In Regulated Work, the Coordination Tax Is a Proof Tax

In regulated finance and legal, the coordination tax is a proof tax. In an industrial case the cost of a handoff is throughput and working capital. In ours it is mostly proof. A handoff in a regulated firm does not stall because nobody can make the judgement. It stalls because nobody can evidence that the upstream step was done correctly, by an authorised person, against an approved source, at a known time. The reviewer re-performs the work not from distrust, but because the record does not let them do anything else.

That single distinction changes what an AI agent has to do at the interface. This post is a chunk of the wider field note, Who owns the wait?. For the shape of the tax and the numbers behind it, start with The 80/37 gap.

The reviewer re-performs because the record leaves no choice

Picture where the proof tax shows up. Reconciliation breaks re-checked by hand. Client letters re-read in full because the draft has no provenance. Onboarding files re-verified at every hop. Matters that wait for a partner who is really waiting for a paper trail. In every case a person is redoing work that was already done, not because they doubt the judgement but because the judgement arrived without its evidence.

What it costs is the same three things McKinsey found in the industrial case. Dedicated checkers you can see. Cycle time you feel. And revenue that never arrived because the mandate, the close or the filing came too late. The proof tax is not a compliance overhead sitting to one side of the business. It is the coordination tax, wearing the specific clothes a regulated firm makes it wear.

Most AI in regulated firms dies at exactly this point: where someone has to sign for it. The model produced a good answer, but the person who has to attest cannot see how it got there, so they redo it. The fix is not more review. It is a workflow that produces the thing the reviewer needed to see, at execution time, every time.

The agent has to emit its evidence as it runs

Verify, reconcile and route are the right mechanisms at a handoff, and we set out how they map onto regulated work in Verify, reconcile, route. They are necessary and not sufficient. Each action has to emit its evidence as it runs. For every step the record should carry:

The source it relied on, and confirmation that the source was approved and current.
The rule it applied.
Who signed off.
The timestamp.

Emit those as the work happens and the human arrives to a curated set of decisions with the working already shown. Leave them out and you have moved the coordination tax, not removed it. The wait simply relocates from the doing to the signing.

Audit trail as a property of the system, not a report bolted on

This is a substrate question, not a reporting one. If the audit trail is a report someone assembles afterwards, the handoff still waits for it, because the evidence is being reconstructed after the fact. If the audit trail is a property of the system that executed the step, the next step can start immediately, because the proof was produced alongside the work.

Design for verification, not generation, and make the verification write its own record. Thresholds, accountability and the boundary between human and machine belong in the system that executes the step, not in a policy document that describes it. That is the same substrate argument as the company brain: the context and the record have to carry, or every hop pays the tax again.

Start with the 65%, and start with one interface

The proof tax is real, but it does not have to be paid off all at once. When McKinsey's team took one interface apart, the work inside it split three ways:

65% routine verification against known rules, thresholds and evidence. Ship this first.
25% constrained agent judgement within limits.
10% people.

McKinsey quotes the full redesign at six to twelve months, with 70% of the effort on people and process, sequenced as a COO-sponsored programme. We agree that most of the effort is organisational: who owns the interface, what the thresholds are, who may change them. That 70% on people and process is our forward-deployed model, and nobody buys a platform and gets it. Where we differ is speed. The 65% is routine verification against known rules. It can be built, evaluated and running on one interface in weeks, with the record attached.

The gate is what makes weeks safe. Before the routine 65% runs untouched, it runs in parallel against the process it replaces, on real volume, until the difference is explainable line by line. Only then does the threshold move. You are not asking anyone to trust the machine. You are showing them the machine agreeing with the current process until the disagreements are understood, and then handing it the routine share.

Why it compounds across interfaces

Start with one interface for the same reason it makes sense to build any substrate once. The connectors, definitions, permissions and evidence model from the first interface are reused by the second. Each build costs less than the last. Their programme redesigns the whole workflow at once. Ours redesigns one interface, leaves the proof behind, and moves to the next. Every quarter of waiting is a quarter of coordination data a competitor is accumulating, so the record compounds while the wait does not.

Which handoff in your firm waits longest? We map the interfaces in a fortnight, build the first one inside your stack, and prove it in parallel before it runs alone. Download the field note, or tell us where the work waits and we will tell you whether it is a proof problem.

Figures, the work split and the industrial example are from McKinsey & Company, "Cutting the 'coordination tax': how agentic AI can reshape workflows", Industrials Practice, September 2026.