Review the Whole System Every Quarter

An AI stack decays quietly. Nothing breaks with an error message. The model behind your assistant gets replaced, a price moves, a privacy default resets during an account migration, and an automation keeps firing on a rule that stopped matching your business six months ago. You find out when a customer does.

Ninety minutes a quarter is enough to catch all of it. The value is not in the checking, it is in doing the same list in the same order so you can see what moved since last time.

Vendors publish very different guarantees about advance warning

Start here, because it tells you how much warning you are actually entitled to, and the answer varies by more than an order of magnitude.

Anthropic's model deprecation page defines a four-stage lifecycle: Active, Legacy, Deprecated and Retired, with Retired meaning requests to that model will fail. On warning, the page states: "Anthropic notifies customers with active deployments for models with upcoming retirements, providing at least 60 days' notice before model retirement for publicly released models." The page also carries a live example of the cycle in motion, showing claude-opus-4-1-20250805 deprecated on June 5, 2026 with a retirement date of August 5, 2026, and both claude-sonnet-4-20250514 and claude-opus-4-20250514 already retired on June 15, 2026.

OpenAI publishes a longer commitment with a large exception. Its deprecations page gives at least 6 months of notice for generally available models and at least 3 months for specialized variants, then states: "Preview models, identified by preview in the model name, may be retired with much shorter notice, such as 2 weeks." The current entries show the gpt-5 family with a shutdown date of December 11, 2026 announced on June 11, 2026, and the audio and realtime models shut down January 20, 2027 announced July 20, 2026.

Google publishes the dates without publishing the promise. The Gemini API release notes list specific shutdowns, including the imagen-4.0 image models on August 17, 2026 and gemini-robotics-er-1.6-preview on August 31, 2026. What that changelog does not carry is a stated minimum notice period comparable to Anthropic's 60 days or OpenAI's 6 months. The dates are there; the guarantee is not. If your business depends on a specific Google model, that absence is the reason to read the changelog rather than wait for an email.

One cross-vendor change from this quarter is worth flagging because it hit two vendors at once. Anthropic's page lists temperature, top_p and top_k as deprecated for Claude Opus 4.7 and later, returning a 400 error when set to a non-default value. Google's release notes record the same three sampling parameters deprecated as of July 21, 2026. If anyone built you an integration that sets those, it is now on borrowed time in two places.

The ninety-minute review, in the order that surfaces problems fastest

1. Open the three deprecation pages. Read only the top table on each. You are looking for a model name you recognise from an integration someone built for you, and for any date inside the next two quarters. Five minutes.
2. Pull your own usage record. Anthropic documents the exact path: go to the Usage page in Claude Console, click Export, and review the downloaded CSV, which breaks usage down by API key and model. Its own page describes this as the way to "locate any instances where your application is still using deprecated models". Do the equivalent on any other platform you pay usage-based fees to.
3. Re-check the privacy switches. Confirm the training controls covered earlier in this series are still where you left them: "Improve the model for everyone" in ChatGPT Data Controls, Model Improvement at claude.ai/settings/data-privacy-controls, and Keep Activity in Gemini Apps Activity. Also re-check the auto-delete period on the Google side, since the documented default is 18 months and a new account will not inherit whatever you chose on the old one.
4. Read the meters, not the invoices. Open the usage screen on each usage-billed product and compare this quarter with last. Remember the units differ: Zapier counts a task "whenever Zapier successfully completes a unit of work for you", while on n8n "An execution is a single run of your entire workflow." A rising execution count with flat revenue is the signal to investigate.
5. List every automation and name its last useful output. If nobody can say what a workflow produced in the last month, it is a candidate for deletion rather than repair.
6. Re-read one AI-drafted document end to end. Pick the customer-facing one you would be most embarrassed by. Prices, hours, service area and staff names go stale silently, and a document that was accurate in January is a liability in July.
7. Write four lines and date them. What changed, what you fixed, what you deferred, and what you will check next quarter. This is the only part that compounds.

Price cliffs deserve their own line in the diary

Two of the prices in the previous part of this series come with published end dates rather than open-ended terms. Microsoft's promotional Copilot rate of $18.00 per user per month is described as applying through September 30, 2026 and to the first year only, with automatic renewal at standard rates afterwards. Anthropic's Claude Sonnet 5 introductory API pricing of $2 and $10 per million input and output tokens is stated as running through August 31, 2026, after which standard pricing of $3 and $15 per million takes effect.

Neither of those will be announced to you as a price rise. Both are documented today. A quarterly review that catches them a month early is worth more than one that catches them on the invoice.

Retiring an automation is a better outcome than fixing it

The instinct in a review is to repair everything on the list. Resist it for the workflows.

Three questions settle it. Did this run at all in the last month? Did anyone act on what it produced? If it stopped tomorrow, would anyone notice within a week? Two nos and it goes. Automations accumulate because deleting one feels like admitting the time you spent building it was wasted, which is exactly the reasoning that keeps the maintenance burden growing every quarter.

The same test applies to seats and subscriptions, and it is faster to apply than any cost analysis. Cancel, then wait. If nobody raises it inside a month, you were right.

What a good quarter looks like on paper

The record you want after four of these is short and specific. A dated list per quarter showing which model you moved off and why, which setting had drifted, which automation you killed, and which price changed. Four entries a year, each half a page.

That document does two jobs. It tells you whether the stack is getting simpler or more complicated, which is the only trend that matters for a business without an IT department. And when a vendor retires something with 60 days of notice, or 2 weeks, you already know whether you use it, rather than starting the investigation from scratch while the clock runs.