You were told AI is getting cheaper. In late September, Anthropic released a new model and said it costs less to run than the one before it. OpenAI cut its prices the same day. So why can one automation cost a few dollars a month to run while another, doing the same job, quietly runs up a bill nobody expected?

The answer is rarely the model. It is the loop: how many times the automation calls the model, and what stops it when something goes wrong.

Same model, different bill

On September 22, Teradata announced a product it calls Tera Harness. A harness is the layer around an AI model. It decides what the model is asked, in what order, and when to stop asking.

Teradata says that in its own tests, the same model did the same work for much less money once its harness managed the steps. Those are the company’s own tests, and the product is not due out until the fourth quarter of 2026. Treat the result as a vendor claim.

The way it says it saves money is worth knowing, though. According to Teradata and CIO’s coverage, the harness makes a plan before it calls the model. It groups independent work together. It drops calls that do not move the task forward. And it limits how many steps a job can take, based on whether the job is actually making progress.

None of that is a smarter model. It is a better-run job.

Why the bill jumps

Advait Patel, a site reliability engineer at Broadcom, put the problem in one line to CIO: “the same task can take five calls one day and fifty the next.”

That is how an automation surprises an owner. Not because the AI got worse. Because nothing told it when to stop.

Here is a made-up example from a shop’s night. A plumbing company uses an automation to answer after-hours emails. A customer writes in at 11 p.m. asking for a morning slot. The automation reads the email, checks the calendar and drafts a reply for the office to approve.

At 11:05 the calendar connection drops. A loosely built automation keeps trying. Each try is another call to the model, and each call costs a little. By morning it has tried hundreds of times, the bill has a new line on it, and the customer still has no reply.

A well-built version tries a few times and stops. It sends the owner one text: “Calendar is down. One email is waiting for a reply.” Same model. Same night. A very different morning.

A price cut helps the second shop a little. It barely helps the first. A runaway job is still a runaway job, just slightly cheaper per try.

Five questions to ask any automation vendor

You do not need to understand the technology to ask these. You only need the answers in writing.

  1. What does one finished job cost? Not per message or per token. Per answered email, per sorted invoice, per quote sent. Ask them to include retries and the time your staff spends checking the work.
  2. What is the limit on each run? Every automation should have a maximum number of steps, a maximum spend and a maximum time. Ask what happens when it hits one. The right answer is: it stops, and a person is told.
  3. Which steps use AI, and which are plain rules? Looking up a customer, checking business hours or adding up an invoice does not need an AI model. Paying one to think about it costs more and adds a way to be wrong.
  4. Can I see what every run did? You want a record of each run: what it read, what it tried and where it stopped. If nobody can explain why a job took fifty steps, nobody can fix it.
  5. Who watches the bill? Ask who gets an alert if running costs jump in a week, and how fast.

A vendor who answers these clearly has thought about your bill. A vendor who talks only about which model they use has not.

Ask about the setup, not the model

Model prices will keep falling. That is the labs’ job. Whether your automation wastes those savings depends on how it was set up.

The model is one line on the bill. The setup decides how many times you pay that line. When you compare quotes, compare the limits, the records and the alerts, not just the model name.

Our published pricing shows what a build costs and what the monthly fee covers to keep it running, so you can put it next to any other quote and ask the same five questions.

Your one step this week: pick one automation you already pay for, or one you are being quoted. Send the vendor question 2: “What is the limit on each run, and what happens when it hits it?” The answer tells you most of what you need to know.

Sources

Kush AI Automation builds in accounts you own, and the AI usage is billed to your own account at cost, never marked up, so the running bill is always yours to see. Bring one process you want handled and book a free call.