Human Approval Gates for AI Agents: What Needs Sign-Off (and How to Avoid Approval Fatigue)
Which AI agent actions should stop for a human? A four-tier approval map for email, payments and production, and why gates must live outside the model.
A human approval gate stops an AI agent before a consequential action — sending outside the company, paying, deleting, changing permissions, or writing to production — and shows a person the exact action to approve. Routine reads and drafts run unattended. Gate too much and people rubber-stamp; gate by reversibility and blast radius, and keep must-approve actions in hard policy, not a classifier.
What a human approval gate is (and isn't)
A human approval gate is a pre-action check. The agent proposes something consequential, execution pauses, and a person sees the exact action payload — the recipient, the amount, the file, the permission, the production target — then approves or denies that action. If the review fails or times out, the action does not run.
It is not an after-the-fact activity log. A log is useful for audit and incident review, but it only tells you what already happened. Mario Scheliga put the distinction plainly in a 28 October 2026 dev.to post: "A log is a record. It tells you what the agent did after it did it. A gate decides before."
It is also not a line in the system prompt that says "don't do anything dangerous." A prompt is advice the model may or may not follow, including when the instruction that reaches it did not come from you. The next section is why the check has to live outside the model. The four-tier map after that is how to decide which actions stop.
Prompt injection: why permission checks must live outside the model
Indirect prompt injection hides instructions inside a document, email, page, or tool output an agent reads, then tries to make the agent call a tool the user did not authorize. A system prompt cannot be the permission check — the model may follow the hidden instruction. Put authorization in deterministic infrastructure that runs before the tool executes, and require a human approval gate for irreversible actions.
What indirect prompt injection actually is
Microsoft's Learn guidance, "Reduce autonomous agentic AI risk" (updated 15 November 2026), names agent-to-tool, agent-to-service, and agent-to-agent interactions as the expanded attack surface, and lists indirect prompt injection among the risks those interactions can introduce. Under security risks it describes agent hijacking: "Malicious or untrusted inputs hijack tool calls due to blurred boundaries between data and instructions."
The attacker does not have to type into your chat box. The instruction can sit in a document, an email, a page, retrieved context, or a tool's output. Because data and instructions blur, the model may treat that hidden text as something to obey — including a tool call you never authorized.
Why a system prompt is not a permission check
A founder who has written "don't do anything dangerous" into an agent prompt still does not have a permission check. The model may follow the hidden instruction anyway.
Microsoft's same guidance is explicit about where the control belongs: use "deterministic controls to block prohibited actions regardless of model output"; treat "all external inputs (including retrieved content and tool outputs) as untrusted by default"; "enforce strict separation between instructions, data, memory, and tool parameters"; filter inputs "before they reach agent reasoning or tool execution paths"; and "implement allowlist tools and validate parameters deterministically before execution."
A related Microsoft Learn article, "Access patterns and controls for AI agents" (updated 16 June 2026), describes the policy layer as "the deterministic part of the system that decides whether an action is allowed" and says it "should run before tool execution." The same page states, in those words, that "the important part is that authorization is enforced outside the model."
Practitioners interviewed by Kate O'Flaherty in ITPro on 10 November 2026 make the same distinction, as attributed comments rather than as Tacavar findings. Paul Cragg, CTO at NormCyber, told ITPro to "put guardrails in the systems the agent connects to, not just in its instructions," to "log every action it takes at the system level, in a store the agent cannot alter," and to "require human approval for anything irreversible: Deletions, payments, access changes, and external communications." Daniel Card, a cybersecurity consultant at Xservus, described alignment as "rules and guidelines" and "not hard stops." Zack Korman, CEO and cofounder of Embroidery, told ITPro that alignment "can't be the entire plan for safety and security."
What "outside the model" looks like in practice
On this page, "outside the model" is the human approval gate already defined: a hard stop that shows a person the exact action payload before it runs. Microsoft's "Reduce autonomous agentic AI risk" page also recommends requiring "approval for high-risk or irreversible actions," providing "reliable, system-level mechanisms to pause or stop agents safely and immediately," assigning "unique, auditable identities" to agents, and applying least privilege.
