AI Productivity Is a Work Design Problem

TL;DR

• AI reduces the effort required to produce work. It does not reduce the effort required to choose, review, coordinate, and own that work.
• If output grows faster than review capacity, the result is a larger queue, more rework, and higher cognitive load. That is not productivity.
• Sustainable AI work requires explicit workload budgets, stop-doing decisions, protected improvement time, and managers who measure impact rather than volume.

---

AI productivity is usually measured at the wrong boundary.

A team generates eight times more code, twice as many campaign variants, or five prototypes in the time it once took to make one. The increase is visible, so leaders call it productivity. Then the review queue grows. More ideas become active projects. People spend their day supervising agents and their evening finishing the judgement-heavy work that never disappeared.

The organisation created more output. It did not necessarily create more value.

This is a work design problem. AI changes the capacity of one part of a system, which exposes constraints elsewhere. Leaders who respond by raising every target convert technical capacity into workload pressure. Leaders who redesign the system convert it into better outcomes.

The productivity paradox

Every knowledge workflow has at least four kinds of work:

1. Choosing: deciding which problem deserves attention.
2. Producing: creating code, analysis, copy, designs, or recommendations.
3. Reviewing: checking quality, risk, coherence, and fitness for purpose.
4. Owning: acting on the result and accepting its consequences.

AI compresses production first. The other three move more slowly because they depend on context, judgement, authority, and trust.

Imagine a team that could produce ten meaningful changes each week and review ten. An agent increases production capacity to forty, while review capacity rises to fifteen. The team has not gained thirty units of throughput. It has created a queue of twenty-five unreviewed changes every week.

The queue shows up as longer pull request waits, shallow approvals, abandoned prototypes, duplicated work, or decisions made from outputs nobody properly interrogated. Eventually quality drops or people work longer to close the gap.

The constraint moved. The operating model did not.

Output has an expansion effect

Efficiency rarely returns time to the employee who created it. Organisations tend to reinvest spare capacity into more scope.

A PM who can prototype in an afternoon is asked to explore five options. An engineer using coding agents receives a larger area to own. A designer who can generate variations is expected to test all of them. Each local improvement looks reasonable. Across the week, the person now carries more active decisions, more review obligations, and more context.

This expansion effect explains why teams can feel faster and more overloaded at the same time.

The answer is not to suppress the tools. It is to budget the scarce work explicitly. Every plan should account for:

• Human review time
• Decision-making capacity
• Agent setup and maintenance
• Coordination across affected teams
• Recovery when generated work fails

If a plan only estimates generation time, it is not a capacity plan. It is a demo estimate.

Review capacity is a product constraint

Review is not administrative overhead. It is where accountability becomes real.

The AI accountability principle requires someone to define success, interrogate assumptions, test important claims, and own the result. Better agents can automate parts of verification, but they also generate more work to verify. Review capacity therefore has to scale with the risk and volume of AI-assisted output.

Use three review modes:

Do not apply full review to everything. That destroys the speed benefit. Do not automate every review either. That creates accountability without understanding.

Managers own the pace

Individual practitioners own the work they approve. Managers own the system in which that work is produced.

That includes workload, review capacity, incentives, escalation paths, and the pressure created by targets. A manager cannot tell the team to use AI, multiply the delivery goal, and then treat fatigue as a personal resilience problem.

Effective managers make four decisions explicit.

What stops

Every meaningful productivity gain should fund a stop-doing decision. Remove a report. Retire a ceremony. Reduce work in progress. End a low-value feature line.

If nothing stops, AI adoption becomes scope inflation.

What receives the saved capacity

Reinvest deliberately. The return may go to faster customer learning, reliability work, technical debt, or a shorter queue. It should not disappear into an unbounded expectation to do more.

What pace is sustainable

Set limits on concurrent bets, after-hours agent supervision, review queues, and incident load. A team that can sprint at a new speed for two weeks has not proven it can operate at that speed for a year.

Who can say no

High-agency teams need the authority to reject low-value generated work and challenge targets that exceed review capacity. Otherwise agency is only permission to take on more responsibility.

Protect the social system

Agent-assisted work can become isolating. People spend more time in private loops with a model and less time seeing how colleagues frame problems, debug decisions, or develop judgement.

That loss matters. Apprenticeship is built from observation. Taste develops through critique. Shared standards emerge when people examine the same work and explain why one option is better.

Preserve deliberate human contact:

• Pair on unfamiliar or high-risk work
• Hold short critique sessions around consequential decisions
• Review failures together, including agent traces and human assumptions
• Rotate ownership so knowledge does not collapse into one person and their private agent context

The goal is not more meetings. It is enough shared work to keep judgement, trust, and learning social.

Measure sustainable productivity

Volume metrics encourage volume. Use measures that connect capacity to value and expose the hidden work.

Track:

• Output-to-impact ratio: how much generated work produces a measured customer or business result?
• Review queue age: how long does work wait for an accountable decision?
• Rework rate: how much AI-assisted output requires material correction after review or release?
• Agent maintenance time: how much capacity goes into prompts, context, permissions, failures, and upgrades?
• Work in progress: how many active bets does each team carry?
• After-hours load: how often are people supervising or recovering agent work outside normal hours?
• Stop-doing rate: which recurring tasks, reports, or ceremonies were actually removed?

Pair these with the customer-facing measures in AI Product Metrics. A team is not productive because it shipped more AI features. It is productive when the features create value without degrading quality or consuming the people responsible for them.

The weekly capacity review

Add five questions to the weekly product review:

1. Where did output increase this week?
2. Which constraint moved as a result?
3. What is waiting for human judgement?
4. What work can stop?
5. Is the current pace repeatable for the next quarter?

These questions turn AI adoption from a tool rollout into an operating discipline.

Anti-pattern: productivity by exhaustion

The team adopts coding agents. Delivery metrics jump. Leadership doubles the roadmap because the team has demonstrated additional capacity.

Three months later, the pull request queue is longer, prototypes are being abandoned, and senior people spend most of their day reviewing generated work. The dashboard still shows more output. Customer outcomes have barely moved.

The tools worked. The work design failed.

AI should remove toil and increase ambition where the opportunity justifies it. It should also create room for deeper judgement, better recovery, and fewer low-value obligations. If every efficiency gain becomes a higher quota, the organisation has automated production and preserved the worst part of the old operating model.

The AI-native team design chapter covers the structures around this work. Every Agent Needs an Owner covers the stewardship load that must appear in the capacity plan.