DEEP DIVE

COMPANION TO ISSUE #162 · 18 NOVEMBER 2026


The spend ceiling

You’ll see one team’s full map of what its AI tools could spend, which limits actually stopped spending, and what broke when they tested one. Doing it took the example agency about six hours over two weeks.

6 MIN READ · FICTIONAL WORKED EXAMPLE, INVENTORY AND CHECKLIST

AI THAT SPENDS / WORKED EXAMPLE

Eleven things could spend. Two could be stopped.

On a Sunday in October, an automation at the agency in this example got stuck retrying the same request. By nine on Monday morning it had spent $1,140 of API credit. Its project had a $300 monthly budget, set up as an alert. That alert went to a billing inbox nobody reads at weekends.

Kestrel Lane, a nine-person marketing agency, is fictional, and so is every figure that follows. It uses AI for drafts, research and images, runs a few automations on API keys, and spends about $900 a month on AI across all of it.

The director wrote two rules. First: nothing that can spend runs without a limit that stops spending. An email doesn’t count. Second: limits get tripped on purpose before anyone trusts them.

How they did it, about 5 hours 40 minutes in all: three months of card statements and billing pages (45 minutes); the issue’s prompt, run with the operations lead (25); a TruffleHog scan with a developer (1 hour); checking what each limit does when reached (1 hour); worst-day estimates (40); reading the terms (30); setting limits (1 hour); one deliberate test (20).

The inventory

Month is a normal month’s spend; together they make $900. ‘Tested’ means someone saw the limit stop something. ‘Configured’ means it’s set but hasn’t been tripped.

API project A: drafting tools

Owner: Developer

Month: $220

Before: Alert at $400

After: Hard limit $440

Status: Configured. Same switch tested

API project B: automations (the Sunday one)

Owner: Ops lead

Month: $180

Before: Alert at $300

After: Hard limit $360. Retries stop after 5

Status: Configured. Same switch tested

Prepaid API account: research

Owner: Strategist

Month: $90

Before: Auto-reload $200 when under $50. No limit

After: Auto-reload off. Limit $180

Status: Configured

Image generator, pay per use

Owner: Designer

Month: $120

Before: Card saved. No cap

After: Monthly cap $240 in the tool

Status: Configured

Automation platform plan

Owner: Ops lead

Month: $60

Before: Plan task cap. Stops

After: Unchanged

Status: Already a stop

Router key for model testing

Owner: Developer

Month: $25

Before: Per-key limit $50. Stops

After: Unchanged

Status: Already a stop

Transcription API

Owner: Account manager

Month: $70

Before: Alert at $150

After: Hard limit $140

Status: Configured. Same switch tested

Video tool, pay per minute

Owner: Designer

Month: $40

Before: Card saved. No cap

After: Monthly cap $80

Status: Configured

Browser agent buying office supplies

Owner: Ops lead

Month: $95

Before: Company card saved. No cap

After: Virtual card, $150 a month. Orders over $60 wait for a person

Status: Tested: $70 order held

Former freelancer’s key, old project C

Owner: Nobody

Month: $0

Before: Live. No limit

After: Revoked

Status: Confirmed revoked

Old trial account with card saved

Owner: Director

Month: $0

Before: Card saved. No cap

After: Card removed. Account closed

Status: Confirmed closed

Before: 2 limits that stopped, 3 alerts, 6 with nothing. After: 8 that stop, 1 capped card, 2 shut off. The test ran on a spare project using the same hard-limit switch as A, B and the transcription API. Those three are set the same way, not tripped themselves.

The worst day, estimated. In the example, the loop spent $1,140 in about 14 hours, roughly $80 an hour. Left alone, that’s about $1,950 a day against a normal month of $180 on project B. With the hard limit, the month tops out near $360. ‘Near’, because limits like this aren’t enforced instantly: OpenAI, for one, says recorded spend can slightly exceed the limit.

The agent with a card. The browser agent got the office-supplies order right 14 times. Once, the listing changed and it bought a 12-pack instead of a 2-pack. The agent stayed. It now pays with its own capped card, and the $70 test order sat waiting for approval, as designed.

