Build Workflows That Survive When Your AI Vendor Doesn't

If a workflow you rely on only works because of one AI vendor's specific product, pricing, or leadership staying exactly as it is today, you have a dependency you haven't priced in. That's not a hypothetical this year — it's happened four separate ways, to four different vendors, for four different reasons.

Four incidents, one mechanism

In August, Google's DeepMind CEO moved to chairman on the same day four senior technical leads left to start a rival — including the engineer who led Gemini's technical direction. Nobody's workflow broke that week. But anyone who'd built a process around "however Gemini currently behaves" had just watched the person who shaped that behaviour walk out the door.

In July, Anthropic drew an ethical redline on weapons and surveillance and lost government access as the direct consequence — Fable 5 was suspended for nineteen days with no restoration timeline. The same month, both OpenAI and Anthropic had their best models gated by government approval rather than a pricing tier or a waitlist, for the first time simultaneously. And in May, an analysis of Anthropic's and OpenAI's own numbers showed both companies are pricing subscriptions well below what the compute costs to run — a gap that closes eventually, and GitHub Copilot's move to usage-based billing was the first crack.

None of these four coordinated. A leadership departure, a policy dispute, a regulatory gate, and a pricing model under strain are four different failure modes. Read together, they say the same thing four times: the vendor you chose is not a fact about the world. It's a decision you made, under conditions that already changed once this year and will change again.

What to build instead of what to worry about

The fix isn't switching vendors or hedging with three parallel subscriptions nobody has time to maintain. It's knowing the answer to a small set of questions before you need it under deadline pressure, not while you're finding it out:

Which steps are actually tied to this vendor, and which just ended up there? Some tasks genuinely need a specific model's context window, reasoning depth, or output format. Most don't — they landed on whichever tool was already procured. Separate the two lists before a disruption forces you to do it in a hurry.

Does every AI-dependent process have a tested fallback? Not a theoretical one — an actual second option you've run once, with the same prompt templates, so a switch is a config change instead of a rebuild.

What would this cost at three times the current price? Subsidised pricing is a phase, not a floor. Knowing which of your AI uses you'd keep at 3x and which you wouldn't means you make that call deliberately instead of by surprise invoice.

Could you explain your vendor choice to someone who asked "why them"? Procurement, legal, and clients now ask this question. "It tested well" stopped being a complete answer the week a vendor's ethics became a documented, public position.

An HR lead at a 90-person nonprofit

She'd built the organisation's grant-application first-pass review around one vendor's API two years ago — mostly because it was the tool a previous hire had already set up, and nobody had revisited it since. After reading about this year's run of vendor disruptions, she spent an afternoon actually tracing the workflow instead of assuming it was fine: the step that summarised long funder guidelines into checklists turned out to need the vendor's larger context window, and stayed. The step that drafted boilerplate acknowledgement emails didn't need anything special — she tested a second, cheaper model against it and got identical results. That step now runs on whichever tool is cheapest that quarter, and switching costs her ten minutes instead of a rebuilt pipeline.

A finance manager at a 25-person creative agency

He'd modelled next year's budget assuming AI subscription costs stayed flat at current per-seat pricing, because that's what every tool had charged for the eighteen months his team had used them. After he read that the labs themselves are losing money on those subscriptions, he rebuilt the model with AI costs at double the current rate — not because he expects that exact number, but because "flat forever" was never a real assumption, just an unexamined one. The revised budget didn't change what his team uses today. It changed what he'll say in the next pricing conversation, because he's already done the arithmetic instead of doing it live in the meeting.

The one thing

None of these four incidents were the same kind of disruption, and the next one won't be either — which is exactly why the fix isn't predicting which failure mode comes next. It's building workflows where the vendor is a swappable component, not the foundation.