Who Checks the First AI Draft?

By Leonard Lewis, Co-founder of Caydev

I spoke on the AI panel at the Spectrum conference this month, and one question came up that caught everyone’s attention. As I see it, the issue is no longer whether we use AI but how, and more importantly, who is signing off on the drafts it produces.

For most of my career, the slow part of a proposal or a report was the first draft. Someone had to gather the source material, work out a structure and write from a blank page, and by the time a draft existed, a lot of the thinking had been done along the way. That part has become cheap. With suitable notes and a few examples, an AI tool can now put a draft together for someone to work through, which means, hopefully, that more care goes into reviewing what it produced.

Where the work moved

When you write a paper yourself, the writing forces you through the material. When the draft arrives first, that effort has to go somewhere else, and most of it now goes into checking whether the draft represents the source properly, which is harder than it sounds. A paragraph can read beautifully and still leave out the one qualification the client needs to see, and a confident tone tells you nothing about whether the figure underneath it is right.

The risk I would put first is over-reliance, which in practice means people trusting an answer they have not checked, with the firm’s name on it. The version that worries me most involves no careless individual at all, just a team where everyone assumes somebody else did the checking, so the faster draft simply gets an unchecked answer to the client sooner.

Three decisions I would make before the tool goes in The time saved on assembling a first version is real, and I would like to see it spent on judgement. That only happens if the review step is designed into the work, because it will not appear on its own. Before a team starts relying on AI for drafts that leave the building, I would want three answers written down:

What information are we allowed to put into the tool? If client information is going in, I would want the firm to establish that the use is permitted before it happens, so individual staff are not making up the rule as they go.

Who checks the facts, figures and omissions against the original material? For work that matters, I would ask the person to show me the source and how they checked it.

Who signs off the finished work? If a junior analyst made a mistake in a client report, the first questions would be who checked it and who signed it off. I would apply the same discipline when the first draft came from an AI model.

Of the three, the second is the one I would watch most closely. Checking for what is missing is slower than checking what is there, and a qualification that never made it onto the page is easy to miss.

Who learns to check

There is a longer-term version of this question that matters a great deal in Cayman. I would expect the routine parts of onboarding, reconciliation and report writing to change first: checking whether documents are complete, comparing records and preparing a first draft. Those are also tasks where junior staff learn how the business works, and my concern is that we automate the task and forget to replace the learning it provided.

I would turn checking into part of the training. I would teach juniors to test the AI’s work against the underlying documents and to explain the exceptions they find, with experienced people still reviewing their judgement. That makes experience more useful, because judgement is what checks the output. The open question for every firm is how someone becomes a good senior when their first few years of work look different from ours, and I have not yet seen a complete answer to it.

One step to take back to the office

If I had to reduce all of this to one practical step, I would pick one use of AI in the business, name the person responsible for it and agree how the work gets checked before anyone relies on it. It is a short conversation, and a far easier one to have now than after a client spots an obvious error in the work they received.

Every AI draft that leaves an office goes out under the firm’s name, and somebody decided it was good enough to send. No one will accept “the AI got it wrong” as an excuse when a mistake goes out, so the person who signs it off owns it, as surely as if they had written it themselves.

Leonard Lewis is the co-founder of Caydev, a Cayman Islands AI consultancy and product studio.