Switching AI models and avoiding vendor lock-in
By the ChamberOne team 4 minute read
A firm that picked an AI tool two years ago may already be using a model that its provider has retired or replaced. A firm that picks one today should assume the same will happen again. The sensible question is not which model is best this month. It is how hard it would be to change. Lock-in rarely arrives through one dramatic contract clause. It builds up through habit: prompts saved in someone's account, outputs in a proprietary format, workflows that only work with one provider's quirks. This article lists where lock-in hides and how to keep your options open.
Key points
- Assume any model you adopt will be changed or retired, and plan for it.
- Store prompts and test samples as firm property, with an owner and version.
- Keep documents and accepted work product outside any single AI provider's screens.
- Prefer platforms where you can choose the model per feature and switch AI off.
- Run a short comparison drill each year so that switching has a known cost.
Why models change under you
Providers release new versions, retire old ones and change prices and terms. Sometimes a new model is better at one task and worse at another. Sometimes it behaves differently for the same prompt, so a workflow that was reliable in March needs retesting in June. None of this is a scandal. It is how the technology currently works.
For a firm, the practical consequence is that any AI-assisted process has a shelf life. You can either plan for replacement or be surprised by it.
Where lock-in hides
When leaders think of lock-in they usually think of the contract. In practice, the contract is only one of five places.
- Prompts and templates: instructions your best people refined over months, stored only inside one tool.
- Outputs and history: summaries, analyses and annotations that cannot be exported in a usable form.
- Workflows: steps designed around one model's strengths, such as a long context window or a particular output format.
- Integrations: connections to your document store, email or billing that only the vendor can maintain.
- People: staff trained on one interface who resist any change.
Keep prompts as firm property
Treat prompts like precedents. A good prompt for summarising a lease or checking a contract against a playbook is knowledge the firm has built, and it should live in a place the firm controls, with an owner, a version number and a note of what it was tested on.
Write prompts in plain language, without relying on tricks specific to one model. Keep a handful of sample inputs with the output you judged acceptable. When you change model, run those samples through the new one and compare. That is a small, repeatable test, and it is more useful than any vendor claim about quality.
Keep data portable
Your documents, matter records and notes should not depend on any one AI provider. The AI layer should read from your records and write drafts back to them, not become the only place where your work product exists.
Ask whether the platform stores documents in ordinary formats, whether you can export a matter with its notes and history, and whether AI-generated content is clearly separated from content a lawyer accepted. If an output is only visible inside the vendor's screen, you have a dependency.
- Store documents in ordinary formats
- Be able to export a matter with its notes and history
- Keep AI-generated content separate from accepted content
- Let the AI layer read from your records and write drafts back
- Make the AI provider the only home of your work product
- Leave outputs visible only inside the vendor's screen
Choosing platforms that let you swap models
A platform that lets you choose the model for each feature, or change it later, gives you bargaining strength when terms change. It also lets you match a model to a task: a careful, slower model for a contract review, a cheaper one for sorting enquiries.
Questions to put to any vendor:
- Which models can we choose from, and who decides when one is retired?
- Can we change the model for one feature without changing others?
- Can we switch AI features off for the whole firm?
- What do we have to retest, and who does the retesting, when a model changes?
- Where is our data processed when we choose a different provider?
- Do we get notice before a model, a price or a data term changes?
A small switching drill
Once a year, run a drill. Suppose a three-lawyer firm uses AI for document summaries and matter briefs. The knowledge lawyer picks ten representative documents, runs them through the current model and the alternative, and compares the outputs against the originals. A second person scores each for accuracy, omissions and tone. The firm records the result, the date and who did it.
The drill does not need to end in a switch. It shows what switching would cost, and it keeps the firm able to react if a provider changes terms or has an outage. It also surfaces assumptions that nobody knew they had made.
- 1Pick ten representative documents
- 2Run them through both models
- 3A second person scores each output
- 4Record the result, date and who did it
- 5Repeat next year
- Then back to step 1.
Who decides, and when
Put one person in charge of model changes. That person does not need to be technical. They need authority to say 'we test before we switch' and the habit of asking vendors for notice. Without an owner, the change happens when the provider decides, and the first you hear of it is a lawyer saying that the summaries look different.
Keep a short register: which features use AI, which model each uses, the date it was last tested and who tested it. A single page is enough. When a provider announces a retirement date, the register tells you in minutes what is affected. It also helps when a client or insurer asks which tools touched their work.
Budget a little time for each change. Retesting ten sample documents and updating the register is an afternoon, not a project, but only if someone has been asked to do it.
What switching will not fix
A new model will not repair a vague process. If your firm has no standard for what a good summary looks like, every model will disappoint in a different way. A switch also brings new data-handling questions, since a different provider may process data in a different place and on different terms. Check those terms before moving, not after.
Finally, remember that the lawyer's duty to review does not change with the model. Whichever one you use, a person must check the output before relying on it.
How model choice works in ChamberOne
- You choose the model for each AI feature, GPT-5 or Claude, or switch AI features off for the firm.
- Each AI output stays a draft until someone accepts it.
- Every AI call is recorded in the audit log with the model and token counts.
What it does not do
- AI output can be wrong or incomplete and does not replace a lawyer's review.
- When AI features are on, the text needed for a request is sent to the AI provider for that feature, outside the country where ChamberOne is hosted.
- It does not retest your prompts for you when you change model.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
How a small law firm can start with AI without a big project
Pick two low-risk uses, set review rules and measure before expanding, with a list of what to postpone.
5 minute readA 90-day AI pilot plan for a law firm
A week-by-week plan to choose a use case, set measures, run a small trial, review errors and decide.
5 minute readChoosing AI models and firm-level controls: what a practice should decide
Model choice, an off switch and a usage record are three decisions a firm should make deliberately.
4 minute read