AI agent approval fatigue: how many approvals is too many?

An AI agent should ask you only about actions that are risky and hard to undo: paying, posting publicly, deleting, and contacting someone new. Everything else, such as reading, research, notes, drafts and routine updates, should run on its own. If you're approving more than a handful of things a day, the fix isn't to approve faster; it's to approve by risk and reversibility instead of by action count, using trusted contacts, templates you approve once, batched approvals and a daily digest.

This guide gives you a framework for deciding what needs a person, and shows which settings in KutrimBots relax approvals safely and which never should.

Why does my AI agent ask me to approve everything?

Because many agents treat every action that changes something as equally risky, so you get the same prompt for "add a note" as for "pay ₹40,000".

That isn't just annoying; it's unsafe. Anthropic, describing its own coding agent, warned that constant approval prompts lead to "approval fatigue", where users might not pay close attention to what they're approving. By setting clear boundaries up front instead, it cut permission prompts by 84% in internal use.

The pattern is decades old in medicine. A 2006 review of 17 studies in JAMIA found that clinicians overrode computerised drug safety alerts in 49% to 96% of cases, a problem now widely known as alert fatigue. When everything is an alert, nothing is.

Is approving everything safer than approving less?

No. A pile of approvals trains you to click "allow" without reading, which removes the protection the approvals were meant to give.

Human-factors researchers call this automation bias and complacency: people come to trust an automated system and stop checking it, and a review by Parasuraman and Manzey found it affects novices and experts alike and isn't cured by simple practice. A card you approve 60 times a day becomes a reflex.

Security guidance points the same way. The 2025 OWASP Top 10 for LLM Applications lists "Excessive Agency" as risk number 6, caused by too much functionality, too many permissions or too much autonomy. Its fixes are structural: give the agent only the tools it needs, require a person to approve high-impact actions, enforce permissions outside the AI, and rate-limit what gets through. Fewer, better-placed approvals plus hard limits beat approving everything.

For the wider list of agent risks, see are AI agents safe?.

How should I decide what needs my approval?

Ask 2 questions about each action: how bad is it if it's wrong, and can you undo it? Then place it in 1 of 4 boxes.

Then apply 3 rules:

1. One approval per decision, not per step. Approve "send these 12 reminders", not each of the 12 emails, where the tool allows it.
2. Thresholds beat blanket rules. "Ask when the amount is ₹10,000 or more" removes most prompts and keeps the ones that matter.
3. Every approval must be readable in 10 seconds. Who, what, how much. If you can't decide that fast, the request is too vague.

How KutrimBots maps to the 4 boxes

These are KutrimBots' built-in rules, as of November 2026:

How do I cut approvals without losing control?

Use 5 tools: trusted contacts, templates approved once, per-action rules, batches and a daily digest.

1. Trusted contacts. With Email Inbox, where each agent has its own address such as sales@yourcompany.com, an agent's email goes straight out to people who wrote to it first and to contacts you've allowed; only the first email to someone new waits for you. New people added on a reply-all or forward wait too. Once a pattern is reliable, allow specific contacts rather than turning approval off. Our guide to the AI email assistant covers inbox setup.

2. Approve a template once. In an automation, you approve the whole workflow before it goes live, including any risky step set to run on its own, with the reason shown to you. After that, each run follows the template without asking. Built-in plugins work the same way: Invoices & Billing emails payment reminders automatically once you turn them on in its Settings, and Bookings lets you choose between auto-confirming bookings and approving each one.

3. Per-action rules. Each approval card offers Allow once, Always allow or Deny. In each plugin's rules you can also set any action that changes something to Always ask me, Never ask or the plugin default. KutrimBots enforces these rules itself, so an agent can't skip them, and reading is never held. For website sign-ins through the Security Vault, "for this sign-in" covers that site for 15 minutes, so the agent doesn't ask per field.

4. Batch approvals. Build an automation with one Approval step that lists today's items ("12 follow-up emails, drafts attached"), followed by a step that acts on each item, up to 100 at a time. Some actions are designed to ask every time, such as emailing an invoice; batching doesn't remove that.

