Business continuity when an AI vendor changes the rules

Models are retired, prices change, terms are rewritten. What to prepare so a vendor's decision is an afternoon of work, not an outage.

3 minread 704words last updated

The short answer

AI vendors change the rules as a matter of routine. A model your system depends on is retired with a few months’ notice. A price list changes. Terms on data use, retention or regions are rewritten. A provider is acquired and a product is shut down. None of it is malicious; it is what a fast-moving market does. Business continuity means that when it happens, it is an afternoon of work for you rather than an outage or a rewrite. Four things make that true, and all four are prepared in advance: an alternative model already tested on your evaluation set, an adapter layer that makes switching a configuration change, your data and prompts in your own accounts, and a monitored channel for vendor announcements.

What vendors change, and what it does to you

ChangeTypical noticeEffect if unpreparedEffect if prepared
Model retiredMonths, sometimes weeksCalls fail on the retirement date; a scramble to find and test a replacementSwitch to the tested alternative; rerun evaluation; done in a day
Model updated in placeSometimes noneBehaviour drifts; quality changes silentlyMonthly sample catches it; evaluation rerun; prompts adjusted
Price changeWeeksBudget overrun or a rushed switchCost per case alert; switch if the alternative is cheaper at equal quality
Terms changeWeeksData handled under terms you would not have acceptedOwner reviews against policy; switch if unacceptable
Product withdrawn or company acquiredVariableFeature disappears; data locked in someone else’s systemData and prompts already yours; alternative already tested
Rate limits or quotas changedLittleCalls throttled at peakCaps and alerts; alternative for overflow

The one-page continuity note per system

  1. What breaks if the current vendor changes: which features, which users, which processes.
  2. The tested alternative: provider, model, evaluation score, cost per case, date tested.
  3. How to switch: the configuration change in the adapter, the retest, the announcement to users.
  4. Who does it and how long it takes.
  5. Where the announcements come from and who reads them.
  6. Last tested switch: the date the switch was actually performed on a preview environment.

Testing the switch

Once a year, like a backup restore, actually perform the switch on a preview environment: change the adapter configuration to the alternative, rerun the evaluation set, run a few live cases, note what broke and how long it took. The first test always finds something: a provider-specific parameter, a prompt that behaves differently, a data format. Finding it in a scheduled hour is the point.

What this means for you

Assume your AI vendor will change the rules within the life of your system, because they will. Keep a tested alternative, an adapter, your own data and prompts, and a named reader of vendor announcements. Write the one-page continuity note per system and test the switch yearly. Then a vendor’s decision is a footnote in your monthly report rather than the reason a process stopped.

Written by the CivSec S.M.A.R.T team

We build and run websites, software and AI systems for businesses. We write about what we see in that work, in plain language, and we update articles when things change.

Last checked . Spotted something outdated? Tell us.

Frequently asked questions

How much notice do vendors give before retiring a model?

Usually months, sometimes less, occasionally none when a company fails or a product is withdrawn. Notice arrives through release notes, developer emails and status pages that nobody in a small business reads by default. Someone must own that channel, or the notice arrives as an error.

What if the vendor changes its data terms?

Terms can change in ways that matter: training on inputs, retention, regions, subprocessors. The internal owner reads terms changes when notified, checks them against the policy and the data map, and triggers the switch if the new terms are unacceptable. Being able to switch is what makes reading the terms worthwhile.

Is this a real risk or a theoretical one?

Real and recurring. Models have been retired on schedules that broke integrations that were not updated, prices have moved in both directions, terms have been rewritten, and AI products have been shut down after acquisitions. None of it is malicious; it is a fast-moving market. Preparation is cheap; discovering the change in production is not.