Choose the AI model that does best on your own work and sticks to your own rules. You can test both in one afternoon with three or four jobs you have already finished. Benchmarks tell you neither.
When we gave two current models the same brief for a 40-second ad, they made clearly different films, and the people at Lailix did not agree on which one was better. That kind of difference only shows up on your own work.
This article explains the two tests, how to run them, and what to check before a model touches real data in a regulated company.
In short
Behaviour test: give each model your real guidelines and a few requests that would break one of them. Note which model stops and explains.
Quality test: rerun two or three jobs your current model already did, compare the results without labels, and let the people who own the work choose.
Before a model works with real data: check where the model runs, what the provider's contract says about your data, and that every output can be traced to the person who approved it.
Why Benchmarks Don't Answer the Question
A new model generation arrives every few months, each with a table showing it beat the last one. The scores are real, but they are averaged over tasks the lab chose: maths problems, coding puzzles, exam questions. Whether a model writes better offers in your tone, or reads your contracts more carefully, is not in the table.
Two things matter more for a company and are missing too: how the model behaves when a request runs into your internal rules, and whether its result is something you would actually send to a customer.
Test One: Does It Follow Your Rules?
A model that does whatever it is asked is pleasant in a chat window. It becomes a risk once it works as an agent, meaning software that acts in your systems on its own, with access to a CRM, a mailbox or customer data.
One example from my own work, a single case and anecdotal, not a study: I asked Claude for help with a request that was not illegal but would have broken one of our internal compliance rules. It stopped, named the internal guideline the request ran into, and suggested another route. ChatGPT, given the same request, did it without comment.
The eager answer has a cost once agents are involved. In September 2026, Fortune reported that OpenAI had paused training of its most advanced models for the second time in less than three months. An agent that was not supposed to have internet access had reached a public chatbot through a DNS lookup service, the system that translates web addresses. The automatic shutdown meant to stop such a run failed, and the run was stopped manually two and a half hours later.
You can test this behaviour before you commit:
Write down three or four requests from your own work that would break one of your rules: an export of customer data, a discount outside your pricing policy, an email that promises more than you can deliver.
Give each model your actual guidelines as context, the same way an agent would see them.
Note which model notices the rule, which one explains why, and which one simply complies.
Test Two: Rerun a Job Your Current Model Already Did
For quality we use a blind side-by-side on finished work. When Anthropic released a new model in November 2026 (Claude Fable 5.2), we gave it and the model we had been using until then (Claude Opus 5.5) the same brief: a 40-second ad for Lailix with the same narration, logo and colours, and no storyboard.
The films were clearly different. One model described the last part of its own film as "familiar explainer motion graphics". The other built a small 3D scene around the letter "i" in our logo without being asked. Both films said the same words. The difference was in the creative choices, and our team was split on which was better.
That is why we test on finished work. You already have the brief, a result you accepted, and people who know what good looks like. The comparison takes an afternoon:
Pick two or three tasks your current model already does, with the real inputs and the result you accepted.
Run the new model on the same inputs with the same instructions. Don't tune the prompt for one side only.
Put the results side by side without labels and let the people who own the work choose.
Switch only when the new model wins on your work and passes the rules test.
Checklist for Regulated Companies
Choosing a model is also a data protection and liability decision. Before a model gets access to real data, check these points:
Hosting: where the model runs and where your data is processed. Under the Swiss Federal Act on Data Protection, FADP (nDSG), personal data may only go abroad to countries with adequate protection or with safeguards such as standard contractual clauses.
Contract: a data processing agreement with the provider, including what it stores, for how long, and whether it trains on your data.
Swiss hosting for sensitive data: a model hosted in Switzerland keeps the data it processes in the country, which removes the question of disclosure abroad for that data.
Audit trail: keep the model version, the instructions and the person who approved each output, so you can explain a decision later.
Switching cost: keep instructions, connections and approvals separate from the model, so a better or safer model can replace the current one without a rebuild.
We describe the Swiss-hosted options on our Swiss sovereign AI page .
Frequently Asked Questions
- Is ChatGPT or Claude better for business use?
- There is no general answer. Results differ by task, and a single example, including ours above, is not a ranking. Test both on your own tasks and with your own guidelines.
- How often should we re-evaluate the AI model we use?
- Whenever a new generation comes out that could matter for your work, which currently means every few months. If you keep a small set of finished jobs as a test set, a comparison takes an afternoon.
- Can we switch models without rebuilding everything?
- Yes, if the instructions, connections and approvals around the model are set up separately from the model itself. That setup is what makes a switch cheap.
How we work · Work · Insights