1 — The first draft
Generate test cases from a specification or a story. The benefit is not the quality of the text, it's no longer starting from a blank page.
- Proposed load cases and boundary conditions
- Mandatory human proofreading, always
AI applied to testing
We are running it in delivery for two years, not just as a demonstration. What has changed is not the nature of the job: it is the volume a consultant covers in a day.
Make an appointmentWhere she reports
Generate test cases from a specification or a story. The benefit is not the quality of the text, it's no longer starting from a blank page.
Fixing a selector that shifted after a screen redesign. It's the most thankless task in automation, and the most mechanical.
Differentiate application faults from unstable environments, and group fifty failures that share the same cause.
Comparing thousands of values against a reference document—a brief, a catalog, an interface contract.
What she does not do
Three judgments remain beyond his reach, and they are the ones that commit.
What deserves to be tested. Weighing whether to cover a payment flow or an administration screen requires knowing what the company loses in each case. No model knows this figure.
What an anomaly costs. A display issue on an internal screen and the same issue on a customer kiosk do not carry the same weight. Prioritization is a business decision.
If a version goes to production. It is a commitment, with an identified person in charge. It cannot be delegated to a probability.
Let's add a more prosaic limit: a model produces the plausible. On a test, the plausible that is false is worse than nothing — it creates a trust that no human is going to reopen. Hence our rule: everything produced by the AI is reviewed, and everything that fails is checked by hand before being declared a defect.
In practice
What this actually produces: coverage that goes from a sample to the totality. What we were verifying on a few restaurants, we verify across an entire network — not because the tool is smarter, but because it doesn't get tired and doesn't skip line 431.
Frank questions
No — and we are well-placed to say so, since we have been running it in delivery for two years. It writes a first draft, maintains selectors, sorts failures, and compares in bulk. What deserves to be tested, what an anomaly costs, and whether a version goes into production remain matters of judgment. What has changed is the volume covered.
Not without your explicit consent, and never by default. Depending on your policy, we work with anonymized games, on a model hosted on your servers, or without any AI at all for sensitive areas. This is a decision that must be made at the outset, not along the way.
By the same proofreading as for a handwritten test: is it replayable, is its answer binary, does it test something valuable? A generated case that fails these three questions is deleted, not kept because it's free.
Through anchor maintenance and failure triage: two invisible tasks for the business, with no risk to quality, that free up time starting the first week. Test generation comes afterwards, once review is in place.
A no-obligation audit of your testing process to find out where you really stand.
Make an appointment
English 
French 
Hebrew