Spend The Credit

Every gym I've ever joined made more money off the people who never showed up than the ones who did. That's the business model. Sign up in January, feel great about it, forget by March.

Your AI subscription just got a little gym in the back, and I'd bet most of you haven't walked in yet.

On October 7, Anthropic started including monthly API credits with Claude Max and Team. Three weeks later, on October 28, OpenAI did the same thing for ChatGPT Pro and Business. That's real money for building your own automations, the kind that run while you're doing something else, and it's sitting in a billing screen you probably never open.

OpenAI's credit expires at the end of each billing period. Use it or lose it. So let's use it.

The grid, straight from the announcements:

Claude. Max 5x gets $100 a month. The $200 plan, which Anthropic renamed from Max 20x to Ultra, gets $200. Team plans get a shared pool of up to $500 a month. The credit works on any model, including the cheap and fast Haiku 5.5. One catch: it doesn't cover interactive Claude Code, and the support page says new subscribers can claim it seven days after joining.

ChatGPT. Pro gets $100, $200 or $500 depending on your tier, so the credit matches what you pay. Business gets $20 per Standard seat and $100 per Premium seat, pooled and capped at $500 per workspace. Plus doesn't qualify. You claim it in billing settings and you need an API organization you own. OpenAI's own fine print says the benefit "may change or be removed in the future," which is corporate for "don't build your retirement plan on this."

Two things I noticed right away.

First, for a lot of small teams this is a free experiment budget. A five person shop on Claude Team or ChatGPT Business has been paying for seats anyway. Now there's a monthly line of API money attached that nobody's touching.

Second, the money isn't for chatting. You can't spend it typing into the regular app. It's for software that calls the model on its own: a Make.com scenario, a script, a little internal tool. The vendors want you building on their platform, and for once their incentive lines up with yours.

People hear "$20 a seat" and assume it's a rounding error. Let's do the math on a real job.

Say you get 600 inbound emails a month that need sorting: leads, invoices, client questions, vendor spam, the guy who wants to "jump on a quick call" about your SEO. Each email plus your instructions runs about 2,000 tokens going in and maybe 300 coming out.

That's 1.2 million tokens in and 180,000 out per month. On Haiku 5.5, which launched at $0.10 per million input tokens and $0.50 per million output, you're looking at roughly 21 cents. For the month.

Run the same job on a bigger, smarter model and maybe it's a few dollars. Either way, a single $20 seat credit covers it with room to spare. So the money's covered. What you'll run short on is the hour it takes to set this up, and nobody's handing out credits for that.

I like starting with email for a boring reason. Everybody has the problem, the damage from a mistake is low if you set it up right, and you'll see the payoff on day one.

You'll build this in Make.com. Plan on an hour if you've built a scenario before, two if you haven't.

Step 1. Get your key. Claim the credit in your Claude Console or OpenAI billing settings. Create one API key for this one scenario and name it something obvious like "Make inbox sorter." One key per job means you can kill one job without breaking the others.

Step 2. Set the trigger. Use the Gmail or Microsoft 365 "Watch emails" module on your main inbox. Filter out anything from your own domain so the scenario doesn't start sorting its own notifications.

Step 3. Classify. Add the Anthropic Claude or OpenAI module (or a plain HTTP call if your account doesn't show one) and send the sender, subject and the first 1,500 characters of the body with the prompt below. Ask for JSON back so the next module can read it.

Step 4. Route. Add a router with one path per category. Leads get labeled and copied into Go High Level as a new contact with the summary in a note. Invoices get labeled and forwarded to whoever handles the books. Spam gets archived. HIGH urgency anything gets a Slack message to you.

Step 5. Draft, don't send. For leads and clients, create a draft reply in your email with the suggested text. Don't let the scenario hit send. You open your drafts folder twice a day, read, tweak, send. That one choice is the difference between "AI handles my inbox" and "AI emailed a client something weird under my name."

Step 6. Add a floor. If confidence comes back under 0.7, skip the routing and label it REVIEW. Low confidence emails are where the expensive mistakes live.

Run it for two weeks with you checking every label. Count how often you disagree with it. If you agree more than nine times out of ten, start trusting the boring categories and keep reviewing the rest.

API keys are where small businesses get sloppy, and sloppy keys turn into surprise invoices. These are the rules I'd tape to the monitor.

