Model-Agnostic by Design: Big AI Bets Without Vendor Lock-In

AI vendor lock-in is an architecture choice, not a vendor trick

The email lands on a Tuesday. The AI model under your lead-scoring workflow is being retired. You have a few months.

One team reads it, swaps in the new model, runs its tests and is done by Friday. Another team starts a project that eats a whole quarter.

Same vendor. Same email. The difference is what each team owns.

That difference has a name: AI vendor lock-in. It means your work is built so tightly around one model that changing it costs months.

Being model-agnostic is the cure. It does not mean using every model. It means you can swap the main one in weeks without rewriting your workflows.

Nobody traps you into lock-in. You move in, one handy feature at a time: the vendor's memory, its hosted agents, its file store.

Each one is small. Together they tie you to one vendor. Call it a contract you never signed.

Just picking your first AI tool? Start with AI marketing workflows and come back when one runs every day.

You will change models even if you never change vendors

Most teams think lock-in only matters if they leave their vendor. That is the wrong worry.

A market survey by Menlo Ventures found only 11 percent of teams changed model providers in a year. But 66 percent moved to a newer model from the vendor they already had.

Every one of those upgrades is a swap. New model, same prompts. And a real chance that what worked last month now works a little differently.

Then there is the calendar. Vendors publish model retirement policies, and every model has an end date.

Loyalty does not save you. The calendar decides.

Sometimes nobody even sends the email. Chip Huyen warns in AI Engineering that a vendor can update the model behind an unchanged name, without telling you.

Your prompts drift off course, and you have no idea why. She calls building on these models riding a bullet train. You do not pick the stops.

The same model update can help one company and hurt another

Here is the part most advice skips. You cannot borrow anyone else's test results.

Huyen's book reports one update, from an older version of a model to its newer version. It cut one company's intent sorting by 10 percent. It improved another company's support chatbot.

Same update. Opposite results.

Outside researchers saw the same pattern. Two releases of one popular model, three months apart, scored very differently on the same tests.

The model is not the only moving part, either. The same model served by two providers can score differently, because each one tunes it for speed and cost.

So what does a leaderboard tell you? Which model is good on average. Not which one is good at your job.

Your tests tell you, or your customers do. Pick one.

Own four things and rent the model

Own four things. Rent one. That is the whole strategy.

Own the harness. The harness is the code and rules around the model: tools, permissions, checks and recovery steps. Write it against your own interface, so the model is a part you can pull out.

More on that layer in what an agent harness is.

Own the prompts. Keep them in your own files with version numbers, like code. Not inside a vendor's console.

Huyen argues that secret prompts are a burden, not a moat. Every model change means updating them. So we put the value in the system around them.

Own the golden set. A golden set is a list of real cases with known good answers. It is what makes switching cheap.

Own the data. Your documents, search index and logs live in systems you control. If your data sits in ten tools, start with one company brain.

Say you cut your documents into chunks with one model's token counter, its own way of splitting text. Switch models, and you rebuild the whole index. Small choice, big bill.

Rent the intelligence. The model is the part you pay for by use, so keep it cheap to leave. On purpose.

Cheap for you to leave. Expensive for the vendor to lose. That is where you want to sit at renewal time.

Put one door in front of every model

The door is one internal gateway. Every AI call in the company goes through it. Nobody calls a vendor directly.

What does the door do? First, it gives every workflow one way to call any model. When a vendor changes its interface, you fix the door once, not every app.

It holds the keys. Staff get access to the door, not copies of the company's vendor keys scattered across laptops.

It sets limits per person and per app, and it logs every call. When one vendor is down or slow, it sends the request to a backup model.

Software teams know this shape. In Team Topologies, Matthew Skelton and Manuel Pais describe a platform that can be one thin layer over a vendor's product. They expect platform teams to swap parts for better outside options as the market moves.

The Accelerate research team found a related rule in software delivery. The biggest driver of fast, safe releases was whether a team could change and test its own system without waiting on other teams.

They also called the which-tool debate the wrong question. We read the AI version the same way. The model you pick matters less than whether you can swap it alone, on a normal workday.

Tool connections can use an open standard like the Model Context Protocol instead of one vendor's format. That is cheap portability.

How do you know a swap is safe? Build the golden set first

A swap is safe when your own tests say so. Not the vendor's launch post. Not a leaderboard.

