AI, Productivity and Decision-Making

The binding constraint on better AI decisions is rarely the technical team's capacity. It is the senior leadership's ability to evaluate what the technical team brings forward. The course is built for the latter.

By the end of the programme, senior leaders can put precise questions directly to vendors and internal AI teams. These include:

• Substitution and risk. Where in our operations is AI substituting for human work, and where would it be unsafe to let it act on its own?
• Augmentation and productivity. Where is AI complementing our existing workforce rather than replacing it — and what does the workflow have to look like to capture those productivity gains?
• Real oversight versus rubber-stamping. If a deployment requires a human sign-off, what test would tell us whether that sign-off is real or just a rubber stamp?

From Tools to Judgement

Most AI courses for senior leaders fall into two camps: short workshops on prompting and tool use, or technical programmes that assume a quantitative background. Neither matches what most boards actually need.

This course sits in the gap between them. It treats senior decision-making, not the technology itself, as the binding constraint. The objective is to give the leader the discipline to read AI claims critically, the framework to decide what to invest in, and a way to design how humans stay in charge when an AI is acting on its own.

What You Leave With

The workshops are where the delegate develops their organisation's position on AI. They leave with four short documents that capture it:

• Tier Map. A one-page sort of the firm's expertise into three tiers by codifiability. Identifies where AI substitutes for, complements, or has no effect on human work.
• Workflow Audit. One workflow decomposed into steps, each step classified by AI's role: substitute, complement, or augment.
• Human-Decision Map. For each step of the audited workflow, who makes the call — the AI or the human. Includes a test for when the human's involvement is real, not just rubber-stamping.
• Strategic Memo. One page setting out the firm's position on AI: what it will do, what it will not do, and why.

Format, Audience & Dates

Each day follows the same pattern:

• Morning. A 90-minute briefing on the framework that day rests on, followed by discussion.
• Afternoon. A workshop applying that framework to the AI decisions the delegate's own organisation actually faces.

The binding constraint plays out differently across firm types. How the course applies, by firm type:

• Large firms with internal AI teams. Equips you to challenge the AI team's recommendations rather than approving them. Where in the firm is AI substituting for human work? Where is human sign-off real, and where is it a rubber stamp?
• Smaller firms with no internal AI team. At small scale, a misallocated AI investment is harder to absorb. The course builds the judgement to avoid the most common mistakes before they are made.
• Market-facing firms. AI is a source of potential advantage. The questions: where is the firm's edge defensible against AI, and which workflows should the firm compress with AI before rivals do?
• Regulated monopolies. AI is a regulated input. The questions: which AI-driven efficiency gains will the regulator incorporate into the next cost review, and at which decision steps does the regulator expect a human to remain in charge?

Day 1 opens with three questions a senior leader should be able to answer about AI: what are the different kinds on offer? What does each cost to deploy at scale? What value can each be expected to deliver?

The aggregate forecasts disagree sharply. Goldman Sachs projects cumulative GDP gains of 6.1% from AI over a decade; Acemoglu estimates 0.9%. Both are plausible readings of the same evidence. The gap is methodological — about whether AI augments or substitutes for human labour, and which cognitive tasks fall within its reach. Day 1 introduces the framework the course offers for working through that disagreement at the firm level.

Morning · Reading the technology narrative

The morning opens on Yann LeCun's (ex Chief AI Scientist at Meta) argument that intelligence, at its foundation, is the ability to learn. Each successive generation of AI learns differently — which means each one automates different things, and leaves different things to humans.

• Machine Learning. Learns from large quantities of labelled examples. Reaches work that can be written down as rules, procedures, or labelled cases.
• Generative AI. Learns based on vast text and image corpora. Impact on tasks that mix pattern-recognition with context, judgement, or tacit know-how — work that is not fully codifiable but where the pattern can be inferred from large-scale exposure.
• Agentic systems. Generative AI extended with planning, tool use, and memory across multiple steps. The system acts across a workflow rather than producing a single output per prompt. Raises the question of where human judgement that AI cannot reach must remain in charge.

Morning · The Three Tiers framework

Each AI variant above reaches a different kind of work, distinguished in part by codifiability — the degree to which expertise can be written down as rules, procedures, or labelled examples.

The Three Tiers framework maps the three categories to a single question — one whose answer is shifting: does AI substitute for this work, complement it, or leave it untouched?

