An AI vendor isn’t a one-time purchase. It’s an ongoing relationship where the other side keeps writing the terms. Most businesses never read those terms until something changes and the change lands on their desk with a deadline attached.

That’s a solvable problem, but only if you know what you’re planning for. “AI risk” isn’t one thing — it’s four fairly distinct categories, each with its own trigger, its own warning window, and its own fix. Here’s how to think about each one.

1. Pricing changes

API pricing is not a fixed cost the way your office lease is. Vendors reserve the right to change it, and they use that right. OpenAI’s Services Agreement states that price changes take effect fourteen days after they’re posted to the pricing page — a real, documented clause, not a hypothetical. Other major vendors build similar short notice windows into their commercial terms. Two weeks is not a lot of runway if your pricing to customers assumes today’s per-token cost.

It isn’t just the sticker price, either. Tier structures get rebuilt — what counts as a “included” request, what triggers overage billing, which usage level unlocks which rate limit — and a tier restructuring can change your effective cost per customer even when the headline price per token doesn’t move.

What to do: model your cost per customer interaction assuming per-token pricing moves, not stays flat. Build a small margin cushion into anything you resell or bundle. Put the vendor’s pricing-change notification on someone’s calendar as a real task, not an email to skim — a two-week window means the day it lands is the day you act, not the day you notice it.

2. Model deprecation and sunsetting

Every major AI lab retires older models on a rolling schedule, and the notice period is genuinely different vendor to vendor. Anthropic’s documented policy is to notify customers with active deployments at least 60 days before a publicly released model retires. OpenAI commits to a longer runway for generally available models — at least six months, or three months for specialized variants like chat- or coding-tuned snapshots — but drops to as little as two weeks for anything explicitly labeled “preview.” Google’s Gemini API follows a similar pattern: preview and “latest”-style model aliases can get as little as two weeks’ notice before a breaking swap, while stable models get more warning but without one fixed, universal number across the board.

The gap between “60 days” and “2 weeks” matters a lot if you’ve hardcoded a preview model ID into a production workflow and assumed it behaves like a stable release.

What to do: before you build anything customer-facing on a specific model, go read that vendor’s actual deprecation policy page, not their marketing copy. Know which of your integrations point at “preview” or “latest” aliases versus pinned, stable model IDs — they carry very different notice guarantees. When a deprecation notice arrives, treat it as the trigger to test the replacement model immediately, not a date to circle for later.

3. Policy and usage-restriction changes

Rate limits, usage tiers, and content or usage policies aren’t fixed either, and they don’t always move against you — which is part of why they’re easy to ignore. In 2026, Anthropic restructured its Claude API usage tiers, moving away from spend-based tiers and substantially raising rate limits for many accounts. That was good news for the businesses it affected. But it’s also proof that the ceiling you’re building around today — requests per minute, tokens per day, what tier your account sits in — is a policy decision the vendor can revise, not a law of physics.

The same is true of usage policies that govern what you’re allowed to build: restrictions on automated decision-making, certain regulated industries, or specific customer-facing use cases can tighten with a policy update, which can force a live product to add guardrails or disclosures on short notice.

What to do: don’t architect a product around today’s rate limit as if it’s permanent capacity — build basic backoff and queuing so a lower limit degrades your product gracefully instead of breaking it. Know how your vendor’s tier-increase request process works before you’re under pressure to use it. And reread the usage policy for your specific use case periodically; it’s revised more often than most businesses check.

4. Regulatory shifts in your market

The most overlooked category is the one that doesn’t come from your vendor at all. Regulation can change what you’re allowed to do with an AI vendor’s product in a given market, even if the vendor itself hasn’t changed a thing.

The clearest current example: the EU AI Act’s transparency obligations under Article 50 apply from August 2, 2026. They require that AI systems designed to interact directly with people — chatbots, voice agents, avatars — be built so users are informed they’re talking to AI, disclosed before or at the very start of the interaction. Crucially, this obligation reaches deployers, not just the labs building the underlying models — meaning the business running the chatbot, not just the company that built it. It also applies extraterritorially: a business based anywhere, serving users located in the EU, can be in scope. Penalties run up to €15 million or 3% of global annual turnover, whichever is higher.

What to do: if you run any AI-driven customer interaction — chat, phone, or otherwise — and you serve customers in the EU, add an explicit “you’re talking to AI” disclosure at the start of the interaction now, and don’t assume your AI vendor’s own compliance work covers your obligations as the deployer. This is one live example of a pattern that will keep recurring: as AI-driven customer interactions become common, the regulatory attention increasingly lands on the business deploying the tool, not only the company that built it.

Where this leaves you

None of these four categories are edge cases. Pricing moves, models retire, tiers get rebuilt, and regulators are actively writing rules for exactly this kind of AI-driven customer interaction right now. The businesses that get caught off guard aren’t usually the ones using AI aggressively — they’re the ones who never read the notice policy until the notice showed up.

If you want help mapping which of your workflows are exposed to which of these four categories, and what a realistic response plan looks like for each, get in touch or take a look at how we approach this in our AI strategy work.

Frequently Asked Questions