Connecting your tools

Connecting Stripe to an AI agent

As soon as an agent touches money, the question is no longer what it can do but what it is forbidden. The Stripe connector gives access to customers, payments, subscriptions, invoices, refunds and reports. It is plugged in with a deliberately restricted key — which is what Stripe recommends for autonomous agents — and the gestures that move money stay behind human approval.

1. What the agent does in Stripe

The connector goes through Stripe's official server. The scope covers the objects a small company's billing rests on:

Read the real state. Which payments went through, which subscriptions are active, which invoices are unpaid and for how long. That is the material of a fair follow-up, rather than an approximate one that offends a customer who is up to date.
Keep the billing. Consult and operate invoices and subscriptions, prepare what needs preparing.
Pull the figures. The reports, to answer “where do we stand this month” without opening three tabs.

The connection requires designating the account and the environment: you do not work in production by mistake when you meant a test.

2. A restricted key, and why Stripe recommends it

Unlike most connectors, this one is not plugged in with a simple account sign-in: you create a Stripe key restricted to the only actions your agents need, and you paste it once.

It is a little more work at the start, and it is deliberate: Stripe explicitly recommends restricted keys for autonomous agents. The reasoning is sound. A full key gives everything, forever, to whoever holds it; a restricted key carries the limit within itself, on Stripe's side — that is, somewhere nullbot cannot contradict.

Consequence: your real ceiling is the one you wrote into the key, and it stays true even if a setting changes on our side. It is the best guarantee available, and it is worth the five minutes of setup.

3. What must never leave human approval

Stripe is classed at the highest stakes level in the catalogue. By default, any action that writes or deletes requires human approval before execution. Three gestures deserve to stay that way for good:

The refund. It moves money outwards and cannot be taken back. Stripe itself advises keeping write and refund permissions behind a prior review.
Changing a subscription. Altering an amount or a frequency commits the commercial relationship over time, often without the customer noticing straight away.
Cancellation. Ending a subscription is a commercial decision, not a maintenance operation.

Our position is clear: those three should not be relaxed, even when the rest is. An agent that prepares a refund and waits for approval saves time without transferring the decision. An agent that refunds on its own transfers the decision.

To which the rule that holds everywhere is added: each agent's spending is reserved before the model call, and going over is made impossible — the mechanism is detailed separately.

4. What it changes over time

Payments is the area where irregularity costs most, because it is paid for in cash flow.

Week 1. The real state of unpaid invoices is established — often a surprise, one way or the other.
Month 3. Follow-ups go out at the right moment, to the right customers, for the right amount. Nobody is chased wrongly, which is the main reason manual follow-ups end up being abandoned.
Year 2. Payment delays become tracked data rather than a feeling, and you can act on causes rather than symptoms.

No quantified improvement is promised: it depends on your volume, your payment terms and your customers.

5. Frequently asked questions

Can an agent refund a customer on its own?

Not by default, and we recommend never allowing it. A refund requires human approval before execution; the agent can prepare it, gather the elements and propose it, the decision stays with a person. That is also Stripe's recommendation for autonomous agents.

Why a key rather than an account sign-in?

Because the restricted key carries the limit on Stripe's side. You write into the key the only actions allowed, and that ceiling stays true whatever happens elsewhere. A classic account sign-in would give a broader scope, controlled only by settings outside Stripe.

Does the agent see my customers' banking details?

It sees what Stripe's interface exposes, that is, the billing objects: customers, payments, subscriptions, invoices. Full card numbers are not exposed by Stripe — that is the very principle of their product. And nullbot connects to no bank account and triggers no transfer.

Can it be tested without risk?

Yes: the connection requires designating the environment, and Stripe provides a complete test environment. That is the right way to tune an agent's behaviour before letting it near production — and to check concretely what approval before writing blocks.