Say you set up an AI helper that answers new web inquiries. It runs on one AI company’s service. It works well, so you build more on top of it.

Then that company changes its prices, its limits or the model you get. What happens to your setup?

For most small shops the honest answer is: you pay the new price, or you rebuild. That is not a plan. This post shows how to keep a third choice open.

Why this is worth thinking about now

On November 3, TechCrunch reported on an updated filing from Anthropic, the company behind Claude. The filing covers the quarter ending September 30. I could not open the filing itself, because the SEC site blocked my request. So everything below about it comes from TechCrunch’s report.

The report says Anthropic is growing fast, and that the cost of the computers that run its AI took a bigger share of its money than the quarter before. It says the filing ties part of that to computing power bought under short-term deals, which “costs significantly more” than power bought under long ones. It also says the filing expects computing costs to stay the company’s largest expense for the foreseeable future.

TechCrunch also says this may be a blip, and that demand is strong. Nothing in the report says any price will go up. I am not predicting that.

The point is simpler. Your AI bill rests on someone else’s costs, and you cannot see those costs. A company with a hard quarter may change a price, a limit or which model is on your plan. So build in a way that lets you switch.

What “switchable” looks like

  1. Keep your instructions in your own account. The text that tells the AI how to answer your customers is valuable. Store it in a document or workflow you own, and not only inside one vendor’s settings page.
  2. Make the AI one step you can change. In n8n, for example, the AI model is its own step. n8n’s documentation lists model steps for Anthropic, Google Gemini and OpenAI, among others. If the model is one step, switching it is one change and not a rebuild.
  3. Keep customer records in your own system. Job history, call notes and contact details belong in your CRM or spreadsheet. Give them to the AI when it needs them. Do not let the AI’s own memory become the only copy.
  4. Set a spending cap and an alert. OpenAI’s documentation says you can set spend alerts and a hard monthly limit. Anthropic’s says spend limits set a maximum monthly cost. Turn yours on, so a price change shows up as an alert and not a surprise bill.
  5. Test a backup now, while nothing is broken. Pick ten jobs your AI does often, such as quote follow-ups, review replies and intake notes. Run them through a second company’s model. Have a person read the results, because different models can word things differently. Write down which jobs hold up. That list is your fallback.

A made-up example

This is a made-up example. An HVAC company in Glendale uses one AI step to draft follow-ups after estimates. The owner tested a second model last month and kept a list of what worked.

Then the company’s plan changes and the owner’s usual model gets limited. The owner switches that one step to the backup, reads five sample emails, and fixes one greeting that sounded stiff. The work continues the same afternoon. No one rebuilt anything.

Compare a shop that stored its instructions only in the vendor’s app, and has no test of any other model. That owner has to start from nothing while the follow-ups pile up.

Your one step this week

Pick the AI job you use most. Find where its instructions are stored. If they live only inside a vendor’s app, copy them into a document you own. It takes about ten minutes, and it is the first thing you would lose in a forced switch.

You already avoid buying all your materials from one supplier. Treat your AI the same way.

If you want help checking how tied your setup is to one company, see what we build or book a free call.

Sources