The freelancer’s key. TruffleHog found it in an old project folder and marked it verified: it still worked eight months after the project ended. Three other matches came back unverified. That isn’t proof they’re dead, so those were rotated too.

What broke in the test. A hard stop is an outage you chose. When the spare project hit its limit, one automation retried every few seconds and buried its error log. Retries now stop after five attempts and post to the team channel.

Who pays, by situation. The agency read its terms and sorted three cases. A purchase someone approved is the agency’s. The agent’s 12-pack was too, though the supplier took the return. Spending on a stolen credential depends on what was stolen: card disputes for the card, each provider’s fraud policy for API keys, and one reseller states it gives no refunds.

AI THAT SPENDS / FROM THE LABS

Failed, and bought anyway

The 12-pack is invented. The pattern behind it has been measured. On 5 November Visa published its Agentic Commerce Benchmark: 22 AI models from 12 labs, 15 shopping tasks such as choosing a TV or a laptop, three attempts each. A run passed only if the model found every requirement and made the right buy or no-buy call. Overall they passed 57.2% of runs, from 82.2% for the best model to 28.9% for the worst. About 54% of failed runs still ended in a purchase. Only 34.2% of model-and-task pairs passed all three attempts.

The caveats. It’s Visa’s own benchmark, the release doesn’t name the models, and it tested buying decisions, not checkout or payment. Visa also sells the rails these agents will pay on. My inference, not Visa’s finding: if roughly one run in four ends in a wrong purchase decision, approving each purchase is doing real work today, and pre-approved limits should start small.

AI THAT SPENDS / THE CHECKLIST

Carry it over

Last week’s companion listed what a task contains before you hand it over. This is the same habit pointed at money: find out what a setting does before you trust its name.

An alert is not a ceiling. Check what happens when the limit is reached. If the answer is an email, keep looking.

A separate key is not a limit. It lets you see and revoke one job’s spending. Only a limit caps it.

Auto-reload has no ceiling of its own. Turn it off, or put a hard limit above it.

Agents that buy get their own card. One with a monthly limit, not the company card saved in a browser.

Cap retries as well as spend. A loop that hits a limit can still flood your logs and alerts.

Leave headroom. Limits lag a little. Twice a normal month is a starting point, adjusted for what an outage would cost you.

Trip one limit on purpose. On a spare project, not in front of a customer.

The strongest objection

‘Caps break things. A hard stop at the wrong moment takes a product down, and a runaway bill can be refunded.’ The first half is true, which is why the limit has headroom and gets tripped on a spare project first. The second half is sometimes true. At Railway, refunds for the stolen keys depended on each provider: two have policies that cover fraud, others review case by case, and one enforced a no-refund rule, so Railway is paying that share itself. A refund you have to negotiate isn’t a ceiling. In my view, an outage you planned for costs less than a bill you didn’t.

YOUR MOVE

Run the issue’s prompt with whoever pays your AI bills. Put a hard limit on the top three items, with headroom, and switch off any auto-reload that has no limit above it. Trip one limit on a spare project. Reply with one line, even before you’ve fixed anything: how many things could spend, and how many could be stopped. Never paste a key.

A limit you didn’t switch on is an email.

THIS WEEK’S FORECAST

By 30 June 2027, OpenAI will let at least some ChatGPT Wallet users pre-approve purchases within a spending limit, so an agent can buy without a tap each time, and OpenAI’s terms will still say you’re responsible for transactions made on your behalf. The first half is likely: OpenAI has said limits by amount, category and time period are planned. The second half is the bet, that autonomy arrives before liability moves. If card networks, a regulator or OpenAI itself shift that first, I lose.

Counting. It counts if OpenAI’s announcement, help centre or terms describe pre-approved purchases as live for some users, not planned, and on 30 June 2027 its service terms or Wallet terms still put responsibility for transactions made on your behalf on the user. A pilot open only to OpenAI staff doesn’t count. If the feature launches but the terms move that responsibility to OpenAI, a card network or the merchant, the forecast fails. Card-dispute and fraud rights, which already exist alongside the terms, don’t change the result either way.