John Berryman and Albert Ziegler, in Prompt Engineering for LLMs, describe a product team whose very first code was its test suite. Before the prompts. Before the screens.

That suite told them whether each later change helped, hurt or did nothing. It set the pace for everything after.

Start small. Pull 20 real cases from last month: real leads, real tickets, real briefs. Write down what a good answer looks like for each one.

For a model swap, test the whole workflow, start to finish. A new model touches every step, so a regression test across the full chain catches what breaks between steps.

How many cases is enough? It depends on the gap you want to see. A rough rule in Huyen's book: each time the gap shrinks by three, you need ten times the cases.

So 20 cases catch a big, obvious break. Trusting a 10-point difference takes about 100.

One trap. Never let your current model write the test cases. Berryman and Ziegler warn that tests made by model A tend to favor model A.

No golden set? Then model-agnostic is a slogan, not a plan. Our guide to evaluating AI agents shows how to grade the answers.

Multi-model means a tested backup, not a collection

A good multi-model setup is boring. Boring on purpose.

One main model does most of the work. One backup stays warm, tested against your golden set every month. Small, cheap models handle high-volume, low-stakes jobs like tagging and sorting.

Routing rules send each job by cost and stakes. Keep the router itself small and cheap, or it eats the savings.

Big companies are piling on models anyway. In an Andreessen Horowitz survey of 100 enterprise tech leaders, 37 percent used five or more models, up from 29 percent a year before.

The same survey found switching was getting harder as AI took on longer, more complex work. More models. Harder exits.

How to lock yourself in, in four easy steps:

1. Paste your prompts straight into the vendor's console. Never save a copy.

2. Give every team its own copy of the vendor key.

3. Use the vendor's built-in memory as your only customer record.

4. Skip the tests, because the new model is surely better. Then ask customers what the hell changed.

Collecting models is a hobby. Routing them is a business. An alternative you have never tested is a slide, not a strategy.

Sign vendor contracts as short as the market moves

Write vendor deals so you can walk away. Annual terms, not five years. Then get three things in writing.

One, your data is not used for training. Two, you can delete it. Three, prices are set per unit of use, not tied to exclusivity.

Put those terms in your AI governance rules so nobody signs around them.

Why so short? Because the leader changes. In the Menlo survey, the vendor with half of enterprise usage at the end of 2023 held a quarter by mid-2025, and a rival had taken first place.

Nobody signing a long deal back then saw that coming. So do not try to call it. Plan to be wrong about who is best.

Take partner credits when they are free. Turn them down when they tie your build to one vendor's format.

Free credits are fine. Free handcuffs are still handcuffs.

Clayton Christensen's team gives the deeper reason in Competing Against Luck. The job a customer needs done barely changes, while the ways of doing it keep changing.

This next part is our reading, not theirs. Write down the job each workflow does. Treat the model as today's tool for that job.

Run the Swap Test once a quarter

The Swap Test is our quarterly check of AI vendor lock-in. It turns a vague worry into a number of weeks.

Pick one live workflow. Route it through the door. Run the golden set on your current model and on one candidate, then compare the answers.

Time the whole thing, from start to decision. Under two weeks means you are portable.

Over a quarter means you are locked in. Something you should own has slipped into the vendor's hands.

Then find the layer that slipped: the harness, the prompts, the golden set or the data. Pull that one back first.

Each quarter, also ask three drift questions. Did our prompt change without anyone noticing? Did users change how they ask?

Did the vendor change the model under the same name? Rerun the golden set every month and you will know.

Is this extra work? Some.

But the swap is coming either way. The test just lets you pick the date.

Run the Swap Test with AI: the copy-paste prompt

Paste the prompt below into your preferred AI. It asks six questions about one workflow.

Then it scores each owned layer and plans your first Swap Test. It never invents numbers.

You bring the facts. It brings the method.

What to do this week

1. List every AI workflow you run, with the model and the vendor-only features each one uses.

2. Pick the one that would hurt most to lose. Pull 20 real cases and write the good answer for each.

3. Put that workflow behind one door. Run its first Swap Test against one other model.

Back to that Tuesday email. Next time it lands, you forward it to whoever owns the door and go to lunch.

More models bring a bigger bill, so read how to control AI token spend next. Then decide who owns AI, including the door.

Marry the job. Rent the model.