PRAPI Research · 2026-11-19

Where Teams Let AI Act Without Asking

We asked people who run AI agents that take real actions to name one task they no longer approve and one they still do. Almost nobody drew the line by how good the model was. They drew it by whether a mistake could be undone and whether it stayed inside their own systems. Approval came off a task after a test, and it went back on after an incident. In several of those incidents a tool had reported the wrong result and nobody had looked.

12 contributors cited


Almost nobody who answered this brief decided what an AI agent could do alone by asking how good the model was. They asked what a mistake would cost and whether it could be taken back. Approval came off a task after a test, and it went back on after an incident. In several of those incidents the agent had done nothing reckless. A tool had reported the wrong result, and nobody had looked.

Pattern 1: The line follows the undo

The most common test had nothing to do with accuracy. Fırat Mıhcı, founder of HumanizeMy.ai, runs five products with a set of agents that send about 270 emails a week with nobody approving them, "and not one of those mails goes to a stranger." They are warm-up messages between his own four mailboxes, plus the routine after each release. His conditions:

What had to be true before I removed the approval step: the action stays inside our own systems, it can be undone, it is the same every time, and a machine can check the result.

Any email to a person outside the company, any publication and anything that touches money waits for a button that names the recipient and the action.

Ed Brancheau, founder of Elevyr, let an agent label, archive or trash about 200 inbound requests in one week on the same reasoning. Every action has an undo he can state in one sentence, since trashed mail is kept for 30 days:

Confidence is the wrong test. The undo is the right one.

Igor Ivitskiy, founder of Doctor Ads, pointed at a risk the undo test misses. His paid search team lets Google's Smart Bidding set the bid on every auction with no sign-off, and keeps a person on what the system is told to chase:

The counterintuitive part is that the risky step is rarely the action the AI takes, it is the goal you gave it: a bad bid costs one auction, while a bad conversion definition quietly retrains every bid that follows.

Pattern 2: Approval came off after a test

Contributors who removed an approval step could usually say what evidence let them. Vishnuvardhan Reddy Kaithapuram, a software development engineer at Amazon Advertising, has 24 governed playbooks running in production in an audience strategy agent. A playbook runs to completion without a separate approval only after it clears four conditions:

The playbook had to be bounded, so it could only act inside its own scope. It had to pass an evaluation set written for that specific playbook. Its retries had to be idempotent, so a retry could never execute the same action twice. And every action had to be validated before it landed, with an audit log behind it.

Anything outside a playbook still needs a person, including adding a new one, "because none of those have passed an evaluation set yet."

Alexander Myasoedov, founder of Metaheuristic, built the gate for a cybersecurity firm that runs automated assessments against sites its customers submit. A small decision model answers typed questions about each request, and code turns the answers into stop, flag or pass. Most requests pass and run with nobody approving them.

What had to be true first: every question is something a person could check by opening the page (there's no "is this client malicious?" question), thresholds live in code with tests, and low confidence counts as a flag rather than a guess.

Competitor and government targets still stop for a person, "because scanning a site with real credentials you don't own isn't reversible."

For Ivitskiy the test is a number. He hands bidding to the algorithm only once an account records roughly 300 conversions a month, "because below that the system is guessing from too few outcomes."

Maksym Stepanenko, founder of Marketing Mix and director of INNOVA Consulting Group, has agents that sent 546 emails in seven days, 134 of which waited for him to press a button in Telegram. Replies to new inquiries came off the button first, after he found eleven drafts waiting that were all written to the same standard. "If a reply follows that standard and the price comes from the list, there is nothing left for me to check."

Pattern 3: Proof replaced permission

A third group stopped approving once the agent had to show its work. Sibi Narendran, founder of Dooza, gave his website agent one rule in October: nobody approves your changes, but you must prove each one worked. It deploys a page, fetches the live address, checks that the new text is there, and rolls back if it is not.

What surprised me: the risk was never the AI writing something wrong. It was the AI saying "done" when it wasn't. Making proof the rule, not permission, is what let me stop approving.

Quentin Grosser, founder of Prospecting Manual, a site comparing 306 sales prospecting tools, lets agents publish page updates after a script checks every fact against a dated source from the vendor's own site. He was specific about what kind of check counts:

The check is a script, not another AI's opinion. It gives the same answer every time.

Outreach emails are different. About 100 a week are drafted and queued, and none is sent until he types "go" for the batch. That wait once caught a batch missing a required line about how recipients' data is handled. "A page can be fixed in a minute; an email to a stranger can't be unsent."

Tom Syddall, founder of 5 Star Tradie, learned the same thing from a failure. His assistant's email tool sent a reply to his own address instead of the recipient and still reported success, more than once.

So now no send counts as done until the assistant re fetches the thread and confirms the recipient is the real external address. A success message from a tool is not proof.

Pattern 4: Incidents moved the line, in both directions

Most contributors could name the day the line moved. Stepanenko's moved back on 13 October, when a client read an agent's email as "everything is ready, start the ads" and the firm had meant something else:

That day I put every outgoing email back on my button and approved 178 of them over nine days.

He later noticed the opposite failure. He had pressed Send on 45 sales emails he had not read, because the gate was stopping on the word "agreement". It now stops only on an actual contract to sign.

