Decide · AI trial sprint
Two weeks to test one AI process on your own data
You get a written yes or no, and you know what the next step means before you commit.
Two weeks, one process, your own data. A working prototype measured against known answers, and a written yes or no.
- Who it is for
- A company with one process to test, and its own data to test it on.
- Length
- Two weeks
What you get
A working prototype on your data
Built in a sandbox, for one process and one data source.
A test set of known answers
20 to 50 real examples with known correct answers. The test set is yours.
Measured results
Accuracy against the threshold you agree on day 1, plus latency per call.
A production plan for the next step
For the production sprint, with scope and approach agreed in advance.
A written yes or no on day 10
A decision meeting with a written report: what works, what fails and why, and the data gaps.
Everything handed over
The code, the prompts and the test set are yours.
How the two weeks run
Days 1 to 3: scope and test set
Kick-off on the process, the success threshold and the risks. Data access is set up, and the test set is built with your process owner.
Days 4 to 8: prototype and measure
A first runnable version on a hosted model, then iterations on retrieval, prompts and guardrails, with eval runs and a mid-point check-in.
Days 9 to 10: plan and decide
The production plan and the basis for the next step, then the decision meeting: demo, results, and a written yes or no.
What we need from you
- A process owner: available on day 1 and at the mid-point check-in, and about 30 minutes a day in between.
- One data export by the end of day 2: a file or a read-only share from one source. No system integration.
- 20 to 50 real examples with known correct answers by day 3.
- A decision-maker at the day 10 meeting.
- An environment: our sandbox, or your cloud account (AWS, Azure, GCP or GleSYS).
Not included
- Production deployment, and integrations with ERP, ticket or CRM systems.
- Authorisation model, automatic gate (CI gate), monitoring and operating guide (runbook). These belong to the production sprint.
- Write actions by agents. The trial is read-only.
- More than one process or data source.
- User training, a security audit, or fine-tuning. See private model and fine-tuning.
What you keep
A written yes or no, with the evidence behind it, and a plan for the production sprint.
Questions
- Why one process?
- The trial covers one process, one data source and read-only access. Anything wider is scoped for the production sprint.
- What if the answer is no?
- You still get the written report: what works, what fails and why, and the data gaps.
- How is success measured?
- You and your process owner agree the success threshold on day 1, so the yes or no rests on measured results.
- How is the commitment set?
- You get a written scope after the call. No hourly work without a written cap.
- Are you a data processor?
- When we process personal data on your behalf we sign a processing agreement. It is part of the written scope.
- Where is data stored?
- The trial runs in a sandbox or in your cloud account. You choose the account. An EU region is possible when the provider offers one.
- Does it work with our business system?
- The trial reads a file or a read-only share. A connection to the business system belongs to a later step, if the measurement holds.
Context
at least 30%
In 2024 Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025.
Next step
Bring one process.
30 minutes. An honest answer on whether AI fits.