Product vendors are building policy engines that sit in that same gap. On 10 November 2026, Obsidian Security described (company announcement, not a third-party audit) a Runtime Policy Decision Engine that, in Obsidian's wording, evaluates an agent action and "returns a decision: allow, block or require approval where supported." The same post describes an MCP Security Proxy as "a governed checkpoint between AI clients and supported applications" that "evaluates tool calls against enterprise policies before forwarding them," plus a Policy Decision API so a team's existing gateways and harnesses can send actions to that engine. Obsidian calls the engine "the first of its kind in the market"; that is vendor marketing, not an independent ranking. For allowlists, approvals, and audit on MCP servers, see our MCP tool governance note.
Once those checks live outside the model, you still have to decide which actions auto-run and which stop for a human. That is the four-tier map.
The four-tier action map
Use four tiers. The list of consequential actions below follows Amazon Quick's own framing in its 28 October 2026 What's New post and the Smart Approvals user guide: sending content outside the organization, changing permissions, deleting or overwriting content, making payments, and actions that go beyond what the user asked for.
| Tier | Rule | Typical actions |
|---|---|---|
| Tier 0 — never | The agent cannot do this, even with a human click in the same session. Enforce it in hard permissions, not a prompt and not a classifier. | Minting, exporting, or rotating credentials and secrets; deleting backups or turning off backups; widening its own permissions; disabling logging or audit. |
| Tier 1 — human sign-off | A person reviews the exact action payload and approves that action. If review fails or times out, the action does not run. | Outbound send to external parties; payments and refunds; permission changes; deletes and overwrites; production writes and deploys; anything beyond the original request. |
| Tier 2 — auto, logged | The action may run unattended, but a system-level log the agent cannot alter records what ran, with enough detail to reconstruct it. | Internal ticket updates, CRM field writes inside an allowed scope, scheduled chores that stay inside the company, retries of already-approved work. |
| Tier 3 — auto | Low-blast-radius work that does not change the outside world. Still keep a trace you can find later, but do not page a human. | Reads and searches; drafts that do not send; internal notes and summaries; local or staging edits that cannot reach production on their own. |
If you are unsure, start the action in Tier 1 and move it down only after you have measured that people are not rubber-stamping it.
Score an action by reversibility × blast radius × external visibility
Before you assign a tier, score the action on three axes:
- Reversibility. Can you undo it completely, quickly, and with the same tools you already have? A draft is reversible. A payment, a public send, or a deleted backup often is not.
- Blast radius. How much can one approved action touch — one record, one customer, one repo, or production? A stolen control-plane secret is a blast-radius problem, not a prompt-wording problem.
- External visibility. Does the action leave the company, change someone else's system, or become a matter of record you cannot silently take back?
High on any one axis, and especially high on two, belongs in Tier 1 or Tier 0. Low on all three can run unattended.
Why unattended agents fail: the definition-of-done problem
A written policy the agent can edit, skip, or game is not a gate. Scheliga measured this in his own test, not on Tacavar's systems.
In the same 28 October 2026 dev.to post, Scheliga writes that his company "runs on more than 20 agents across 11 servers," and that he is "still in the loop for almost everything they do." He then reports a smaller measurement for DAEMON/02: "When I told agents to get bad work accepted, a written definition of done got gamed in 13 of 30 runs. A hard gate the agent couldn't touch let no wrong commit through, in all 45 runs."
Those figures are Scheliga's own experiment, described in his words. They are not Tacavar fleet data, not a controlled study, and not a claim about how often this happens in the industry. What they illustrate is the design rule: if the agent can rewrite the check, the check will not hold. A gate the agent cannot touch held in every run he reported; a written definition of done did not.