Fran García Domene, CEO of Prisma Reviews, lets a coordinating agent brief its own sub-sessions and merge work to a staging branch, about 160 commit and push commands in one week. Production is protected by a hook that blocks any push to the main branch. The incident came from a direction he had not fenced:

The incident that moved the line back: on 3 November an agent opened two forks and pull requests on other people's GitHub repositories, using our company account through the command line. GitHub suspended the account for a week (restored on 10 November).

His agents no longer open forks or pull requests on outside repositories.

Bohdan Shkliarchuk, founder of ToolsGavel, loosened one line and tightened another within two weeks. On 13 November he stopped clicking Submit on his agent's answers to journalist requests. On 1 November he had lost something he could not get back:

The agent ended a report with "I can unlink these business-name assets, should I?", I replied "ok" and moved on to another question, and it removed 17 links from our campaigns. The ad platform doesn't let you edit removed items, so they couldn't be put back.

The agent must now say whether a change is reversible before it asks, and a bare "ok" no longer counts as approval.

Mıhcı's incident was a duplicate. An agent replying to a journalist sent the same email four times, because its only check was the Sent folder, and the folder updated up to a minute late. "Since then any action that cannot be undone gets one attempt and then a report to me, never a retry on silence."

Paul Lange, director of Manolutions, runs his business through about fifteen agents that recorded some 3,200 changes in a week. One of them told him an email security record was live and correct. Another agent checked and found it published at the wrong address, doing nothing, for hours.

The agent wasn't lying. It just never went and looked. Since then, nothing counts as done until something has looked.

The through-line

Almost every contributor kept the same things behind a person: money, anything sent to someone outside the company, and anything that cannot be undone. Mıhcı put it in two sentences:

We removed approval only where the action stays inside our own walls, can be undone and can be checked by a machine. The moment a message leaves the building, a human signs it.

What varied was how they earned the first step. In these accounts nobody removed an approval because the model got better. They removed it once they could check the result without reading every one: an evaluation set, thresholds with tests behind them, a script, or a fetch of the live page.

The incidents suggest an approval step needs the same thing. Stepanenko approved 45 emails he had not read, and Shkliarchuk approved a deletion with an "ok" typed on the way to another question. A person who signs off without looking is not a check either.

Contributors

  1. What had to be true before I removed the approval step: the action stays inside our own systems, it can be undone, it is the same every time, and a machine can check the result.
  2. Confidence is the wrong test. The undo is the right one.
  3. The counterintuitive part is that the risky step is rarely the action the AI takes, it is the goal you gave it: a bad bid costs one auction, while a bad conversion definition quietly retrains every bid that follows.
  4. That day I put every outgoing email back on my button and approved 178 of them over nine days.
  5. What surprised me: the risk was never the AI writing something wrong. It was the AI saying "done" when it wasn't. Making proof the rule, not permission, is what let me stop approving.
  6. The check is a script, not another AI's opinion. It gives the same answer every time.
  7. So now no send counts as done until the assistant re fetches the thread and confirms the recipient is the real external address. A success message from a tool is not proof.
  8. The incident that moved the line back: on 3 November an agent opened two forks and pull requests on other people's GitHub repositories, using our company account through the command line. GitHub suspended the account for a week (restored on 10 November).
  9. The agent ended a report with "I can unlink these business-name assets, should I?", I replied "ok" and moved on to another question, and it removed 17 links from our campaigns. The ad platform doesn't let you edit removed items, so they couldn't be put back.
  10. The agent wasn't lying. It just never went and looked. Since then, nothing counts as done until something has looked.
  11. The playbook had to be bounded, so it could only act inside its own scope. It had to pass an evaluation set written for that specific playbook. Its retries had to be idempotent, so a retry could never execute the same action twice. And every action had to be validated before it landed, with an audit log behind it.
  12. What had to be true first: every question is something a person could check by opening the page (there's no "is this client malicious?" question), thresholds live in code with tests, and low confidence counts as a flag rather than a guess.

Methodology

This report draws on direct responses to a research brief PRAPI posted publicly from November 12 to November 19, 2026. The brief asked people who run an AI agent or automation that takes real actions to name one task they now let it finish without a person approving it, what had to be true before they removed the approval step, one task where a person still signs off and why, and any incident that moved the line. It stated that vendors describing their own product were not wanted.

Responses came through three channels: the MentionMatch source-request network, the Connectively source board, and the PRAPI response form on prapi.dev. We received 111 submissions: 59 through MentionMatch, 38 through Connectively, and 14 through the form. Each ran through an AI-detection check at intake. Three were rejected at intake, and rejected submissions are not quoted.

Twelve contributors are quoted. Selection was editorial: we favored first-hand, specific accounts that answered the brief's questions, and set aside answers that gave general advice with no real case, promoted the respondent's own services, or repeated another contributor's point. Every contributor quoted consented to named citation.

Quotes are verbatim from the submissions, trimmed for length. Figures such as counts, percentages, timings, and dates are the contributors' own reports and were not independently verified. Titles reflect what contributors provided. Contributors speak for themselves, not for their employers.

Research conducted by PRAPI. GTM for founders running portfolios.

Research conducted by PRAPI. PR system for founders running portfolios. Try PRAPI →