What Happens to Deep Work When AI Produces the Drafts

AI makes finance work quick to produce and slower to understand. When models, memos and board decks arrive as finished-looking drafts, the working day moves from building things to reviewing them, and reviewing uses attention differently. Each draft that comes back makes the reviewer rebuild enough of the problem to judge it. The sustained concentration in which people notice what is wrong (usually called deep work, or flow) gets broken into pieces.

The practical answer for a firm is to choose on purpose which work to automate. Hand over the extraction, formatting and assembly. Keep people inside the parts of the work where building the thing is how they form a view: working through a model, writing the explanation of a variance, deciding what a board pack should say. And stop producing material nobody uses, since making it faster does nothing to make it useful.

Why the working day turns into review

Once a first pass costs almost nothing, people start far more things. Another analysis, another version of the deck, another cut of the numbers. Each one looks promising enough to finish, and every one of them comes back to the same person for a decision.

What builds up is a longer queue of nearly useful material. There is always something waiting: a draft to read, an analysis to check, a revised version to compare against the last. Output rises, and so does the time spent inspecting output, though only the first shows up anywhere. The waiting between people already limits how far faster individuals can move company output. This is the same constraint, sitting inside one person’s day.

Polish makes it worse. A well-structured document used to be evidence that someone had worked through the argument inside it. Now it can arrive before anyone has. Producing it got cheaper, and the work of establishing whether it deserves attention moved to whoever reads it. A confident answer can be well formed and still wrong about the business, so a finished-looking draft needs the reader to ask what should have been there, which is harder than checking what is.

“That’s how stuff sticks out to me”

In a conversation about automating his financial models, one client drew the line precisely. He wanted the data extraction automated. Pulling numbers out of documents and into a usable structure was an obvious place to save time. He also wanted to keep working through the model himself.

“That’s how stuff sticks out to me,” he said. Putting the numbers into a familiar structure was how he spotted changes, inconsistencies and the assumptions he wanted to challenge. He expected automation to help, and to catch things he might miss. He still wanted to take part in the thinking.

Writing works the same way. You start with a rough explanation of what happened and try to support it. Then two pieces of evidence disagree, and the explanation has to change. The finished document is one output of that process. The author’s improved understanding is another, and it exists only because the author got stuck somewhere. Often you don’t know which question matters until you reach that point.

Review asks for a different kind of attention. Holding a problem in your head and following a thread builds on itself over an hour or two. Reviewing several parallel pieces of work means switching in and out, and every switch costs the time it takes to reconstruct the context. With enough things moving at once, reconstruction takes up much of the day. People describe finishing a day having moved a great deal forward and feeling none of the satisfaction that used to come with it, because most of the day went on checking.

The same pattern at the portfolio company

This is easy to picture in a deal team, with an associate reviewing AI-drafted IC memos. It matters at least as much at the operating level.

A monthly reporting pack is the clearest example. The finance team that assembles one by hand is slow, but assembly is also where someone notices that gross margin in one region moved against the trend, or that a cost line was coded differently this month. Automate the whole pack end to end and it arrives on time, formatted, with commentary attached. Nobody inside the business has handled the numbers, so the first person to question them properly may be a board member three weeks later.

Board decks and KPI reviews follow the same logic. When a controller walks through the variance bridge before the review, part of the value is the bridge itself. The rest is that the controller now understands why the number moved and can answer the follow-up question in the room. If the bridge and the narrative both come from a model, the controller becomes its reviewer, and the meeting gets a slide that nobody present actually worked out.

What to automate and what to keep

The useful split runs through individual tasks, inside each role.

Extraction, re-keying, reconciliation and formatting are safe to hand over. They take hours, they rarely teach anyone anything, and mistakes in them can be checked against a source. A serial brokerage acquirer had analysts transposing target financials by hand from more than twenty carrier and agency formats, at a cost of days per deal. Extracting and normalizing those formats automatically made deals about three times faster from financials to a first view. The part that went was the part that took days and told nobody anything about the target.

Keep people in the steps where doing the work is how a view gets formed: working through the model on a new deal, writing the explanation of a variance, choosing what the board needs to hear this month. AI still helps inside those steps, checking arithmetic, flagging an anomaly or finding the comparable deal from two years ago, while the person builds the argument.

Then look at what the new capacity is being spent on. If ten versions of an analysis barely improve the decision compared with two, the analysis was doing less than anyone assumed. A report earns its place when someone uses it to understand a problem or act on it. Where nobody does, producing it in thirty seconds is a small win, and dropping it altogether is a bigger one.

Some work is worth keeping for other reasons. It keeps people sharp, or it is the part of the job they enjoy, and both are legitimate. A firm that ignores them will find its best people doing more review and less of the work that made them good in the first place. Having a capability puts no obligation on anyone to use it everywhere.

Firms that handle this well use AI to remove the work nobody needed to do by hand, and give people back uninterrupted time on the problems that need them. The relationship side of the same question is covered in what AI can’t automate in dealmaking.