The demo had gone perfectly. Priya, who led a six-person product team building an AI assistant for a logistics firm, watched the vendor’s model summarise a messy shipping thread in four seconds and thought: this is it. The sales rep sent the order form that afternoon. It was a one-page summary with a price, a start date and a link to “full terms”. Her finance lead wanted it signed by Friday to catch the discount. This is the story of what that link contained, and why a team that normally reviews every pull request approved a contract nobody had read.
It is told as a story because the details feel familiar to almost anyone who has bought an AI platform. The lessons are practical and fit on a page.
Friday: the order form looks fine
The summary page listed a monthly platform fee, a generous-sounding allowance of processing credits, and a twelve-month term. Nothing on it looked unreasonable. The team had already spent a week testing the product and, like most engineers, trusted what they had seen work. They also had a project plan that assumed the model would be live in six weeks.
What they had not seen was the document the order form pointed to. It ran to thirty pages, written for the vendor’s lawyers rather than for people deciding on a tool. Priya skimmed the first page, saw nothing alarming, and signed. Anyone in a hurry would have done the same, which is exactly why the habit of reading the terms first is worth building deliberately.
Month two: the surprises start arriving
The first surprise was minor. The credits that had sounded generous were consumed far faster than the demo suggested, because longer documents counted for more than one credit each. Overage was charged at a higher rate, and the bill rose accordingly.
The second surprise was bigger. The team wanted to try a different model from another provider for one task and discovered that exporting their fine-tuned configuration was not supported. The terms described it as part of the service, not as something the customer owned. Their work stayed on the vendor’s platform.
The third surprise was about data. A security review asked where customer shipping data was stored and for how long. Nobody on the team could answer, because nobody knew what the retention clause said. When Priya finally went looking, she found a paragraph allowing the vendor to retain inputs for a period to improve its services unless the customer opted out in writing. The opt-out step had a deadline that had already passed.
For a fuller view of how businesses should think about AI-assisted choices in general, the piece on what AI changes in day-to-day business choices is a good companion read. The point here is narrower: the best model in the world still arrives wrapped in terms.
Month ten: the renewal notice
The real lesson came late. Deep in the agreement was a clause saying the term renewed automatically for another twelve months unless the customer gave written notice sixty days beforehand, sent to a specific address. The team had planned to review the tool at the one-year mark. By then the window had closed.
Priya spent several weeks negotiating, and the vendor eventually agreed to a shorter extension at a modest cost. It could have gone worse. But the episode took time from the roadmap, strained the relationship with finance, and left everyone with a lingering sense that they had been careless. They had, but only in a way thousands of teams are.
What the team changed afterwards
After the dust settled, they wrote a short rule: no AI tool is purchased until someone has read six specific clauses and put the answers in a one-page note. The six were these.
- Term and renewal. How long is the commitment, does it renew itself, and what notice is needed to stop it?
- Usage and overage. What counts as one unit, and what happens to price when you go over?
- Data handling. What is kept, for how long, who can see it, and can it be used to train anything?
- Ownership and portability. Who owns outputs and customisations, and can you take them elsewhere?
- Liability and uptime. What does the vendor promise about availability, and what are the remedies if it fails?
- Changes. Can prices, limits or the model itself be altered, and how much warning do you get?
The note is short on purpose. Anyone on the team can read it in two minutes, and the person approving the spend can see in one glance whether a discount is worth its strings.
Practical habits that make the reading quicker
- Search, then read. Open the terms and search for words like renew, terminate, retain, indemnify and fees. Read the paragraphs around each hit.
- Ask for the full terms before the demo ends, not after you have mentally committed.
- Put the notice date in a shared calendar the day you sign, with a reminder well ahead of it.
- Test the exit. Before going live, try exporting your data and configuration once, so portability is a fact, not a promise.
- Bring in the right professional when the stakes are high. Contract wording is a specialist subject, and an hour of the right person’s time is cheap by comparison.
Teams working with visual or high-throughput workloads should be especially careful, because usage-based pricing climbs quickly there. The article on scaling fast visual pipelines without choking shows how quickly throughput assumptions can change, and a bill is only one clause away from reflecting that.
A note on finding legal help
Not every team has a lawyer on call, and many contracts do deserve a professional’s eyes. For readers in India, WakilSahab is a website where people look up lawyers by practice area and read legal news, guides and forum threads for India. It is somewhere to begin looking, not a source of advice, and the scope and fees of any engagement still need to be settled directly with whoever you choose.
Where Priya’s team ended up
They kept the vendor, in the end. The product was good, and the renegotiated term was fair. What changed was the process. The next purchase, a vector database, took an extra afternoon and surfaced a clause about storage fees that would have doubled the cost at their expected scale. They chose a different plan and moved on.
The moral is not that vendors are villains. Most are simply writing agreements that protect themselves, as you would. The remedy is a short, repeatable act: before you commit, read the parts that describe the bad day, not just the good one. The headline is the pitch. The small print is the contract, and it is the contract that you will be living with in month ten.