Operating Notes
A trace is a record, not a check
This week the Financial Times reported on an experiment KPMG UK calls Project Exponential. More than 100 AI agents worked through every phase of a parallel audit of an unnamed FTSE 250 company, while the regular team did the formal audit. According to the FT report, as carried by The Irish Times, Cath Burnet, KPMG’s UK head of audit, said the agents sometimes went “so much deeper” than the human team and asked “better questions.”
In one area, an agent failed to match a set of regulations to the company’s documents. KPMG also said its system gives “full visibility and full traceability,” recording agent interactions with documents down to individual characters and timestamps.
KPMG says the trace was complete. The matching still failed, and a person had to step in. The reports don’t say how the failure was found, and the trace may well be how the reviewer found it. KPMG’s check was people reviewing the agents’ work, and that review was planned from the start. My question is where to point them.
Much of the on-topic discussion in the Hacker News thread asked whether this says more about AI or about big-firm audit staffing. One commenter made the point I’d build on: working through the checklist takes the time; the skill is knowing what belongs on it.
Where matching goes wrong
A trace tells you what the agent did. It doesn’t tell you whether the agent was right. You can replay every tool call and still miss that rule 4(b) was never matched, because the agent never listed it.
In this kind of step, the agent reads a set of rules, reads a set of documents, and claims that each rule is satisfied by something in the documents. That claim can fail in two ways. The agent can cite evidence that doesn’t say what it claims. Or it can skip a rule entirely, and nothing in the trace looks wrong, because nothing happened. I don’t know which of these happened at KPMG.
The second failure is harder to see. A reviewer gets a tidy list of matches, and a missing row looks like nothing.
A worked example
Here is a hypothetical from airline operations, a domain I know. An agent reviews a crew hotel contract against the hotel requirements in a pilots’ union agreement. The agreement has, say, fourteen requirements: distance from the airport, quiet rooms, food available at odd hours, transport times. The agent reports the contract compliant, citing clauses for twelve.
Without a fixed checklist, a reviewer on the crew hotel team skims twelve good-looking matches and approves. The two missing requirements appear nowhere.
I’d build it differently.
First, the requirement list comes from somewhere other than the agent that answers it. Someone extracts the fourteen requirements and a person approves the list. It is checked once, against the agreement’s own section numbering, and re-checked when the agreement is amended.
Second, every claimed match must quote the contract text it relied on, and a small piece of code confirms the quote exists in the document after normalizing whitespace. That catches invented quotes only. An existing clause quoted accurately can still fail the requirement, such as a 30-minute transport time against a 15-minute limit.
Third, any requirement with no match, or with a match the agent marks as partial, goes to a person with both texts side by side.
Fourth, a person samples the matches the agent marks as full, or reviews all of them for the requirements that matter most.
The reviewer still reads what the agent said, but spends most of the time on what it couldn’t show.
My own research comes at this from the other end. A paper I’ve submitted to arXiv, which isn’t public yet, marks a tool-using agent’s work complete only when the systems it changed agree. So far it has been tested on a synthetic pilot. The idea is the same: a task isn’t done until the world has changed, and a review isn’t done until every rule has a row. Often the check doesn’t need a model at all, which is why the LLM call you can delete is worth looking for first.
What to ask this week
If your team uses agents to check documents against rules, ask where the list of rules comes from. If the agent writes it, a missing rule leaves no mark in the trace, and nobody is prompted to look for it.