• Tier 1 · Codifiable. Anything that can be written down as a rule, a procedure, or a labelled example. Machine Learning reaches here. AI substitutes for human labour.
• Tier 2 · Partially Codifiable. Work that mixes pattern-recognition with context, judgement, or tacit know-how. Generative AI reaches here. The worker remains in the task, but the labour input required per unit of output falls — and with it the time per task, the headcount per unit, and the unit cost.
• Tier 3 · Residual. Knowledge that resists codification — the framing of problems, the weighing of trade-offs that have no agreed metric, and the judgement calls that carry institutional and ethical consequence. AI amplifies those who operate well inside it.

Morning · The economics of AI

Three Tiers names where AI reaches. The economic frame names what it costs to put it there, and how those costs will move.

• Capital-labour shift. Classical economics predicts diminishing returns to labour as capital is fixed. AI breaks that constraint — it is scalable digital capital that does not degrade. The distributional consequence follows directly: a capital-biased technology shifts returns toward the owners of infrastructure and away from the people who work with it. For senior leaders, the question is not only where productivity gains appear, but who captures them — and whether the firm is on the capital or the labour side of that divide.
• Tokens, not requests. AI infrastructure scales on tokens, not requests. A typical enterprise request consumes thousands of tokens before the user types anything — system prompts, retrieved context, multimodal inputs. Budgeting AI on per-request prices misses the real cost picture; budgeting on tokens names the actual unit of compute, memory, latency, and cost.
• Tokenomics is the new headcount. Token allocation is now a leadership decision in the same shape as headcount allocation: who gets tokens, how many, for what work. The relevant comparison for an AI deployment is no longer a software-line-item cost; it is the cost of a human doing the same work.
• The recalibration gap. Token prices have fallen sharply under competitive pressure. What an AI task costs to run today will not be what it costs next quarter. The Jevons paradox cuts the other way — cheaper tokens induce more token-intensive workflows — so total cost can rise even as unit prices fall. As token prices fall, the set of economically viable AI tasks expands — but only for organisations that periodically reassess the threshold, not those locked into static cost assumptions.

Afternoon · Application

Two exercises:

• Tier mapping workshop. Each delegate maps their organisation's distinctive expertise across the three tiers, with named examples in each, and produces the first completed template — the Tier Map.
• Critical reading exercise. The cohort works through three contemporary AI claims — a consultancy essay, a peer-reviewed study, and a vendor pitch — and applies a common critical standard to each.

Day 1 ends with the Tier Map complete — a working classification of where AI reaches the organisation and where it does not.

Day 2 takes the question many senior leaders are now facing: where are the AI productivity gains, and what does capturing them actually require?

Morning · The productivity question

The morning works through the empirical productivity evidence:

• The productivity-paradox lag. Productivity gains from general-purpose technologies lag adoption. Electricity was commercially deployed in the 1880s; the gains did not appear until the 1920s — because the unlock required factories to be redesigned around the new input, not retrofitted to it. The lag was not technological. It was organisational. The same dynamic applies to AI now: productivity is conditional on workflow redesign, not tool adoption.
• The translation gap. Distinguishing between a productivity claim made in a controlled study and the same claim translated into a real organisation, where adoption depends on cost, scale, and the fact that most firms do not operate at the volumes a study assumes.
• The heterogeneity of gains. A large customer-service study illustrates the point directly: an AI-assisted conversational system raised overall productivity by 14%, but bottom-quintile workers gained 35% while top-quintile workers gained just 2%. Transformative for some, marginal for others — in the same firm, on the same tasks, in the same week.
• The jagged frontier. The Harvard-BCG consultant study quantifies what AI can and cannot do. Inside the frontier, AI-assisted consultants completed tasks 25% faster and produced work rated 40% higher in quality. Outside the frontier — on tasks that appeared similar but required reasoning the model could not reliably perform — AI-assisted consultants performed 19% worse than those without AI.

Morning · The economics of deployment

Two cost dynamics that condition how the productivity evidence translates into a real deployment:

• Volatile costs. Token consumption is uneven — a small share of requests does most of the damage, and budgets planned on the average miss the upper-tail surge.
• Agentic cost dynamics. A single agentic workflow can consume hundreds of times more tokens than a one-shot query, and enterprise AI budgets intended to last a year are routinely exhausted in the first quarter. The morning closes on the governance needed to monitor this.

