significantly enhance your organization's AI capabilities. Remember that AI adoption is an ongoing journey. Continuously monitor progress, iterate on your strategy, and adapt to changing technology and business landscapes. By following this plan, your organization can establish a robust AI CoE that drives innovation, efficiency, and value across the board.

Choosing an operating model

A CoE is one way to organize AI work. Most large organizations end up with a hybrid that changes as they mature.

Model How it works Fits when Watch out for
Centralized One team owns platforms, talent and delivery Early stages; scarce skills; high regulatory exposure Becoming a bottleneck far from the business
Hub-and-spoke (federated) A central hub sets standards, platforms and governance; business units build with embedded teams Scaling across many functions; the most common pattern today Unclear decision rights between hub and spokes
Decentralized Each business unit runs its own AI Very different businesses under one owner Duplicated effort, inconsistent risk controls, shadow AI

Microsoft's Cloud Adoption Framework recommends a version of the hub-and-spoke model: a central function such as an AI CoE owns organization-wide AI policy and responsible-AI requirements, while business groups innovate within those boundaries. Government shows the same pattern: under OMB memorandum M-25-21, U.S. federal agencies must appoint a Chief AI Officer and the larger agencies must run AI governance boards.

Here's an outline to guide your organization in adopting AI effectively and maximizing its value:

  • Create the AI Vision:
    • Define a clear and compelling vision for AI adoption within your organization. Understand how AI aligns with your business goals, mission, and long-term strategy.
    • Engage key stakeholders, including executives, business leaders, and technical experts, to ensure buy-in and alignment.
    • Name an accountable executive owner. Many organizations now appoint a Chief AI Officer; McKinsey's high performers are about three times as likely to have senior leaders who strongly own AI initiatives.
  • Identify Use Cases:
    • Conduct a thorough assessment of your organization’s processes, pain points, and opportunities.
    • Identify specific use cases where AI can add value. Prioritize use cases based on their potential impact, feasibility, and alignment with strategic objectives, using the portfolio criteria in Strategy.
  • Determine Ambition Level:
    • Define the level of ambition for your AI initiatives. Consider factors such as budget, resources, and risk tolerance.
    • Decide whether you want to start with small-scale pilots or aim for broader, organization-wide AI adoption.
    • High performers aim for transformation and growth, not only efficiency, but they also invest in the foundations that ambition requires.
  • Create a Data Architecture:
    • Establish a robust data infrastructure to support AI initiatives.
    • Ensure data quality, security, and accessibility. Consider data governance, privacy, and compliance requirements.
    • Explore cloud-based solutions for scalability and flexibility.
    • Treat key datasets and knowledge sources as products with owners, and add the access permissions and freshness rules that retrieval-based AI depends on.
  • Manage External Partnerships:
    • Collaborate with external partners, such as AI vendors, research institutions, and industry experts
    • Leverage their expertise, tools, and technologies to accelerate AI development
    • Establish clear communication channels and expectations
    • Check vendor claims (Gartner's "agent washing" warning), data-use terms, model portability and support for open standards such as MCP and A2A.
  • Identify AI Champions:
    • Identify individuals within your organization who are passionate about AI and can drive its adoption.
    • These champions can be from various departments: data science, IT, business units, etc.
    • Empower them to lead AI initiatives, advocate for AI adoption, and share best practices.
  • Use AI Technology:
    • Explore AI tools, frameworks, and platforms
    • Choose technologies that align with your organization’s needs and goals
    • Start with a small-scale project or proof of concept
    • Validate AI concepts and demonstrate feasibility
    • Provide a shared platform: approved models, an evaluation harness, logging, guardrails and cost dashboards, so each team does not rebuild them.
  • Share Success Stories:
    • Communicate wins and achievements related to AI projects, perhaps by sharing them on a wiki.
    • Showcase how AI has improved processes, efficiency, customer experiences, or revenue.
    • Use success stories to build momentum and encourage broader adoption.
  • Plan for Change Management:
    • Communicate the AI roadmap to stakeholders.
    • Address concerns, manage expectations, and ensure smooth implementation.
    • Train broadly (step 3 of Ng's playbook), redesign roles rather than only adding tools, and publish what AI will and will not be used for.
  • Govern and Measure:
    • Run an AI governance board, keep an inventory of AI systems, and classify each use case by risk (for example, "high-impact" uses under OMB M-25-21 or "high-risk" systems under the EU AI Act).
    • Report value and incidents together, so the portfolio is judged on both. See AI Governance.

Load video
YouTube
Load video
YouTube

Business Case

YouTube search... ...Google search

An AI business case answers four questions: what will change, what it is worth, what it costs in full, and how you will know. Recent evidence makes the last question the most important: many organizations use AI widely but cannot show an effect on profit, so a business case without a measurement plan is a hope, not a case.