5. A daily digest instead of pings. Let routine information arrive once a day instead of as it happens:

• the CRM daily brief: hottest leads, overdue follow-ups, today's plan;
• today's bookings and yesterday's form responses, each morning in Needs you;
• your morning plan and evening wrap-up from Tasks & Reminders;
• a monthly finance report on the 1st;
• a scheduled automation that sums up anything else and posts it to a channel.

The digest is for information; real approvals still wait in Needs you. Unanswered approvals don't run later by surprise: in automations they time out as declined after 24 hours by default.

Example scenario: from 46 approvals a day to 9

The numbers are assumptions for illustration, not customer data. A three-agent setup for a small travel agency: a sales agent, an inbox agent and a finance agent.

Those 9 approvals at about 10 seconds each take 90 seconds a day, and every one of them is worth reading.

As a rule of thumb, once the first 2 weeks are over:

If you're above these numbers, look for repeated requests you always approve unchanged; each is a candidate for a trusted contact or a template.

Which approvals should I never relax?

Payments, public posts, deletions, first contact with new people and anything involving secrets should keep asking, at least until you've watched them work for 4 weeks or more.

• Money out. Paying a bill, a refund, voiding an invoice.
• Public posts. Anything with your brand on it, on any network.
• Deleting or merging records. Lost leads should be marked lost, not deleted.
• New contacts. The first email to anyone new is where a mistake or a manipulated instruction does the most damage.
• Secrets. Card numbers, CVVs and PINs in the Security Vault are for you only; agents can't use them at all.
• Server changes. Deleting folders, rebooting, changing users or the firewall, which ask every time even on a server you've "Always allowed".

Be careful with Always allow on anything broad. Allowing one contact or one action is different from allowing everything an agent might do.

How do I tune approvals in KutrimBots?

Spend the first two weeks approving everything, then relax rules one at a time based on what you actually approved unchanged.

1. Days 1–14: keep every default. Approve carefully, and note which requests you approve without edits.
2. Allow trusted contacts for regular customers and partners whose emails you never change.
3. Turn routine sends into templates: payment reminders in Invoices settings, booking auto-confirm, and automations whose risky steps you've seen work.
4. Add thresholds and batches for bulk updates and daily follow-ups.
5. Move information to digests and turn off real-time notifications you don't act on.
6. Review every 30 days. Look at what you approved, what you denied and why. Tighten anything you denied twice.

Each agent's approval rules also apply inside automations it runs, so you don't need separate settings for each workflow. For how agents choose tools and when they pause, read how AI agents work.

When should you keep approving everything?

Keep full approvals on with a new agent, a new automation, a new channel, after any mistake, and for work in regulated areas such as health, finance or law.

The framework also has limits. A threshold set too high lets an expensive mistake through, and a template approved once will repeat its error on every run until you notice. Approvals also can't catch a wrong fact in your own notes: if your price list is wrong, the agent will politely send the wrong price. Read your digests, and keep the run logs in view for the first month.

You can try these settings on the free 7-day trial, with up to 3 agents and approvals on by default.

Frequently asked questions

How many approvals should an AI agent ask for a day?

For a small business, usually under 10 once it's set up: new contacts, payments, public posts and deletions. If you get many more, move routine work to trusted contacts, approved templates, batches and a digest.

What is approval fatigue?

It's when so many approval requests arrive that people start clicking "allow" without reading. It makes agents less safe, because the checks that matter get approved on autopilot along with everything else.

What does Allow once, Always allow and Deny mean?

Allow once approves this single action. Always allow approves this kind of action from now on, so use it for narrow, trusted actions. Deny blocks it, and a well-built agent doesn't try another way around.

Should an AI agent ever pay or delete without asking?

Not by default. In KutrimBots, agents ask before anything that pays or deletes, and those are the last approvals worth relaxing, if ever.

What is human-in-the-loop for AI agents?

A person approves selected actions before the agent carries them out. Good human-in-the-loop design places those checkpoints only where an action is high-impact or hard to undo.

Sources

• Making Claude Code more secure and autonomous with sandboxing (Anthropic)
• Overriding of drug safety alerts in computerized physician order entry (van der Sijs et al., JAMIA)
• Complacency and bias in human use of automation (Parasuraman and Manzey, Human Factors)
• LLM06:2025 Excessive Agency (OWASP Top 10 for LLM Applications)