Approval limits: keeping humans in charge of AI agents

The first question most business owners ask about an AI agent isn't "what can it do?" It's "what will it do without asking me?" That is the right question. An agent that can issue refunds, offer discounts or raise purchase orders is useful only if you trust where its authority ends.

Approval limits are how you draw that line. They are the same idea you already use with people: a new team member can approve small things on their own and needs a manager for bigger ones. This article explains how to set those limits for an agent, who should sign off on what, and how to adjust them over time.

Why limits matter more than intelligence

It is tempting to focus on how clever an agent is. In practice, what keeps a business safe is not how well the agent reasons but how clearly its boundaries are set. A well-defined limit turns a vague worry into a simple rule: below this amount, the agent acts; above it, a person decides.

Limits also make the agent easier to explain to your team. Instead of wondering what the system might do, everyone knows which decisions it handles, which ones land in their queue and why. That clarity is often what makes a team comfortable working alongside an agent at all.

Decide which actions need a limit

Start by listing every action the agent could take that changes something real. Reading an order or checking a tracking update carries little risk. Anything that moves money, makes a promise to a customer or commits you to a supplier deserves a limit. Common examples include:

• Issuing refunds or store credit.
• Offering discount codes or goodwill gestures.
• Approving returns outside the normal window.
• Waiving shipping or cancellation charges.
• Raising purchase orders or reorder requests with suppliers.
• Changing an order after it has been confirmed.

For each one, decide whether the agent may act alone, may act within a limit, or must always ask. Some actions, such as cancelling a large wholesale order, may be best kept with a person entirely.

Set thresholds that reflect your real policy

The best starting point is the policy your team already follows, even if it has never been written down. Ask the people who handle these decisions today: what would you approve without checking with anyone? When do you ask a manager? Their answers usually give you the first set of thresholds.

A threshold can be more than a single amount. Useful limits often combine a few conditions, for example:

• Value: a refund up to a set amount can go ahead; anything above goes to a person.
• Frequency: a customer can receive one goodwill discount in a set period; a second request is reviewed.
• Customer history: first-time buyers or accounts with unusual activity always go to a person.
• Reason: damaged-in-transit claims with a photo may be approved; claims without evidence are reviewed.
• Total exposure: the agent can approve up to a daily total across all customers, after which everything pauses for review.

Start with limits that feel slightly too cautious. It is far easier to raise a limit once you've seen the agent's decisions than to recover trust after it has gone too far.

Be clear about who approves what

A limit only works if the request above it reaches the right person quickly. Decide who owns each type of approval. Refunds might go to the support lead, supplier orders to whoever manages inventory, and anything unusually large to a founder or finance head.

Think about the practical side too. Where will approval requests appear: in your helpdesk, on Slack, by email or in a simple dashboard? What happens if the approver is on leave? How long should a customer wait before someone follows up? An agent that hands over well but then leaves requests sitting unanswered simply moves the delay somewhere else.

Each request should arrive with enough context to decide in one look: the customer, the order, what the agent recommends and why. Your team shouldn't have to dig through three systems to approve a refund.

Review decisions and adjust over time

Approval limits are not a one-off setting. In the first few weeks, look closely at what the agent approved on its own and what it passed to your team. Ask simple questions: were the automatic decisions ones you agree with? Were people approving almost everything that reached them, which may mean a limit is too low? Did anything slip through that should have been checked?

Keep a record of every change to your limits, who made it and why. That history makes it easy to explain the agent's behaviour later and to roll back a change that didn't work. It also shows your team that the limits are deliberate decisions, not settings someone forgot about.

Building limits into a first agent

When we build an agent, approval limits are part of the design from the first week, not something added at the end. We agree the actions, thresholds and approvers with you during discovery, test them on real examples and set up the hand-over so your team sees every decision that needs them.

A typical first agent goes from discovery to production in about three to five weeks, depending on system access. Pilots start from ₹40,000, and the final price depends on the systems and work involved.

If you'd like an agent to take on routine decisions while your team stays in charge of the ones that matter, book a call with us. We'll help you map your current approval rules and turn them into limits you can trust.