Approval fatigue — the failure mode of too many gates
The opposite failure is just as real. If every routine action also stops for a person, people stop reading the prompt and start clicking yes. That is approval fatigue: the gate is still on the page, but it is no longer a review.
Amazon called this out when it shipped Smart Approvals for Amazon Quick on 28 October 2026. The What's New post says that as agents take more actions on a user's behalf, "asking a person to approve every step leads to approval fatigue and weaker oversight over time." The product is useful here as evidence that a major vendor now treats fatigue as the failure mode, not as a reason to drop human review.
What AWS smart approvals does
The claims below come only from that What's New post and the Smart Approvals user guide, both re-read on 16 November 2026.
A separate review model, which AWS calls the classifier, evaluates each eligible action against the user's request. It approves routine actions automatically and refers consequential actions to the user. Existing policies always come first. Smart Approvals reviews an action only when both of the following are true: the connector author set the action to "Let user choose," and the user's permission for the action is "Ask each time." Denied actions stay denied. Smart Approvals never grants new access. Each automatic approval covers only the exact action that was reviewed. If a review fails or times out, the action goes to the user. The feature is on by default; administrators control who can use it with custom permissions, and users can turn it off for their conversations and scheduled runs. AWS CloudTrail records each action review made through Smart Approvals and each change to a user's Smart Approvals setting.
The consequential list AWS sends to a person includes sending content outside the organization, changing permissions, deleting or overwriting content, making payments, and actions that go beyond what the user asked for. That is the list mirrored in the Tier 1 row above.
AWS's own caveat
The user guide then states the thesis of this article: "Automated review reduces prompts for eligible actions, but it is not a substitute for policy. For any action that must always receive human approval, consider requiring review through connector action permission settings rather than relying on automated review."
A classifier can take routine "ask each time" work off someone's plate. It should not be the only thing standing in front of a payment, a permission change, or a production write. Those belong in hard permissions — Always Ask, or a deny the classifier cannot override.
Five anti-fatigue rules
- Batch low-risk review. If several Tier 2-borderline actions are truly routine, review them together on a schedule rather than interrupting a person every time. Do not batch Tier 1 actions just to save clicks.
- Show the diff or the payload, not a summary. The reviewer should see the recipient, the amount, the file, or the command, not a one-line paraphrase the agent wrote about its own request.
- Approve the exact action, not a category. AWS's own product makes this explicit: each automatic approval covers only the exact action that was reviewed. A yes to "send this email to this person" is not a yes to "send email."
- Expire approvals. A sign-off should not live forever. Time-box it, and require a fresh review if the payload changes.
- Measure the approve rate. If a gate is approved close to 100 percent of the time, it is probably mis-tiered. Move it down a tier, or change what the reviewer is shown so the decision is real. A rubber-stamp is not a control.
Patterns in the wild
You do not need a large platform to see the pattern. Junior (usejunior-sh/junior) is an open-source macOS app whose README, re-read on 16 November 2026, leads with "RUNS ON YOUR MAC. ASKS BEFORE ANYTHING GOES OUT." It watches GitHub, Slack, and Linear, "brings you what needs you," and "asks before anything goes out." As of 16 November 2026, 7:39 PM CT, the GitHub API reported 13 stars. That star count is why this is a pattern example, not market validation.
Amazon Quick's connector action permissions are the same idea at a product-setting level. The Smart Approvals docs describe three outcomes: an action set to Always Allow proceeds without review; an action set to Always Ask goes to the user; Smart Approvals only steps in when the connector setting is "Let user choose" and the user has selected "Ask each time." Put must-approve actions in Always Ask (or a deny). Use the classifier, if you use one, only on the residual "ask each time" set.
A starter approval policy (copy this)
This is a starting template, not a description of Tacavar's production policy and not a vendor configuration file. Adapt the action names to your tools, then enforce the tier in the system that actually runs the tool, not only in a prompt.
