What Is a Document Worth?

What is the value of a document?

Is it the information it holds? Is it the reasoning it carries, worth whatever the thinking behind it is worth? Or is it what the document signals - the effort, the competence, the attention of the author?

These are three reasonable answers, and for most of working life it did not matter much if you knew which was which, because you could not produce a good document without doing the thing underneath it. A tight, well-argued memo was proof that someone had read the material and thought it through. The document stood in for the work, and crafting a good one meant you had done the work. AI has broken that link.

When Freelancer.com added an AI cover-letter generator to its site, the letters got better - more tailored, more closely matched to what each job asked for. They also got less useful. Three researchers looked at what happened across more than five million letters. The correlation between a well-tailored letter and getting a callback fell by about half. The correlation with actually winning the work fell even more, by close to eighty per cent. Employers discounted the letter as evidence, whether the applicant had used AI or not, and shifted their attention to signals that were harder to fake.

Nobody ever wanted a cover letter (the applicant least of all). What an employer wanted was evidence that someone had read the role, worked out whether they fit, and cared enough to spend an hour saying so. The letter was a receipt for that time spent. Once anyone could print the receipt without spending the time, it became proof of nothing.

Your documents are also receipts

Few will mourn the loss of the cover letter. But a great deal of knowledge work is, viewed cynically, the production of documents. For professional services firms the question is even more acute, because a firm produces documents at an industrial scale.

A client cannot see most of what they pay for. They do not see the reading, the discarded drafts, the judgement call made after hours, or the twenty years it took to be able to make it. What they see is the document that arrives. The deliverable is the evidence that the work happened - and an entire commercial model is built on that substitution. In this system, hours stand in for effort, effort stands in for value, and the document is often what sits on the surface. For a long time this has held, because the proxies moved together: you could not produce the document without doing the work, so paying for one was a reasonable way of paying for the other.

That is the link the cover letter just lost. And it raises a question every firm now has to answer, deliverable by deliverable: which of your documents contained the thing the client really wanted, and which were only ever the receipt?

Not every document was doing the same job

It is easy to assume the answer is obvious - that a faster, better version of a document is a better document. But run three cases through and the assumption does not survive.

We've already seen a document that was more receipt than substance. But there are others such as the status update nobody reads, the file note that exists so the file has a note, the summary whose function was to show the matter had been attended to. Their value was the evidence of attention, and AI has made that evidence free. Improving them improves close to nothing.

Some documents carry the thing itself. If I have worked with someone for three years and formed a view of them, then hand my notes and my assessment to a machine to write up, the person I am writing for is probably better off - the letter is likely fairer, more complete, and it arrives on time rather than three weeks late because I could not face the blank page. The judgement is still mine. The writing was only ever the delivery mechanism. I am well aware that some readers will accept that and some will bristle at it, and the disagreement is itself worth noticing: when two competent people cannot agree whether applying AI to a document destroyed it, the value was never sitting in the document, and the norms around it are still being written.

To complicate matters, most documents are neither. A single advice memo can be half receipt and half the thing itself: two pages summarising the facts and the precedent, which AI now drafts well, wrapped around the one paragraph of recommendation the client is actually paying for. Automate the memo as a unit and you improve the wrapper and flatten the centre.

So the sort is real, and it is harder than it looks. Complexity, judgement and cost are all properties of the document. What the client was taking from it is not.

What the documents never contained

There is a limit to the sort, and it is worth naming plainly: everything so far assumes the documents hold what the client is paying for. Sometimes they do. But ask a general counsel why the firm has kept its place on the panel for a decade, or a business owner why they have used the same accountant for fifteen years, and the answer is rarely a document, or even an accumulation of them. It is confidence. Someone who knows the business, who they can call before a decision rather than after, who they trust to say "leave it" as readily as "act". The deliverables are artefacts of that relationship - the visible surface of something that mostly happens off the page.

That value never touches the ledger either. The read of the room, the call not to escalate, the reassurance that a steady hand has looked at the problem - nobody itemises those, and the client could not point to the line where they were charged for them. The most valuable thing a firm sells can be the thing that appears on no invoice and in no deliverable.

Which is exactly why it is exposed. An efficiency programme, pointed at the documents, inevitably reshapes the contact around them - fewer drafts passing back and forth, fewer calls to discuss them, fewer reasons to be in the room. Nothing on the ledger moves, because it wasn't there in the first place. The saving shows up as a line item, but the loss shows up years later as a client who did not come back and could not quite say why.

Software has already had this problem

None of this is a new problem. It is a new frontier of an old one, and another industry has a twenty-year head start on it.

From the early 2000s, the cost of building software collapsed - cheaper tools, then cloud computing, then open source - until teams could build almost anything, almost immediately. Shipping faster did not mean building better products. Eric Ries, in The Lean Startup - a rare book from this world that made it to the airport shelves - put the lesson plainly: the question was no longer whether a product could be built, because almost anything could be. The question was whether it should be.

The discipline that grew out of that lesson runs on two questions, asked before anything gets made: what problem are we really solving, and who are we solving it for? And it asks them at two levels, because they give different answers. A feature can pass the test while the product around it fails, and a product can be loved while half its features go unused. Nobody buys a feature. They buy the product.

That distinction maps onto a firm in a useful way. Each document can be thought of as a feature. The product is what the client is actually hiring: the outcome, and the confidence of having an adviser beside them. Some decisions need making at the document level - the advice memo with the recommendation buried in it. But the level that decides whether the client stays is the product, and a firm that interrogates every deliverable and never the engagement is optimising the features of a product it has not examined.

You do not get to opt out

Go back to the cover letter one last time, because it holds one more lesson, and it is the hard one. Abstaining did not help. Only a minority of applications ever used the generator, yet the signal collapsed for the whole market - the hand-written letters lost their evidentiary value alongside the machine-written ones, because employers could not tell which was which. A signal is not something you own. It is something a market agrees to read, and the market decides when it stops.

The same logic is now running through professional work. Whether or not your firm automates a single deliverable, other firms will, and the document as proof of effort dies for everyone. That breaks something deeper than a deliverable. For decades, time told you where the value was: hours stood in for effort, effort stood in for value, and you could read a client relationship off a ledger. That inference no longer runs. You can keep charging for time. You can no longer learn from it.

The conversation to have

Which leaves one instrument for finding out what a client is actually paying for: ask them. Not whether they mind that you use AI - that is a question about permission, and it gets a shrug. Ask what they do with each thing you send: which documents they read closely, which they skim for a number, which they would notice the absence of. Then ask the product-level question: what did you actually hire us for - and if the documents kept arriving but the people disappeared, what would you miss?

I have to do this in my own practice, on my own deliverables, and the answers are rarely the ones I expect. The document I am proudest of is sometimes the one the client barely glances at. Sometimes what they are paying for is not what you think it is - and that was always true. The difference is that the proxy was established and you never really needed to work it out, because the hours got paid anyway. Now the firms that find out will automate the work that was only ever a receipt and protect the work that was the point. The ones that don't will improve every document equally, and quietly hollow out the things that were never about the document at all.

Sources

Cui, Dias & Ye, Signaling in the Age of AI: Evidence from Cover Letters, arXiv working paper, Yale University (2025, 5.5 million cover letters across 106,714 jobs on Freelancer.com; not yet peer-reviewed)

Eric Ries, The Lean Startup, Crown Business (2011)