Is AI Giving Your Team Time Back, or Giving It More Work to Check?
Is AI Giving Your Team Time Back, or Giving It More Work to Check?
Say you lead a small account team. Every Monday, each account owner gets a short briefing on their biggest customers: what changed last week, which support issues are still open, where the renewal stands and what to raise on the next call. Someone on the team pulls it together from the CRM, the support queue and last week's call notes, and it takes most of the morning.
So you hand the first draft to AI, and it comes back in minutes. Then the account owner reads it. The renewal date was copied from an old note. The pipeline isn't split by region, which she asked for last week. One recommendation offers the customer a discount nobody has approved. She fixes all three before her first call, and the next Monday she fixes two of them again.
The draft got faster. That doesn't tell you whether the team got time back. For that, you need to know whether people now spend less time briefing, checking, correcting and passing the work along before a usable briefing reaches her.
If the checking grows as fast as the output, you haven't found capacity. You've moved the work.
Why it can feel faster than it is
Her Monday isn't unusual. BetterUp and the Stanford Social Media Lab surveyed full-time U.S. desk workers about what they call "workslop": AI-generated work that "looks polished and complete" but turns out to be off the mark or no help to whoever receives it. Forty percent of the workers believed they had received some in the last month. They reported spending almost two hours dealing with each instance, about 20 minutes longer "than if the sender had done the work themselves."
That's easy to miss because the time doesn't show up where the draft was made. The teammate who ran it watched the briefing appear in a few minutes. The account owner spent her Monday morning fixing it.
People also misjudge their own speed. In METR's study of experienced open-source developers, tasks where AI was allowed took 19% longer, and afterward the developers still believed AI had sped them up by 20%. METR has since said those results no longer reflect what AI does for developers, and that developers likely got more speed from AI tools by early 2026 than its early-2025 estimates showed. What still holds is the gap between what was measured and what people felt: METR's later update warns that self-reported speedups "can be quite unreliable." So count the hours the briefing takes instead of asking how it felt.
Count the work around the work
Start by writing down who spends time on the briefing each week, before the AI draft and after it. The teammate pulls from the CRM, the support queue and last week's call notes, explains what the account owner wants and checks what comes back. The account owner reads it, fixes what's wrong and decides what to raise on her calls.
Keep the record light. You need a fair comparison, not a time study that turns into a second project.
| Step | Before AI | After AI | What to look for |
|---|---|---|---|
| Brief | How often does someone restate the goal, the reader, the sources and what to leave out? | Does the work start with the standards and context from last time? | Explaining the same thing every week is work. |
| Draft | Who gathers, compares and writes? | What arrives already prepared? | More output doesn't mean the job is done. |
| Review | Who checks the facts, the tone, what's missing and what a recommendation would commit the team to? | Which of those checks still need a person? | Keep the checks that change the decision. |
| Repair | How many rounds of questions or rework happen? | What sends the work back? | Watch for the same fix two weeks running. |
| Handoff | Who works out what happened and what comes next? | Can the account owner see the result, the evidence and the open question? | A clean handoff saves the owner's attention. |
Run the same briefing for a few weeks. Each week, note how long each person spent on it, how many rounds of review it took, whether the account owner could use what she got and what still slowed her down.
Keep the review that changes the result
Some review is the work. When the account owner decides whether a recommendation suits this customer right now, she's using judgment the briefing can't supply, and it's the reason she reads it at all. A finance lead approving an exception is doing the same thing. So is anyone checking an email to a customer before it goes out.
Other review is a sign the assignment is missing something. She rereads the same rules every week, hunts for the current renewal date, works out where a number came from and fixes the regional split again. None of that makes the briefing better. It repairs work that should have arrived ready, and it's the time you're trying to get back.
OpenAI Academy's workflow design coach puts this sorting before anything gets built. The spec it helps you write makes the human judgment, the boundaries and the points where work should be escalated visible "before developing the solution." For the Monday briefing, that means writing down which calls belong to the account owner before deciding what the AI prepares.
The account owner doesn't need to watch the briefing take shape. She needs it finished by Monday morning so she can spend her review on the recommendation. In Supanova, you can say "do this every Monday" and it happens every Monday, and finished work comes back for your review.
Turn a repeated correction into a standard
The quickest way to build a checking queue is to have people correct the same thing every week without giving the next run a better start.
When the account owner writes "split the pipeline by region" on one briefing, it might be a one-off note. When she writes it three Mondays in a row, it's a standard nobody has written down. Someone should decide where it belongs: in the brief, in an approved source, in the checklist the reviewer uses or in the template the briefing starts from.
It also matters who the standard is for. The regional split may be right for the account team and wrong for Finance, which reports the same pipeline by product line. If she marks one briefing down for the missing split, that mark shouldn't quietly become a rule for the whole company. In Supanova, Finance can be walled off from Sales, including what each department's atoms are allowed to learn from the other, so what the account team's atoms learn about the regional split can stay on the Sales side.
After each review, ask four questions:
- Did the account owner make a business decision, or fix a mistake that could have been prevented?
- What context or standard would have saved that fix next week?
- Does the correction belong to this briefing, this team or the whole company?
- Should the briefing keep running as it is, run with a smaller scope, or stop until the missing piece is in place?
Decide what the workflow deserves next
After a month of Mondays, read the worksheet and choose one of four outcomes.
- Keep it if the briefing is usable, the review that's left is the account owner's real judgment, and nobody is re-explaining the same standards or fixing the same mistakes.
- Narrow it if the AI does one part of the job well, such as summarizing open support issues, but the full briefing keeps sending work back.
- Redesign it if the same correction keeps coming back, like the regional split. Change the brief, the sources or the template before the next Monday.
- Stop it if checking the briefing costs more than the briefing is worth, or if it keeps offering customers terms only the account owner can approve, like that discount nobody signed off.
A single score won't make that choice for you. Anthropic's guide to evaluating AI agents says "no single evaluation layer catches every issue." Alongside automated evals, it describes watching the live system, collecting user feedback, reading through the agent's conversations and having trained raters grade the work. Your worksheet is a smaller version of the same habit: the hours, the rounds of review and the account owner's own view of whether the briefing helped.
If the briefing runs on Supanova, Learning Signal gives you the wider view: across your whole workspace, how often your workforce gets it right the first time, this month versus last. For the Monday briefing itself, whether it comes back with fewer fixes and how much of the account owner's morning it still takes, your worksheet is the record.
Where the time goes next
The account owner still reads the briefing every Monday. What changes is that she stops fixing the same three things, and her morning goes to the customers.
We built Supanova so people could turn ideas into economic value without waiting for capital, headcount or permission.
In this example, the idea is a growth plan for the team's biggest account. Everyone agrees it's worth writing, and nobody has had a free Monday morning to start it. Once the fixes stop, someone does.
Pick one recurring assignment. Run the worksheet for a month, then let what you counted decide what the workflow deserves next.
Explore Supanova at supanova.team.
Sources
- The hidden cost of AI "workslop" and how leaders can fix it, BetterUp (2025)
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR (2025)
- We are Changing our Developer Productivity Experiment Design, METR (2026)
- AI workflow design coach, OpenAI Academy (2026)
- Demystifying evals for AI agents, Anthropic (2026)