Building the case

  • Baseline. Measure today's cost, cycle time, error rate, revenue or customer satisfaction before anything changes.
  • Value levers. Revenue growth, cost reduction, risk reduction, speed, quality, and employee or customer experience. High performers pursue growth and innovation, not only cost.
  • Full cost. Build cost, run cost (model access, tokens, compute, licenses), integration, change management, training, governance and monitoring.
  • Risk. Probability and impact of errors, regulatory exposure, vendor dependency and reputational harm. Include the expected risk cost in the numbers.
  • Measurement plan. Leading indicators (adoption in the redesigned workflow, cycle time, quality scores) and lagging ones (EBIT, revenue), with experiment designs and decision dates.
  • Stop rules. Preset criteria for stopping, pivoting or scaling, so pilots do not linger.
Measure Formula or definition Use it to
ROI (benefit - cost) ÷ cost Compare projects of different sizes
Payback period Months until cumulative benefit exceeds cumulative cost Judge cash and timing risk
Net present value Discounted benefits minus discounted costs Compare long-lived investments
Cost per completed task Total run cost ÷ tasks completed to standard, including retries and human review Compare AI with the current process and between models
Expected value See The Physics Rank a portfolio that includes risk

Lessons from recent cases

  • Count quality, not just volume. Klarna reported in February 2024 that its AI assistant handled about 2.3 million customer conversations in its first month, work it equated to about 700 full-time agents. In 2025 its CEO said the push had lowered service quality, and the company began hiring human agents again. A business case should track customer outcomes and keep a human path for complex cases.
  • Evaluate before you scale. Morgan Stanley built evaluations of its advisor assistant before rolling it out; more than 98% of advisor teams now use it. Adoption inside the target workflow is a strong leading indicator.
  • Price in liability. The Air Canada ruling shows that wrong answers from a customer-facing assistant are the company's responsibility.

Load video
YouTube
Artificial Intelligence: Building the Business Case for AI (CXOTalk #246)
Artificial intelligence can make companies dramatically more efficient, but investing in the technology can come with risks and complications. Tiger Tyagarajan, CEO of Genpact, tells Michael Krigsman of CXOTalk about the best strategies for buying AI to improve your business.

Load video
YouTube
Building a Business Case for your Machine Learning Idea
This presentation will discuss building a business model for your machine learning idea. Our presenter, Neeti Gupta, will provide a 10-step checklist with examples for the audience to build their own business model.

Business Strategy/Consulting

Youtube search... ...Google search

What the evidence says about AI in strategy work

The 2017 question of whether AI will replace elite consultants now has experimental evidence behind it. The best-known experiment on AI in consulting work, the "jagged frontier" study of 758 BCG consultants, found big gains on idea generation, writing and analysis inside the model's capability, and a drop in accuracy on a business-analysis task outside it. The "Cybernetic Teammate" experiment at Procter & Gamble found that an individual working with AI can match a two-person team on idea generation. Together these suggest a practical division of labor:

  • AI is strong at scanning and summarizing signals, drafting scenarios and options, generating alternatives, challenging assumptions (a "red team" role), and turning data into first-pass analysis.
  • People remain essential for deciding what matters, judging situations unlike anything in the training data, weighing values and trade-offs, and being accountable for the decision.
  • So, can AI generate corporate strategy? Today's evidence supports AI as a co-pilot for strategy work, not its author. Strategy teams that use it well treat its output as drafts to test, verify facts against sources, and keep the final choice and its rationale with people.

Note that the 2020 consulting guide above predates generative AI; most firms have since built internal AI tools and AI practices.

Load video
YouTube
How to Apply AI in Business
Raj Ramesh AI is a cool technology, but how do you use it effectively in your business operations. Solving the AI problem is only a small part of the whole solution, and you have to consider other things as well. www.topsigma.com www.linkedin.com/in/rajramesh

In this video, I walk through an example of using an AI system to automate a parking lot business operations

Load video
YouTube
How to Insert AI into business processes and what are the critical factors to consider?
Raj Ramesh Today many businesses are trying to transform to adapt to the new realities of the world. Future business models are often dependent on available technology. For example, the Uber business model is made possible because of smartphones. Technology will be a major component of those future state of any business. One of the main technologies is artificial intelligence (AI) and businesses that master AI will establish themselves in a leading and competitive position. Many studies have indicated that the faster the business integrates AI, the faster it will move forward. This means that if you are late to the game, it is almost impossible to catch up. So the question to ask is how do you establish that lead? Having gone through large scale multi-billion dollar transformations myself, and also being deeply familiar with AI, I offer some secrets on successful AI-based transformations. www.topsigma.com www.linkedin.com/in/rajramesh

Load video
YouTube
Meet Sheldon, the Future of Consulting
By Emmanuel Jusserand, Partner at Accenture Digital Strategy

