How to Teach AI Your Standards Without Making Everyone Work Like You
How to Teach AI Your Standards Without Making Everyone Work Like You
Say your team is preparing a recommendation for a client who wants to know whether to switch shipping carriers. An AI agent writes the first draft, and the comments start.
The founder says it's too long. They want one page they can walk the client through.
The analyst says it needs the rate comparison and the delivery data, because that's what makes the recommendation believable.
The account lead wants both. They know this client will ask where the savings figure came from.
None of them is wrong. They're asking the same document to do two different jobs.
This is where teams get into trouble with AI feedback. Someone corrects one draft, and the correction quietly becomes the rule for every draft after it. Or everyone leaves comments, the draft gets fixed, and nothing the team learned survives past that document.
Sort the comments before they become lessons
Don't flatten all of that into "make future work shorter." Sort each comment into one of three kinds first.
A fix for this draft
"The savings figure doesn't have a source."
"The draft gives the wrong date for when the new rates start."
"The opening doesn't answer what the client asked, which was whether to switch before their busy season."
These belong to the document in front of you. Someone repairs the draft, rechecks the facts and decides whether it's ready to send. One wrong date doesn't call for a new rule about dates.
One person's preference
"When I review a recommendation, put the conclusion first."
"I want a short summary before the detailed analysis."
"Show me the source next to the claim that depends on it."
These are worth keeping. With the conclusion on page one, the founder can walk the client through it without hunting for the answer. But that shouldn't cost the analyst the rate comparison, or quietly become how everyone on the team has to write.
A standard the team shares
"Every recommendation we send a client names a source for each factual claim."
"Never present an estimate as a confirmed number."
"A client recommendation has a named owner before it goes out."
These belong to the whole team, and each one needs someone who owns it. Say the account lead owns the first, since the client calls them when a number looks wrong. It covers the recommendations the founder never reads, too.
Let the disagreement improve the assignment
The founder and the analyst don't need to pick one format and decide the other person is being difficult. Their disagreement can produce a better assignment.
The recommendation opens with a one-page decision brief: switch or stay, what the switch saves and what could go wrong. Behind it sits an appendix with the rate comparison, the delivery data and the assumptions under the savings estimate. The account lead can walk the client through page one while the analyst stands behind every number. And when the client asks where the savings figure came from, someone turns to the appendix instead of promising to follow up.
Nobody had to win that argument, and the team now knows what the document owes each person who reads it. Write that down as the rule: client recommendations open with a one-page decision brief, and when the conclusion depends on factual claims, an evidence appendix goes with it. Then name the exceptions. A short internal update probably doesn't need an appendix. If the recommendation depends on what the client's current contract says about leaving early, that part may need a lawyer to look at it, whichever format it's in.
Be careful what an old example teaches
Say last quarter's carrier recommendation for a different client went over well. It's useful evidence. It shows how much detail the team found helpful, how the headings worked and how the appendix was laid out.
It doesn't tell you what this client needs, though. Before you hand it over as a model, remember what it was for. Was it written for an executive who wanted an answer in five minutes, or a logistics manager who wanted every route? Which parts did that client actually approve, and which did they put up with because the deadline was close?
Use the old recommendation for what it can teach, which is mostly presentation. Keep its client's rates, its delivery assumptions and its conclusion out of the new work. If those carry across, the new client gets advice built on someone else's situation, and nobody decided that should happen.
Make a standard on purpose
Now say the founder has written "too long" on three drafts in a row. Does that make brevity a team standard?
Not by itself. It makes it a strong preference from someone whose opinion carries weight. It becomes a standard when the team agrees it should apply to work the founder will never read, writes it down and puts a name next to it.
That's more work than letting it settle in as a pattern in someone's comments. The payoff is that a written standard can be argued with. The analyst can point out that the client asked three questions in the last meeting that only the appendix could answer, and the owner can decide whether the rule changes. A habit nobody wrote down is harder to question, because there's nothing to point to.
Two of the guidelines for human-AI interaction that Microsoft researchers published in 2019 are useful here. One says an AI system should personalize a person's experience by learning from that person's actions over time. Another says it should limit disruptive changes when it updates how it behaves. The researchers wrote them for people designing AI products, testing them with 49 design practitioners against 20 popular AI products, not for a team reviewing client work. For the carrier team, the first looks like the founder getting the conclusion on page one, and the second like the appendix staying put after the founder writes "too long" a third time.
Where Supanova fits
Say the carrier recommendation was drafted by atoms in your Supanova workforce. Finished work comes back for your review, and the founder, the analyst and the account lead all read it.
Your workforce also learns from what you tell it. When you correct an atom in chat, that learning carries into later conversations in your workspace, not just the one you're in. That's what you want when the correction is "every claim needs a source."
The founder's "too long" carries the same way, so work out what it meant before anyone types it to an atom. If it meant one page for Thursday's meeting, say that. And keep the sourcing rule where the team keeps its standards, with the account lead's name next to it, so anyone can find it.
Try it on your next review
Pick one piece of work your team argued about. Put every comment into one of three piles: fix this draft, help this reviewer, or change the shared standard.
For anything in the third pile, ask who owns it and what would make the team change it later. For a preference, keep the person's name on it. For a fix, finish the repair before you decide it taught you anything.
People don't need to work alike for AI to support them well. The founder can keep reading page one and the analyst can keep checking the appendix, as long as the team knows which rules it actually agreed to.
You can run that first review with your own workforce at supanova.team.