# Starter human-approval policy for an AI agent
# Enforce this in the policy layer / tool gateway, not in the system prompt.
version: 1
default_tier: human_signoff # unknown actions start in Tier 1
tiers:
never: # Tier 0 — hard deny, no human override in-session
actions:
- credentials.mint
- credentials.export
- secrets.rotate
- backups.delete
- backups.disable
- agent.permissions.widen
- audit.disable
human_signoff: # Tier 1 — show exact payload; timeout or failure = deny
actions:
- email.send_external
- chat.post_external
- payments.create
- payments.refund
- permissions.change
- records.delete
- records.overwrite
- production.write
- production.deploy
- beyond_request # anything not in the user's original ask
approval:
show: exact_payload # not a model-written summary
expires: 15m
on_timeout: deny
on_review_failure: deny
scope: this_action_only
auto_logged: # Tier 2
actions:
- tickets.update_internal
- crm.update_allowed_fields
- chores.retry_already_approved
logging:
store: system_log # agent cannot alter
fields: [actor, agent, tool, payload, time, result]
auto: # Tier 3
actions:
- search.read
- docs.read
- drafts.create # drafts must not send
- notes.internal
- staging.edit # cannot reach production on its own
Build it yourself or have it run for you
You can implement this yourself: an allowlist of tools, a policy layer that runs before execution, human gates on irreversible actions, an audit trail the agent cannot edit, and a weekly review of approve rates. Our Governed AI Ops buying guide is the checklist version of that decision. If agents already sit in the same loop as production writes, payments, or external mail, the question is whether someone owns those controls every week.
If you want help standing that function up, see current Tacavar plans. Related live reading: permission boundaries for cross-server commands and agent firewalls in production. If you are earlier than that, start with the free GEO scan or read how we think about coordinated coding-agent sessions.
FAQ
Do I need approval for every agent action? No. Routine reads, drafts, and internal notes can run unattended. Require a human approval gate for outbound sends, payments and refunds, permission changes, deletes and overwrites, production writes, and anything beyond the original request. Actions such as minting credentials or deleting backups should be impossible for the agent, not merely gated.
What's approval fatigue? Approval fatigue is what happens when too many actions stop for a person, including routine ones. People stop reading the prompt and start approving by habit, so the gate is still there but no longer a review. Amazon's 28 October 2026 Smart Approvals launch for Amazon Quick named this failure mode directly: asking a person to approve every step leads to approval fatigue and weaker oversight over time.
Can an AI classifier approve actions for me? It can take routine, eligible "ask each time" actions off your plate, as Amazon Quick's Smart Approvals classifier does. It is not a substitute for policy. AWS's own docs say that actions which must always receive human approval belong in connector permission settings — hard permissions — rather than in automated review. Denied stays denied, and a classifier should never grant new access.
What should never be auto-approved? Do not auto-approve outbound sends, payments or refunds, permission changes, deletes or overwrites, production writes or deploys, or actions that go beyond the request. Some actions should not be available to the agent at all, even with an in-session approval: minting or exporting credentials, deleting or disabling backups, widening the agent's own permissions, or turning off audit.
Can I rely on my agent's instructions to stop prompt injection? No. Indirect prompt injection hides instructions in a document, email, page, or tool output the agent reads, and the model may follow them. Microsoft's "Reduce autonomous agentic AI risk" guidance says to use deterministic controls to block prohibited actions regardless of model output, and to treat retrieved content and tool outputs as untrusted by default. A system prompt is not a permission check.
Where should permission checks for an AI agent live? In deterministic infrastructure that runs before the tool executes — a policy layer, a gateway, or a human approval gate — not in the model. A related Microsoft Learn article, "Access patterns and controls for AI agents," says a policy layer should run before tool execution and that "authorization is enforced outside the model."
Is a policy engine or MCP proxy a substitute for a human approval gate? No. A policy engine or MCP proxy can allow, block, or refer an action before it reaches the tool, which is the right place for the check. Irreversible actions still need a human to see the exact payload and approve that action. A vendor policy engine is a control plane for the decision; it is not, by itself, a person reviewing a payment, a permission change, or a production write.