Tech Retailer ON
ON: AI readiness
What separates an AI pilot that quietly stops being used from one that survives contact with a real process.
Published
September 22, 2026
Topic
Tech Retailer ON
Talk it through
Bring us the requirement and we will price it.
Almost every organisation now has an AI pilot somewhere. Far fewer have one that is still running twelve months later, because the pilot was measured on whether it worked rather than on whether anyone kept using it. Readiness is rarely about the model. It is about the data the model can reach, the process it is allowed to change, and whether the people doing that process were involved in designing it.
Three questions before you start
Before choosing a tool, we ask organisations to answer three questions honestly. The answers usually predict the outcome better than any pilot result.
- Is there a written process, or does the work live in one person's head? Automation of an undocumented process just makes the inconsistency faster.
- Is the data the assistant needs reachable, current and clean? If the answer lives in three systems and two of them disagree, the assistant will inherit the disagreement.
- Who owns the outcome when the assistant gets it wrong? Without a named owner, the first mistake ends the project.
Pilots fail on the last mile
The technical work is usually the shortest part of an AI project. The longer part is fitting the output into a process that already has quality checks, escalation routes and people whose performance is measured on the work you are changing.
- Build the exception path before the happy path, so unfamiliar input goes to a person rather than being guessed at
- Keep a manual review step for anything with a consequence until the output has been stable for a full cycle
- Measure time saved per week, in the team's own terms, rather than counting interactions
- Write down what the assistant may change and what it may only propose
A narrow assistant with a defined job and a clear boundary is trusted faster than a general one with no stated limits.
Where the infrastructure fits
Private assistants grounded in your own documents need somewhere to run. Usually that is a modest server on premises or a small cloud footprint, plus the identity and access tooling to make sure people only retrieve what they are allowed to see.
That is the point at which an AI project becomes an infrastructure project, and it is worth planning the hosting, access and backup before the first document is indexed rather than after.