The AI Company Would Like to See Your D+ Work

There was a time when a polished five-page proposal was at least some indication that the person who wrote it had spent time thinking through the argument. AI has made that assumption a lot less reliable.

Scott Stevenson, cofounder and CEO of legal AI company Spellbook, recently updated the company’s internal AI policy after noticing that employee proposals were arriving at his desk looking considerably more developed than the ideas underneath them. His preference now is for employees to show more of the unfinished thinking. In his words, he would rather see the “D+ version.”

Spellbook, which sells AI software for drafting and reviewing legal contracts, has told its 200-plus employees that AI should not be used to write the main body of project proposals. Employees can still use it for supporting work, including things like tables and spreadsheets, but the central argument is supposed to come from the person making it. The company is also asking employees to disclose when internal messages or documents were drafted with AI, while customer communications are expected to be written by humans.

On its face, none of this is especially radical. A manager wanting to know what an employee actually thinks is a fairly ordinary expectation. What makes it interesting is that ordinary expectation has now become something companies increasingly have to protect through policy.

For the past few years, the corporate message around AI has largely moved in one direction: use it more. Use it to move faster, improve your writing, generate the first draft, summarize the meeting, refine the proposal and turn a few rough thoughts into something polished enough to send upstairs.

Employees responded accordingly.

The problem is that AI is exceptionally good at improving the appearance of an idea without necessarily improving the idea itself. Three mediocre bullet points can become five pages of confident prose, complete with an executive summary, recommendations and the reassuring tone of something that appears to have survived several rounds of review.

The document gets better. The thinking does not always come with it.

That is what makes Spellbook’s policy more interesting than a simple story about an AI company restricting AI. Stevenson is not arguing that the technology is useless. His company exists because he believes the opposite. The distinction appears to be between using AI to help perform a defined task and using it to manufacture the appearance of finished thinking.

That distinction may become increasingly important in the workplace.

For years, polish functioned as a rough proxy for effort. A detailed deck suggested someone had spent time building the deck. A carefully written proposal implied that someone had wrestled with the argument. Even a well-constructed email carried some evidence that a person had stopped long enough to decide what they wanted to say.

AI is breaking that relationship.

The professional-looking document is rapidly becoming the cheap part. What is harder to determine is whether there was a useful idea before the document existed.

That creates an unusual problem for managers. As AI makes it easier for employees to produce work that looks complete, companies may need to become more interested in the messier evidence that thinking actually occurred: the rough notes, the incomplete argument, the uncomfortable question, the half-formed idea that has not yet been transformed into six neatly formatted sections.

In other words, one of the more valuable signals at work may soon be the thing employees have spent years learning to hide.

The unfinished version.

Spellbook’s response is essentially to ask employees to show some of their work before the machine makes it presentable. Not because polish is bad, but because polish has become too easy to produce on its own.

Corporate AI spent years promising to make everyone’s work look better.

Now management needs to see the part before that happened.