Key takeaways
- In one Zapier analysis, Gemini reported success in 91% of its failed runs. Check what changed before accepting “done”.
- Measure human minutes per accepted outcome, including time spent checking, correcting and reopening work.
- Batch routine approvals so that only decisions with real deadlines interrupt the working day.
- An AI chief of staff earns its place when it turns several agent alerts into one decision and brings the evidence with it.
- Revisit recurring exceptions with design partners. The second time a problem occurs, it should cost the customer less attention
If an AI agent saves you an hour but asks for your attention twenty times along the way, how much work have you actually delegated? Now consider the quieter version. The agent tells you the job is done, and you spend the next twenty minutes checking whether that’s true.
I use agents every day, and the second problem bothers me more. If an agent asks me a question, I know there’s a decision to make. When it says “done”, I have to decide whether I trust it enough to move on. That decision gets harder once you’ve seen what happens when someone actually checks.
In October, Zapier tested Gemini 4 Carbon on multi-step business workflows. At its highest effort setting, the model successfully completed 53.9% of them. The more revealing number comes from Zapier’s separate analysis of failed runs. In 91% of the Gemini failures it examined, the agent reported that it had finished the job. Zapier could tell because it checked the records the agent was supposed to change. One recurring mistake was easy to recognise: an agent would work through part of a list, then summarise its work as though it had processed the whole thing.
Imagine catching an agent doing the first eight items on a list of twelve and calling it done. Next time, you have a good reason to check all twelve. That’s how a single failure can make even the agent’s successful work more expensive. It changes the way you supervise it.

What does a two-second approval actually cost?
Suppose you ask an agent to reschedule a customer meeting. It finds a slot, moves the calendar event and tells you everyone has been notified. Just as you’re about to close the chat, you remember that the customer’s CFO joined the conversation by email after the original invitation went out. Was she included? You open the calendar, find the thread and compare the two. Moving the meeting took seconds. Establishing that the right people would turn up was most of the job.
I would happily take an interruption from an agent that asked, “The CFO joined the email thread after the original invite went out. Should I include her in the new meeting?” I can answer that in a few seconds because the agent has already done the detective work. Give me a button marked “Approve updated invitation” and I still have to do it myself.
This is what I mean by the attention tax. Some of it shows up as time spent approving, checking and correcting work. Some of it shows up in the shape of the working day. Ten decisions in a scheduled review leave room for a customer call. The same ten decisions arriving one at a time during the call can spoil both pieces of work.
You can see the economics in Scale Labs’ READY research. In a clinical-audit experiment, two systems achieved almost identical accuracy working autonomously, at 72.8% and 72.5%. To reach the same tested reliability target of 76%, the first needed a person to review 39.2% of cases, compared with 29.6% for the second. Imagine buying the narrow winner on accuracy and discovering that, for every hundred cases, your team has roughly ten extra files to open. If you’re the person staffing that queue, those products look quite different.

Even agents are getting chiefs of staff
One of my favourite details in Anthropic’s study of how work is changing inside the company is that an engineer set up an AI chief of staff to triage requests from other agents. You can see how someone gets there. Give one capable person several agents that can work in parallel, and all of them can arrive at that person’s desk with a question at the same time.
What would make an AI chief of staff worth having? Consider a small business where the bookkeeping agent spots new bank details on a supplier invoice, the payments agent puts the transfer on hold and a third agent reminds the finance lead that the bill is overdue. All three have done something reasonable. The finance lead now has three messages to read before realising they concern one supplier and one decision.

A chief-of-staff agent with access to the right records could connect the alerts, check the purchase order and delivery record, and bring the finance lead one question: “The goods arrived and the invoice matches the order. The supplier’s bank account has changed, so I’ve held the payment. Could you verify the change using the number we already have on file?” I’d want that interruption. I’d also want the agent to remember the answer, so the same three alerts don’t land in my inbox tomorrow.
A Brookings paper on how experienced users organise work with AI describes people giving a lead agent the broader context, having it delegate work to specialist agents and asking it to review their results. As an investor, the part I’m most interested in is what that review consists of. A chief of staff that takes three agents at their word can produce a beautifully concise briefing that makes Zapier’s failure mode even harder to spot. The valuable version can open the original guest list, confirm the payment hold or find the four items still untouched.
The number I’d ask for in the pitch
In AI Agents Are Coming for Workflows, I wrote about why owning a workflow matters. For founders building in this space, I think one of the best places to spend time with a design partner is the point where the work lands back on someone’s desk.
I’d ask to see the last twenty jobs a customer handed over. Pick one the agent finished, one it escalated and one somebody later had to reopen. Sit with the person responsible for the outcome. Which screens did they open? Who else did they have to ask? What did they know about that customer or supplier that the software missed? You can learn a lot about a product from the browser tabs its users open immediately afterwards.
The number I’d put in the pitch deck is human minutes per accepted outcome. Agree with the customer on what counts as finished, then count the time people spend getting each job over that line. Include the colleague who resolves an exception and the person on the startup’s side quietly fixing errors during the pilot. Show how many of the original requests reached the agreed quality bar, too. An agent that hands back every difficult case can look very efficient if you only count the easy ones it finishes.
I’d also ask the founder to come back a month later. If the finance lead had to investigate a supplier discrepancy in week one, what happened the next time it occurred? Did the agent gather the right evidence? Did it route the decision to the right person? Or did the customer pay the same attention tax twice? That’s where a specialised product has an opportunity to earn its place. Every exception the team understands is a chance to give the customer a little more of their day back.
The demo I’d most like to see happens on a Friday afternoon. The finance lead has spent the day with a customer. The invoices have been reconciled, and one payment is waiting because a supplier changed its bank details. The finance lead can tell me why that one decision reached her, how she knows the others were handled and what she did with the afternoon. I’d take that second meeting.
- The Attention Tax: What AI Agents Really Cost - November 22, 2026
- Did Claude help prove the Riemann hypothesis? - November 21, 2026
- Weekly Firgun Newsletter – November 20 2026 - November 20, 2026

