A testing log for leadership and teams

AI Watch: what we tested, and what it changes for you

Kailli's AI Watch is not a news feed. When a model, a tool or an agent comes out, we test it on real projects, then publish a short note: what we tested, the verdict, who it's for and what you can ignore.

Book 20 minutes See the testing method

01The method

How we test a new AI release before writing about it

A new release only makes it into the Watch if it has been tried on real work. The vendor's press release and public benchmarks are not enough.

Observe

Every day we follow releases from vendors (Anthropic, OpenAI, Google, Mistral, Microsoft) and from everyday tools. We keep only what can affect a business task.

Prove

We replay real tasks on our own projects: writing, document analysis, code, agents. We look for where it breaks, not for the demo that works.

Compare

We put the new release up against what you already use, on the same tasks: quality, time, cost, data settings.

Share

We write a dated note with a verdict. What holds up goes into our recommendations, our training courses and the support we give our clients.

02The verdicts

Proven, Compared, Optimized: what each verdict means

Each note carries a single verdict. Proven, Compared and Optimized are the three gates a release that holds up goes through; "Not adopted" covers those that don't make it.

VerdictWhat it meansWhat you do with it
Not adoptedThe test reveals a limitation that prevents serious business use.You can ignore it for now.
ProvenIt holds up on real tasks, without yet beating what already exists.Keep an eye on it, change nothing.
ComparedIt beats the tool in place on at least one specific task.Try it on that particular task.
OptimizedWe found the settings and the way of working that get the best out of it.Ready to enter your workflows, along with the method.

03The format

What each AI Watch note contains

A note takes three minutes to read. It always answers the same questions, in the same order.

The standard sheet

Tool tested
Name, vendor, version and plan used.
Test date
The day we ran the test, not the day of the announcement.
What we tested
The tasks replayed and the projects used, described without client data.
Verdict
Not adopted, Proven, Compared or Optimized, with the reason in one sentence.
Who it's for
The teams and tasks for which it makes a difference.
What you can ignore
Launch promises that don't hold up, or that don't concern you.
Going further
The Kailli guide or offer that extends the note.

Publishing schedule: irregular, as real tests happen. One note per release worth testing, never one note per announcement.

04The notes

The latest AI Watch notes

Notes are listed from newest to oldest. Each one keeps its test date: AI moves fast, and yesterday's verdict can change.

  2. Jev: the AI that picks the right model for each task

    Test date Compared

    Compared by the Kailli team against an AI assistant on its own, from September to October 2026: Jev wins when it chooses between clear options, and loses when it has to search.

Frequently asked questions about the AI Watch

What is a tested AI Watch?

It is a watch that doesn't relay announcements: each new release is tried on real tasks before we comment on it. The note says what was tested, the verdict and who it's for. You know what changes for a business without having to try everything yourself.

How do you test new models?

On our own projects, by replaying business tasks: writing, document analysis, code, agents. We then compare with the tools already in place, on the same tasks. No client data is used in the published tests.

How often do you publish?

There is no fixed schedule. A note is only published once a real test has been done: some weeks there are several, other weeks none.

Can I receive the AI Watch by email?

Not for now: there is no newsletter. Notes are published on this page, and our recommendations are shared with you directly if we are working with you.

Does Kailli sell the tools it tests?

No. Kailli sells no tools: the Watch is there to help you choose, not to sell. That is what lets us publish a "Not adopted" verdict when a test doesn't hold up.

Twenty minutes. And you'll know whether we can help.

No sales pitch. You tell us where you stand with AI, we tell you how far you can go.

Book 20 minutes Prepare for the call