Sourced note
Agent access, ranked by who absorbs the mistake
Whether to let AI agents into your service has no single answer. Rank each action by who absorbs the mistake when an agent gets it wrong, and the access rule for that action follows.
Three answers in two reports
Two reports on websites and AI agents, published a day apart in mid-October 2026, contained three different answers.
- Block. The Verge reported on 17 October that Delta won't let AI agents book flights, and that Amazon is fighting off outside shopping agents.
- License. The same Verge report noted that Yelp had told TechCrunch it only allows AI agents that pay for its data licensing program.
- A lane. On 16 October, Schwab's CEO, Rick Wurster, told Bloomberg Television that the firm stops outside agents that log in with a customer's password. He also said Schwab plans to let AI tools connect to view balances and do research.
Schwab's answer is the most useful of the three because it splits by action. Looking is fine. Moving money isn't.
The ladder
That split works for any service an agent can reach. Put each action on a rung by asking who pays when the agent is wrong:
| Rung | Action type | Examples | Who absorbs the mistake | Control |
|---|---|---|---|---|
| 1 | Read | Prices, availability, public information | The agent's user: a misread costs them | Open |
| 2 | Hold or quote | A reservation, a cart, a draft order | You, for a while: a bad hold ties up capacity until it expires | Open, rate-limited, with an expiry |
| 3 | Transact up to a cap | Buy, book, pay | Shared between you and the user | A separate lane for agents that say who they act for, with a spend cap and an easy reversal |
| 4 | Move money out or change the account | Transfers, payees, passwords, contact details | You: fraud lands here and is hard to reverse | A person confirms every time |
This is why one company can welcome agents at checkout while another blocks them at login. They are on different rungs.
The ladder is my judgment, not an industry standard. The rungs and controls are a proposal to adapt to your own product.
Outside agents, not your own
This note is about outside agents reaching your service on a customer's behalf. If the agent is your own, working inside your systems, the guide on giving an agent production access covers its identity, narrow grants and rotation.
First step this week
List the actions in your product and fill in one row per action:
| Action | Rung | Who absorbs the mistake | Control today | Control it needs |
|---|---|---|---|---|
Until the "who absorbs the mistake" column exists, "block or allow" has no answer. Then look for any action that is open to agents today but belongs on rung 4.
If you want help sorting your own list, get in touch.