Afternoon · Workflow audit and agentic deployment

• Workflow audit. Each delegate takes one workflow from their Tier Map and runs it through a structured audit, identifying where AI substitutes for human work, where it compresses it, and where it augments it. The completed Workflow Audit is the second template.
• Agentic coordination. Standard exposure measures count task-level automation; agentic AI automates the coordination between tasks — the sequencing and handoffs that standard measures do not capture. A live demonstration, led by a senior practitioner, shows an agent operating across connected tools and identifies precisely where human oversight cannot yet be removed.

By the end of Day 2, each participant has a completed Workflow Audit — one process decomposed, each step classified, and the productivity conditions identified.

Day 3 turns the analytical work of the first two days into decisions. It formalises the distinction between what AI can predict and what the leader must judge.

Morning · The prediction–judgement worksheet

The economic foundation for the prediction–judgement distinction comes from Agrawal, Gans and Goldfarb. AI lowers the cost of prediction; human judgement determines the payoffs that prediction is fed into. Their central result: prediction and judgement are complements over hidden costs. As prediction improves, the value of human judgement about downside risk rises, not falls.

Trust in AI is built specifically, not generally. The Johns Hopkins cancer-detection system — a narrow model on a single dataset, focused on a defined task with accountability when it fails — earns trust because of its specificity. An AI agent that can be prompted to do anything earns no trust because it has no edges. The worksheet exists to force the leader to make the trust question specific to the decision in front of them.

AI systems produce recommendations and confidence scores but do not record the reasoning that should accompany a consequential decision — nor the conditions under which the recommendation should be overridden.

Before acting on a significant AI recommendation, the leader completes a short worksheet that records two things:

• Evidentiary threshold. What would have to be true for this recommendation to be wrong.
• Causal mechanism. The link from the proposed action to the expected outcome.

It is not a check on the AI; it is a check on the leader's own reasoning — made necessary precisely because the AI's output can feel more certain than it is.

Afternoon · The human-decision map

Two strategies have emerged for human-AI collaboration. The first maintains a clear division of labour: the human identifies which tasks fall inside the AI frontier and delegates only those. The second integrates human and machine at every step, with each AI output reviewed before the next iteration begins.

The first economises on human attention; the second provides continuous oversight at higher cognitive cost.

The afternoon returns to the workflow audited on Day 2 and specifies which strategy applies where. Across the literature, the human role at the points where AI does not act alone is converging on a four-step cycle: Define what the AI is supposed to do; Configure it to do that well; Measure whether it is doing it; Refine when it is not. The Human-Decision Map specifies, for each step in the audited workflow, who owns each part of that cycle.

• Human-decision map. For each step in the workflow, who makes the call — the AI or the human, and under which collaboration strategy. At the human steps, what the human is specifically responsible for across the Define / Configure / Measure / Refine cycle.
• Rubber-stamp test. A clear, testable rule for distinguishing real human involvement from sign-off that adds nothing.

Day 3 closes with two completed templates: the Human-Decision Map, specifying where human judgement is required and what it must do; and the Strategic Memo — one page setting out what the firm will do with AI, what it will not do, and why.

Each workshop produces one working document, built from the participant's own organisation. The four templates — Tier Map, Workflow Audit, Human-Decision Map, and Strategic Memo — form a connected sequence: each feeds the next, and together they constitute a board position on AI grounded in the firm's own analysis rather than a consultant's generic framework.

Worked example

Participants work through a complete worked example before applying the method to their own firm. The primary example is built for a UK Distribution Network Operator operating under Ofgem's RIIO framework — a regulated monopoly in which AI's role in cost efficiency and service quality is under regulatory scrutiny. Additional examples across commercial and professional-services contexts are in development.

The programme is taught by Cambridge Faculty of Economics, supported by senior industry advisors actively building AI in industry today.

Pricing for open-enrolment delegates and bespoke single-company variants is available on request.

For pricing, dates, or to register interest, please contact the Programme Director, Dr. Melvyn Weeks (Faculty of Economics, University of Cambridge) — mw217@econ.cam.ac.uk.

The programme is currently in development. Certification through the Møller Centre, Churchill College, University of Cambridge is under consideration.