Build so switching AI vendors costs you an afternoon, not a quarter

If your AI vendor raises prices or ships a worse model next quarter, moving should cost you an afternoon of config work — not a quarter of rebuilding.

The news, in one line

TechCrunch reported in August 2026 that new data indicates OpenAI is gaining on Anthropic among business users (TechCrunch, 20 August 2026). We haven't seen the underlying dataset, so take the share numbers as that reporting presents them and nothing more.

Our read (opinion, clearly labeled): the ranking moving at all is the useful part. If the buyers with the biggest integration bills are still comparing labs release by release, "pick a winner and stop thinking" isn't a strategy that's available to you either. The practical response isn't to guess which lab wins. It's to make the guess cheap to be wrong about.

Why a 12-person shop feels this harder than a 12,000-person one

An enterprise has an abstraction layer, a procurement team, and someone whose job is migrations. A small business has three things instead:

• Prompts that live in one person's chat history.
• Workflows built inside whatever agent builder the vendor shipped last month.
• An annual seat contract signed for the annual discount.

Each of those is fine on its own. Together they turn "we'd like to try the other model" into a multi-week project nobody has time for, which means you stop shopping, which means you pay whatever the renewal says. The lock-in isn't technical. It's that nobody can face the work.

Five things you should never let a vendor hold

1. Your prompts. Every system prompt, instruction block, and "here's how we talk to customers" rule belongs in a doc or repo you own, versioned, with a date on each change. If your best prompt exists only inside a vendor's assistant builder, you don't own your most valuable AI asset.

2. Your eval set. Twenty to fifty real inputs — actual customer emails, actual quote requests, actual messy invoices — each with the output you'd accept. In our view this is the highest-leverage artifact in a small AI stack. Without it, switching is a vibe check that takes weeks. With it, switching is: run the file, read the diffs, decide. That's the afternoon.

3. Your data and history. Conversation logs, uploaded files, the knowledge base you spent a weekend cleaning up. Set a recurring monthly export to your own storage. Before you sign anything, open the admin console and confirm you can pull a full export yourself, without a support ticket.

4. Your workflow logic. The if-this-then-that layer — routing, approvals, retries, where the output lands — should sit in your own automation tool or your own app, calling the model as a step. Vendor-native agent builders are great for prototyping and, in our experience, a bad permanent home for business rules, because the rules can't leave.

5. Your customer-facing endpoints. Phone number, email domain, chat widget snippet, webhook URLs. If the number your customers call is owned by the AI vendor, they own your churn risk too. Port numbers to a carrier account in your name and point the widget at your own domain.

What we'd actually build

For a typical small stack — a support inbox assistant, a quoting helper, a couple of automations — here's the shape we use:

• Model name lives in config, not in code. One environment variable, one place to change it. Sounds trivial; it's the difference between a config edit and a code hunt.
• One API surface for many providers. LiteLLM is open source, and its docs describe calling 100+ LLM APIs using the OpenAI input/output format (as of August 2026). A hosted router like OpenRouter puts many models behind one OpenAI-compatible API (as of August 2026). Either way your app speaks one dialect and the swap happens in the middle. Read each one's current pricing page before you route production traffic through it — self-hosting costs you infrastructure, hosted routers charge for the convenience.
• Lean on documented compatibility. Anthropic publishes an OpenAI SDK compatibility layer that lets code written against the OpenAI SDK call Claude by changing the base URL, API key and model name; Anthropic's own docs frame it as a way to test and compare rather than a long-term solution (as of August 2026). That framing is the point: the plumbing side of a swap is cheap when you've kept it thin.
• Evals as a file you can run. Even a spreadsheet plus a short script beats memory.

Our estimate, not a measured benchmark: for a stack that size, that's roughly half a day to a day of setup work.

A worked example (hypothetical — our own illustration, not a client)

Every number below is an assumption we invented to make the arithmetic legible. It is not a client result, not a benchmark, and not a quoted price. Check real seat prices at the vendors' own pages, linked in the next section.

A 14-person home-services company puts 10 people on a business AI plan. Assume, purely for the maths, $30 per seat per month: $300/month, $3,600 on a 12-month commitment. They also build their quote-drafting flow inside the vendor's agent builder, and the prompts live there.

Six months in, the other lab ships a model that's clearly better at their quote emails.

• Locked-in version: rebuild the flow in the new vendor's builder, re-derive prompts from memory, re-test by feel. Assume 30–40 hours spread over a few weeks, plus riding out the remaining $1,800 of committed seats. Realistically, they don't bother — and they don't push at renewal either, because they have no alternative priced out.
• Portable version: prompts in a doc, workflow in their own automation tool, 30 saved eval inputs. Change the model name, run the evals, compare, keep or revert. An afternoon. Seats are month-to-month, so the switch costs one month of overlap — $300 — instead of leaving $1,800 of committed seats on the table.

Swap in your own seat count and rate; the ratio is the argument, not the dollars.

Run the switch drill once a quarter

Block 90 minutes. Run your eval set against the comparable model at the other lab. Write down three things: quality diffs you'd care about, the current list price straight from the vendors' own pricing pages (OpenAI, Anthropic), and how long the swap would take today. Read those pages on the day — tiers and prices change, and a number in a blog post, including this one, goes stale fast.

Even if you never switch, you now walk into renewal knowing what the alternative costs and performs like. That's the whole point. Leverage is knowing you could move.

What lock-in is fine

Don't over-engineer this. If your team loves one product's interface, that's a real reason to stay, and a chat subscription your staff actually uses beats a perfectly portable one they don't. Deep integrations — a vendor's file handling, its meeting notes, its desktop app — are worth some stickiness.

The target isn't zero switching cost. It's a switching cost you could pay on a Tuesday afternoon without asking anyone's permission. Before you sign an annual deal, ask one question: if we had to leave in month four, what exactly would we lose, and how many hours would it take? If nobody can answer, go month-to-month until they can.

Find out what's actually worth automating first

Before you argue about which lab to bet on, it helps to know where AI would pay you back at all. Our AI Readiness & ROI Calculator is a free five-minute assessment that projects your ROI over 24 months — take it, then pick your vendor with numbers in hand.