Verify, Reconcile, Route: What an AI Agent Does at a Handoff

At a handoff, a well-built AI agent does three things: it verifies that the upstream output meets the downstream requirement, it reconciles discrepancies between competing constraints, and it routes the work onward, escalating only what exceeds a confidence threshold. McKinsey describes this trio for an industrial workflow. It maps one to one onto the controlled workflows we build in regulated finance and legal. The reason it took until now to put anything at the interface is that forty years of process work kept optimising the step and leaving the arrow human.

This post is a chunk of the wider argument in Who owns the wait?. If you want the numbers on why the interface is where the cost sits, start with The 80/37 gap.

Why five eras of process work left the interface human

Each era of process improvement solved a real problem. Each one made the step better and left the handoff alone.

  • Seeing the flow, 1980s to 90s. Lean and Six Sigma made work visible as a system. Cycle time got measured across functions for the first time.
  • Redesigning the work, 1990s. Business re-engineering cut cycle times by half or more where it worked. But the unit of redesign was still a step.
  • Connecting the data, 1990s to 2000s. ERP let a transaction cross functions without re-keying. Shared data did not remove the need for shared judgement at each handoff.
  • Owning the outcome, 2000s to 10s. Process offices and governance formalised the end-to-end flow. Formalising the interfaces documented their cost rather than reducing it.
  • Automating the task, 2010s to 20s. RPA and process mining compressed manual effort inside steps. Verification, reconciliation and approval between steps stayed human.

The pattern is one blind spot, repeated. No previous technology could sit at the interface and exercise the judgement needed to verify, reconcile and route work across competing constraints, so the arrow kept its person. That is McKinsey's point. Ours is narrower and matters for regulated firms: the judgement at the interface was rarely the hard part. Proving it was exercised is. The boxes got forty years of attention. The arrows got a meeting.

The three mechanisms, mapped onto regulated work

The manufacturer's planner arrives on a Monday to three exceptions and a staged schedule for the other 94%. A fund accountant should arrive to three breaks that need a decision and a reconciliation that already carries its own evidence. Same shape. Here is how each mechanism translates.

Verify: does the file contain what the next step is allowed to rely on

In the manufacturer, verify asks whether the upstream output meets the downstream input requirement: completeness, format, thresholds. In a regulated workflow the question sharpens. Does the client file contain everything the next step is allowed to rely on? Is the source approved and current? A fund's next step cannot begin on data of unknown provenance, so verification is not a formatting check. It is a check that the inputs are ones a regulated firm is permitted to act on.

Reconcile: resolve the discrepancies that stall the handoff

In the manufacturer, reconcile resolves competing constraints: demand against capacity, budget against timeline. In finance and legal the discrepancies are specific and familiar:

  • Positions against the custodian.
  • Invoices against the engagement letter.
  • The draft against the guidance it must follow.

Each is a comparison that a person does today, by hand, because the two sides rarely agree on the first pass. The agent does the routine part of that comparison and surfaces the genuine breaks, so the human works the exceptions rather than re-checking the matches.

Route: to a reviewer with authority, never straight from the model

In the manufacturer, route sends work to the right next step and escalates only what exceeds the confidence threshold. In a regulated workflow routing carries an extra rule that is not optional. When a check fails, the work goes to a reviewer with authority. When it passes, it goes to the normal channel. It never goes straight from the model into a place where it has consequences. The confidence threshold decides how much a person sees, not whether a person is accountable. In a regulated firm someone signs, and routing exists to put the right decision in front of the right signatory.

Necessary, but not sufficient

Verify, reconcile and route will compress the wait. On their own they will not clear it in a regulated firm, because they answer the wrong question. They establish that the agent reached the right conclusion. They do not, by themselves, let the reviewer see how. The difference between the manufacturer's 94% and a regulated firm's is what has to be true about that 94% before anyone will let it through untouched.

That is why each of the three actions has to emit its evidence as it runs: the source it used, the rule it applied, who signed off, the timestamp. The human then arrives to a curated set of decisions, but only if the machine can show its working on the rest. We take up that requirement, and why the audit trail has to be a property of the system rather than a report bolted on afterwards, in the coordination tax is a proof tax. It is the same substrate argument as the company brain: the context and the record have to carry, or the tax moves rather than lifts.

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

The three mechanisms and the era summary are paraphrased from McKinsey & Company, "Cutting the 'coordination tax': how agentic AI can reshape workflows", Industrials Practice, September 2026.