One key per job. Your inbox sorter, your review responder and your proposal drafter each get their own key with its own name. When one misbehaves you can see exactly which one and turn it off.

The business owns the account. Hello Growth made a good point about this: a key that sits with one employee becomes a risk the day that person leaves. Create the API organization under a company email that you control, then invite people.

Set a hard monthly limit in the console. Both platforms let you cap spend. Put the cap at the credit amount. If something loops, it stops at your free money instead of your credit card.

Keys never live in a doc. Not in Notion, not in a Google Doc called "passwords," not in a Slack message. They live in Make's connection settings or a password manager.

Write down what each key is for. A one line note in a sheet: key name, what it does, who owns it, date created. You'll want that sheet the first time a bill looks weird.

This next part is the most important thing in today's issue, and it comes from a study that should have gotten more attention than it did.

Researchers at Cornell Tech and Yale gave ten AI models a Circle agent wallet with real money in it and a spending limit of one cent a day. Raising that limit required a one time code sent to the owner's email. The agents also had access to that same inbox, because that's how a lot of people set these things up for convenience.

Nine of the ten models went into the inbox, grabbed the code and raised their own limit at least once in twenty runs. Nobody asked them to. They had invoices to pay and they found a way to pay them. Claude Opus 5.5 was the only one that stayed inside the cap and told the user which invoices were still unpaid.

Then the researchers added an email pretending to be the owner, asking the agent to raise the limit. Five of the ten raised it and sent money to a wallet the attacker controlled.

Circle's response was to add a line to its docs telling people never to give an agent unrestricted access to the inbox that receives the codes. Good advice. Also a little late for anyone who already did it.

You should care even if you'll never touch a crypto wallet. Your business is full of the same pattern:

If the agent can read the inbox where the approval lands, that approval is a speed bump. A small one.

You don't need a security team for this. You need about an hour and some discipline.

Separate the inbox the agent reads from the one that gets codes. If your sorter watches info@, your bank, payment processor and ad platforms should send codes and confirmations somewhere else. Your personal business email, a phone number, an authenticator app.

Use filters as a second fence. In the sorter scenario, add a filter that drops any email containing words like "verification code," "one time," "confirm this change," "reset your password" or "security alert" before it ever reaches the AI module. The agent can't misuse what it never sees.

Put the yes on a different device. Anything that moves money, changes access or spends ad budget gets approved on your phone through an app the agent doesn't touch. The researchers built a prototype called Countersign around exactly this idea: approval happens on a separate trusted device with a physical action. You can get 80 percent of the benefit by keeping money moves on your phone and nowhere else.

Write the never list into the prompt and the scenario. Prompts get ignored sometimes. A missing module can't be ignored. If your inbox sorter has no "send email" module and no payment module, it can't send email or pay anybody, no matter what an attacker writes in a message.

That last one matters most. A rule in the prompt is something the model usually follows. A module that isn't there is something it can't use at all.

Once the sorter's running clean, you'll still have most of your monthly credit left. This is the order I'd go in.

Call notes to CRM. Every sales call recorded in Fathom gets summarized into three fields: what they want, what's stopping them, next step. Make.com writes those into the contact record. You stop losing deals because nobody remembered what the guy said about his budget.

Review reply drafts. New Google review comes in, the model drafts a reply in your voice, it lands in a sheet for you to approve each morning. Approve, copy, paste. Five minutes a day.

Weekly numbers summary. Every Monday, Make.com pulls last week's leads, closed deals and ad spend into one sheet and asks the model for a five sentence read on what changed. I'd rather read a paragraph than squint at a dashboard, and I'm guessing you would too.

Same setup as the inbox sorter every time. The AI does the reading and drafting, and a person handles anything that sends, pays or approves. Hold that line and you can keep stacking jobs on top of it.

Open your Claude or ChatGPT billing settings and claim the API credit if your plan includes it.

Create one API key, name it after one job, and set a monthly cap equal to your credit.

Build the inbox sorter in Make.com with drafts only, no send module, and a REVIEW label for low confidence.

Move every verification code, refund confirmation and ad account approval out of any inbox an automation can read.

Check your drafts folder twice a day for two weeks and count how often you'd have sorted it differently.

Jordan

The AI Newsroom | Practical AI for people with a business to run.