Your company bought an AI agent to eliminate busywork. Excellent news. It attended the meeting, transcribed the meeting, summarised the meeting, extracted seventeen action items, generated a slide deck about the meeting, and scheduled a follow-up meeting to align on the action items from the meeting.
The work is now faster and, somehow, there is more of it.
This is not because AI cannot be productive. It can. It is because productivity is usually measured at the easiest place: the task. Organisations, unfortunately, happen after the task.
Faster is real.
In a large field study of 5,172 customer-support agents, access to a generative-AI assistant increased issues resolved per hour by about fifteen percent on average. The largest gains went to newer and less-skilled workers.1
That is useful evidence. It suggests AI can compress the time required to retrieve knowledge, phrase an answer and learn the patterns of experienced colleagues.
But notice what the study measured: a bounded job with a legible output. Customer issue, agent response, resolution.
Most knowledge work is less polite. The output of one person becomes the input of five others. A memo creates comments. Comments create revisions. Revisions create approvals. Approvals create a dashboard so leadership can monitor the revisions to the comments on the memo.
When producing an object becomes cheap, consuming that object does not automatically become free.
Cheap output creates expensive attention.
Suppose a weekly report once took four hours, so the team produced one. Now an agent can produce twelve before lunch: one by region, one by product, one by risk category, and nine because somebody discovered the dropdown.
The production cost collapsed. The verification cost did not. Someone still has to ask whether the source was correct, whether two charts contradict each other, whether the confident paragraph means anything, and whether the report should have existed.
This is the trap: AI does not merely remove labour. It changes which kinds of output are cheap enough to demand.
Email made messages cheaper; we did not respond by enjoying less email. PowerPoint made slides cheaper; humanity remained resilient and produced more slides. Generative AI makes competent-looking cognitive artefacts almost frictionless.
Friction was wasteful. It was also, occasionally, the only thing preventing Kevin from commissioning a forty-page strategy document before breakfast.
The distinction. Task efficiency asks how quickly one object is produced. System effectiveness asks whether producing it improved the condition the organisation exists to change.
Your org chart became software.
In 1968, computer scientist Melvin Conway observed that organisations tend to design systems resembling their own communication structures.2 Four teams make a compiler; the compiler mysteriously has four major parts. A department has six approval layers; the software acquires six permission states and a button marked “escalate.”
ReGrowth extends this into a deliberately coined idea: Conway Debt.
Technical debt is old code still charging rent. Conway Debt is an old organisational arrangement still executing after the reason for it has disappeared.
The monthly report survives the manager who requested it. The approval survives the scandal that created it. The data field survives the programme it once measured. Nobody owns the whole chain, so everybody carefully completes their part.
Then the AI agent arrives. It does not see “obsolete ritual.” It sees a workflow with excellent documentation. So it learns the ritual, accelerates the ritual and offers to personalise the ritual at scale.
AI is an accelerator, and an audit.
This produces a useful double effect. AI can make bureaucracy cheaper. But because it can read across documents, tickets, policies and messages, it can also make bureaucracy visible.
Why are three teams calculating the same number differently? Why does this approval loop return to the person who initiated it? Why are employees translating data from one internal vocabulary into another every Thursday?
ReGrowth calls this AI is the Audit: not because a chatbot is an omniscient auditor, but because automation forces hidden rules to become explicit enough to execute.
Once the rule becomes explicit, a more important question is difficult to avoid: what outcome does this rule protect?
Sometimes there is a good answer: safety, accountability, due process, fraud prevention. Bureaucracy is not a synonym for stupidity. A hospital should not move fast and break people. But sometimes the process protects only the continued existence of the process.
Some workflows need archaeology.
Before giving a workflow an agent, give it four things.
An outcome. What condition in the world should improve?
An owner. Who can judge the whole result rather than merely complete one step?
A mandate. What is the process allowed to change, and what must it preserve?
A stopping rule. Under what condition does this report, approval or dashboard stop existing?
If nobody can answer those questions, the workflow is not ready for autonomous execution. It may be ready for archaeology.
This is why the interesting AI transformation is not simply that every employee gets a copilot. That can make every department faster at being exactly what it already is.
The deeper opportunity is to discover which tasks create an outcome, which tasks protect a legitimate constraint, and which tasks exist mainly to feed other tasks. Automate the first. Respect the second. Delete the third, with adult supervision and a backup.
The question is not whether your agent can do the work. It is whether the work should survive its new efficiency.