Load video
YouTube
How AI is Changing Consulting
Harvard Business Review Facebook Live Presentation

Load video
YouTube
AI for Business: How Should We Frame It?
Boston Consulting Group BCG Partner and Managing Director Philipp Gerbert presents on the benefits of AI to a group of senior leaders at MIT. Welcome to the official YouTube channel for The Boston Consulting Group. Since 1963, BCG’s experts have been helping businesses, governments, and non-profit organizations build lasting advantage. As a leading management consulting firm, BCG empowers clients to make the changes needed to seize the competitive advantage and win. The independent spirit handed down from Bruce Henderson, BCG's founder, always challenging the status quo, has given the firm the courage to look beyond the obvious in pursuit of lasting solutions for our clients. Subscribe to BCG’s YouTube channel: https://goo.gl/hsFsVT Visit us at https://www.bcg.com

Load video
YouTube
How AI revolutionises business strategy
For centuries, the decisions made in a company were the responsibility of the top managers. But when firms harness AI and big data, algorithms can make many more decisions in the same time, and probably better ones. Kenneth Cukier explores how this affects the ways that companies are organized and how they compete and set strategy (as opposed to just execution) as he outlines what is happening now and what to expect in the future.

Load video
YouTube
The Future of Consulting
Recently did a Q&A jam session with CMC-Manitoba on the future of management consulting (one of my favourite topics). Like most industries, we are also being disrupted from all sides: the proliferation of the gig economy, automation, start-ups, the commoditization of information, etc. One of the ways to remain competitive is to hire and train "next-generation consultants". Biography: Shawn Kanungo is a strategist and keynote speaker who operates at the intersection of creativity, business and technology. He has been recognized nationally and globally for his work in the innovation space after 12 years working at Deloitte. Shawn’s mandate at the firm was to help corporate executives to better understand and plan for the opportunities and threats associated with disruptive innovation. Now as General Partner of Queen & Rook Capital, he is focused on applying exponential technologies and new business models to mature businesses. Shawn is a practitioner who has worked hand-to-hand with hundreds of organizations on their journey to digital transformation. He has adopted the concepts of behavioural economics, user-centered design, crowdsourcing, artificial intelligence, drones and film to help create world class client experiences. His work and interviews have been featured in Forbes, The Globe & Mail, The Guardian, CBC and CTV. In 2016, he was recognized as Avenue Magazine's Top 40 Under 40. He has spoken at TEDx in 2017, and named to Inc's 100 Most Innovative Leadership Speakers in 2018. His podcast, The Remix with Shawn & Wang, can be currently found on iTunes, Google Play and Stitcher. Website: https://www.shawnkanungo.com/

Sources and Further Viewing

Additional Viewing

Load video
YouTube

Andrej Karpathy: Software Is Changing (Again) (Y Combinator AI Startup School, June 2025, 39 minutes). Karpathy describes "Software 3.0", where natural language becomes a programming interface, and argues that the most useful near-term products are "partial autonomy" apps with an autonomy slider that keeps people in a fast generate-and-verify loop. It is a grounded counterweight to agent hype and a useful frame for the "start simple, add autonomy gradually" tactic on this page. Watch on YouTube

Load video
YouTube

Andrew Ng: Opportunities in AI (Stanford Graduate School of Business event, July 2023, over 30 minutes). Ng frames AI as a general-purpose technology like electricity, contrasts supervised learning (most of the value created so far) with generative AI (growing fastest), maps where opportunities sit across the AI stack, and explains how his venture studio validates ideas before building. It pairs well with his AI Transformation Playbook cited on this page. Watch on YouTube

Load video
YouTube

Erik Brynjolfsson: The AI Awakening, introductory lecture (Stanford ECON295/CS323, Stanford Online, August 2024, 1 hour 12 minutes). Brynjolfsson, director of the Stanford Digital Economy Lab, opens his course on how AI is reshaping the economy and society; the Lab features this lecture as the course introduction. His research is behind several ideas on this page, including the productivity J-curve and the customer-support field study, so the lecture is good background for The Physics and Recent Developments. Watch on YouTube

Load video
YouTube

Reid Hoffman: Business of AI (Stanford ECON295/CS323, Stanford Online, September 2024, 1 hour 18 minutes). LinkedIn co-founder and investor Reid Hoffman joins Brynjolfsson's course for a question-and-answer session about building and running businesses with AI. Watch it alongside the Business Case section for a founder and investor perspective on where AI creates value. Watch on YouTube

Key Sources

Adoption, value and strategy

Economics and evidence

Agents and delivery

Governance and regulation

Cases

Humor research

Tooling history

Retrieved from "https://primo.ai/w/index.php?title=Strategy_%26_Tactics&oldid=46761"