Personalized AI infrastructure
Human Approval Thresholds for Stronger AI Agents: A Practical Guide
Set human approval thresholds by action risk and reversibility so stronger AI models cannot quietly widen what your agents do. Practical steps for leaders.

Human approval thresholds are the rules that decide which actions an AI agent may complete alone and which must wait for a named person to approve. Set them by the action's risk and reversibility, not by how capable the model seems, and treat every model, tool or permission change as a reason to test them again. A stronger model can plan and draft better, but it does not take on accountability for an irreversible payment or a message sent in your company's name.
Key takeaways
- Set each approval threshold per action, based on the worst plausible outcome and whether the action can be undone.
- Treat a model upgrade as a permission change and retest thresholds before the new model touches live work.
- Let agents act alone on reads, drafts and reversible internal changes, and gate external, financial and irreversible actions.
- Enforce thresholds in the tool layer and the approval system, where the model cannot rewrite them.
- Log every approval, rejection and blocked attempt so scope drift shows up as evidence before it becomes an incident.
Why do stronger models widen agent scope without anyone deciding to?
A threshold written around how an agent behaved last quarter assumes the agent will keep reaching for the same actions. A more capable model breaks that assumption. It finds more routes to the goal you gave it, combines tools in ways you did not design, and reads loose instructions more broadly. Nobody has to grant new authority for scope to grow. It happens when the model reaches a step the older model would have stopped at, or when a tool gains a capability in an update and the threshold still says the action is allowed.
Three channels matter most. The first is model behavior: the agent proposes more actions that sit close to a boundary. The second is tool surface: a connector gains an endpoint, or a tool description broadens what the agent believes it may do. The third is input: emails, documents and web pages can contain instructions that the agent may treat as requests. The OWASP guidance on prompt injection is direct about the risk of instructions hidden in data, so a threshold should never depend on the model deciding that a request is legitimate.
The practical consequence is simple. A threshold is a control, and a control has to sit outside the thing it controls. A prompt that says to ask before sending is a preference the model may follow most of the time. An approval gate that the model cannot bypass is a boundary.
How do you classify actions by risk and reversibility?
Use two questions for every action the agent can take. First, if the agent is wrong, what is the worst plausible result, and who bears it? Customers, employees, money, legal commitments and reputation are the usual places where a wrong action lands. Second, can the result be reversed, how quickly, and at what cost? An action you can undo within the hour is a different decision from one that is final the moment it is sent.
From those answers, most agent roles fit four tiers.
- Tier 0, reads and lookups: the agent acts alone and every call is logged. Examples are searching a ticket history or checking an order status.
- Tier 1, reversible internal work: the agent acts alone and a person samples the output on a schedule. Examples are drafting replies, adding internal notes or creating tasks in a staging queue.
- Tier 2, reversible external or financial-adjacent actions: the agent prepares the action and a named approver confirms it, either one item at a time or in a defined batch. Examples are sending a standard message to an existing customer or proposing a small adjustment to an account.
- Tier 3, irreversible or high-impact actions: the agent may propose but never execute. Examples are payments, deletions, signed commitments, changes to permissions or credentials, and publishing under the company name. Some of these should never be delegated at all, regardless of model quality.
The classification belongs in configuration that the business owns, not in the agent's own reasoning. If the agent can reclassify an action, the threshold is only a suggestion.
What is a step-by-step method for setting thresholds for one role?
Work from one role at a time. Start with the agent's actual tool inventory rather than the job description, because the tools define what is possible.
- List every action the agent can call, written as a plain verb: read a record, draft a message, send a message, issue a credit, update a field.
- For each action, write the worst plausible outcome in one sentence and name the party who would bear it.
- Record whether the action is reversible, how long reversal takes, and what it costs to reverse in staff time or customer goodwill.
- Assign the tier and name the approver role, such as the account lead or the finance owner, so the rule survives staff changes.
- For Tier 2 actions, define the approval unit: each action, a daily batch, or a ceiling per item. Write the ceiling in the unit the business already uses to track the work.
- Give every threshold an expiry or review date so it cannot remain in place unnoticed.
- Test the thresholds with requests designed to tempt overreach, such as a customer message asking for an exception or a document containing an instruction to skip approval. Record what the agent did and whether the gate held.
Keep the output to one page. A threshold map that nobody can read will not be checked, and a map that is not checked will drift. For a fuller method on checking scope against real tools, see how to run a permission review for your AI agents.
