The Productivity Paradox: Why Your AI Gains Aren't Showing Up Across the Organisation

In 1987 the economist Robert Solow described a puzzle that has never quite gone away: the computer age was visible everywhere except in the productivity statistics. Firms were investing heavily in computing, yet output barely moved. Economists called it the productivity paradox. Enterprise AI is now running into its own version of it.

Individually, people are faster. Engineers write code faster, analysts draft faster, operations triage exceptions faster. The gains are easy to feel. Their effect on the organisation as a whole is far harder to see. A 2025 MIT study of hundreds of corporate AI efforts found that about 95% delivered no measurable return.

Inside regulated firms the pattern is familiar. Releases ship on the same cadence, regulatory submissions take the same number of weeks, the change advisory board meets just as often, and the backlog keeps growing. Every individual now has tools to work faster. The organisation moves at the pace it always did.

The spending tells the other half of the story. Enterprises are paying for AI, vendors are selling productivity and capacity, and consultants are selling the transformation programme on top. The one question that decides whether it paid off, where is the return, is rarely answered with evidence.

Where the gains go

The gains themselves are real. In a controlled study, developers using GitHub Copilot completed a well-scoped coding task 55% faster than those without it. Field results are more modest, but they point the same way. The problem is not the size of the individual gain. It is that the gain rarely survives the journey from one person to the whole organisation.

Work in a regulated enterprise is not one person at a keyboard. It is many teams doing many different things, handing work to one another in sequence.

It starts well before code. Someone gathers the requirements, scopes the feature, runs a gap analysis to check whether they are even the right requirements. Then a developer commits. The change waits for peer review, then security review, then the architecture forum that meets on Tuesdays. The ticket waits for the advisory board. The board waits for the release window. The window waits for the next quarterly freeze to lift. The developer's link in that chain is now five times faster. The chain is exactly as long as it was. And after all of it, the requirements often weren't quite right to begin with.

You sped up one link. The chain did not get shorter.

The gains are lost in the gaps between people, not inside any one person's work. They leak away in four places:

• Hand-off friction. Compressing the time inside a step does nothing to the time between steps. A feature built in a day still waits weeks for change advisory, security and a regulated release slot. Speed up one step of a twenty-step process and the end-to-end gain is a rounding error.
• Coordination debt. Individual speed creates a surge of output nobody downstream is staffed to absorb. An analyst team producing three times the credit memos does not make the credit committee meet more often, so the cadence is throttled and the gain leaks away.
• No institutional memory. A developer asks a chatbot how to refactor an unfamiliar piece of code. The answer helps for ten minutes, then vanishes: no record, nothing the next person can reuse, nothing that tells risk which model touched which decision. The gain is personal, not institutional.
• Shallow adoption. Most organisations are experimenting at the edges, not running AI across the operation. The Accenture and Carnegie Mellon SEI AI adoption maturity model places most enterprises in its early stages, well short of AI scaled across the enterprise, and names workflow re-engineering as one of the eight capabilities that scaling depends on. Pilots at the surface do not move the organisation.

None of this is a failure of the people. Engineers in regulated firms are excellent, often more rigorous than their startup peers. The bottleneck is the wiring between them: separation of duties the regulator expects, change gates designed to give risk time to object, vendor cycles measured in quarters, teams split across time zones. Much of it exists for good reason, and most of it was built before AI could do what it now can. You cannot reorganise your way out of structures a regulator requires.

Redesigning the process around AI

So the gap does not close by making each person faster inside a process built for humans handing work to humans. It closes by redesigning the process itself: coordinated agent teams that span the steps people currently hand between, with humans approving wherever judgement is required. The work runs in parallel rather than in series, under control, with the feedback loop measured in hours instead of weeks.

The shape is visible in practice. At a tier-1 bank, a legacy data-pipeline estate of several thousand jobs written in one language had to be rebuilt on a framework that spoke another, far too many to convert by hand. A coordinated team of agents ran the work end to end, discovery, conversion and validation, inside the bank's perimeter with every step on an audit trail. Production bundles came back verified in around three minutes each, every one compiling cleanly under the bank's own toolchain. The process stopped moving at the speed of its slowest hand-off and started moving at the speed of the work.

How Promenaut closes the loop

Promenaut is the enterprise agentic workforce platform: it maps, designs and runs a hybrid workforce of people and agents, inside the enterprise perimeter and model-agnostic, with full audit and control built in. Three things make the difference:

• Agents span the hand-off, not one role. A coordinated agent team, developer, architect, tester, security, compliance, works across the whole lifecycle, so code arrives with its review, test and audit artefacts already attached instead of queuing for them.
• Control is in the work, not bolted on after. Every action passes through approval points and fail-closed gates with a complete audit trail. This is agentic oversight: we govern what agents do at the workflow level, not which model they run on.
• The capability compounds. Because every action is logged and attributed, the work builds institutional capability rather than vanishing into personal toolkits.

None of the controls go away. Agents act, humans approve, the platform logs; separation of duties holds and the regulator still gets its evidence. What goes away is the latency between signal and improvement, not the controls between intent and action.

The question is not whether your people can be faster. It is whether the organisation can be. That changes when you redesign the loop, not just the links.

See the loop running inside an organisation like yours.