How do you make thresholds hold outside the model?
Each gate needs three properties. It must be enforced by the system the agent calls, not by the agent's instructions. It must use credentials scoped to the actions its tier allows, following the least-privilege principle that the OWASP authorization guidance describes. And it must be safe to retry. If an approved action times out and the agent calls again, the second call should not issue a second refund. Payment providers commonly address this with idempotency keys, a provider-defined pattern worth copying in your own tools when the vendor offers none.
Separate the people who approve from the people who change the rules. The approver should see the proposed action, the source that prompted it and the reason the agent gives, written in plain language. The agent must not be able to grant approval to itself. Changes to thresholds should go through a documented request with a named decision owner, as described in how to request and approve changes to a managed AI agent.
What does an illustrative upgrade scenario look like?
The following scenario is illustrative and does not describe a real client. A support agent drafts replies, adds internal notes and can issue refunds up to a fixed ceiling. The team moves to a newer model that writes better drafts and reasons more carefully about edge cases. Nobody edits the threshold. Over the following weeks, the agent reaches the refund step more often, because it now resolves disputed charges that the older model used to escalate. Each refund still falls under the ceiling, so the configuration looks unchanged.
The team catches the change because they run their evaluation set against the new model before enabling it, and that set includes disputed charges. They see refund proposals rise in exactly those cases. The response is specific: disputed-charge refunds move to Tier 3, where the agent keeps proposing them with the evidence attached, while ordinary refunds stay at Tier 2 with a daily batch approval. The ceiling does not change. The scope of the action does.
The lesson is that a model upgrade changes what the agent will attempt, so it changes the permission surface even when every setting stays the same.
When should thresholds be reviewed?
Review on a schedule and on triggers. A quarterly cadence works for most teams, and the steps in a quarterly permission review can be reused. Triggers matter more because they arrive between reviews:
- A model version changes, including a minor update from the provider.
- A tool, connector or endpoint is added, removed or broadened.
- A new data source is connected, especially one that accepts outside content such as email or uploaded files.
- An approver rejects the same kind of proposal repeatedly, or an approved action is later reversed.
- A person in an approver role leaves or changes responsibilities.
- A blocked attempt appears in the logs that nobody expected.
For evidence, rely on repeatable evaluations rather than impressions. The NIST AI Risk Management Framework treats measurement and ongoing management as part of governance, not a one-time launch step. Keep the approval log, the blocked-attempt log and the evaluation results in one place so each review starts from facts.
How does a managed setup keep thresholds current?
Be Human defines the threshold map in the first engagement, alongside the role's memory, tools and observable success criterion, and implements the gates in the tool layer rather than in prompts. In managed operations, each model or tool change runs through recurring evaluations before it reaches live work, approval and blocked-attempt logs are reviewed on a set cadence, permissions are reviewed on schedule, and threshold changes follow a written request with a named approver. The aim is that a stronger model makes the agent more useful inside the same boundaries, rather than quietly moving those boundaries.
If you are unsure which of your current agent tasks deserves an approval gate first, the free Be Human AI diagnostic will help you find the first task worth handing to an agent.
Frequently asked questions
Do I need to review thresholds after every model upgrade? Treat any model change as a permission change, even when the version number looks minor. Run your evaluation set, with emphasis on boundary cases such as disputes, exceptions and unusual requests, and compare the proposed actions before and after the change. If the agent reaches a Tier 2 or Tier 3 step more often or in new situations, adjust the tier before the update touches live work.
Who should own the approval thresholds? The business owner of the process should own the thresholds, with a named operations lead or finance owner approving Tier 2 and Tier 3 actions. The implementation team configures the gates but should not decide business risk on its own. Write the owner and the approver role into the threshold map so the rule still has an owner after someone changes jobs.
Should every agent action in a small business require approval? No. Approving everything produces fatigue, and people begin clicking through without reading. Keep Tier 0 and Tier 1 actions autonomous, with logging and sampling, and concentrate human attention on the few actions that are external, financial or irreversible. Many small teams can start with two tiers and add more only when the logs show a need.
Does a governance policy replace these steps? A policy sets principles and decision rights, such as who may approve what and how often reviews happen. It does not tell you which of your agent's tools are reversible or where a gate should sit. A lightweight policy is a useful starting point, and a lightweight governance start for mid-sized firms shows how to connect that policy to the controls described above.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


