How do we keep it right?
AI that improves with use.
Evidence from everyday use helps your team decide what to improve next. Changes are tested and released through agreed controls.
A live system needs someone checking its answers.
It went live, the project team moved on, and nobody answers for what it says. A system can run like that for years.
Operate keeps a system right, and improving, once it is live.
What changes under a live system
- Suppliers
A supplier changes a material.
- Sensors
A sensor drifts out of calibration.
- People
Experienced people leave and their successors decide differently.
- Models
A vendor retires a model and its replacement behaves differently on the same inputs.
Any one of them can make a working system wrong.
We watch every live system on the same standing measures.
There are two kinds of problem, and they’re handled differently.
When the software fails
An outage, an error, a broken connection to a source system. Normal monitoring sees it, and the team that built the system puts it right.
When the answers are wrong
The system is running and the answers have drifted. That shows only in overrides, deferrals, evaluation scores and what users report, so it goes to review with the people who own the outcome.
An example planning service
Live
- Evaluation score Holding against the reference cases your experts judged Steady
- Drift Inputs shifting as a new product family ramps up Watch
- Override rate A pattern on that product family, sent to review with its owner In review
- Deferral rate People hand back the same share of cases as before Steady
- Business outcome Hours returned, reported against the baseline agreed at the start On track
Change log Model provider update tested against the reference cases before release.
What Operate covers.
It starts before go-live. From the first trial, we record where people and the system disagree.
-
Close support at handover.
For an agreed period after rollout, while your teams take ownership.
-
Watch the standing measures.
Evaluation score, drift, override rate, deferral rate and the business outcome, reported in your own measures.
-
Treat corrections as evidence.
When people override the system, we review the pattern with the people who own the outcome.
-
Change it under control.
Feedback never becomes a production change by itself. Each change is evaluated, reviewed and released through the controls agreed for the system.
-
Never a silent model swap.
When a provider changes or retires a model, the replacement is tested against your reference cases before it goes in.
-
Find the next challenge.
The record of overrides shows where the work still waits, and that’s where the next build starts.
Operate runs on Continuous Intelligence Operations.
Every system we run under Operate is kept to its discipline: people and agents share the work within agreed limits, important decisions are kept with their reasons, and that record decides what changes next.
What people ask about life after go-live.
Isn’t this just support?
No. Support fixes failures. Most of the work is the other kind of problem: answers that drift while the system keeps running, which need review by the people who own the outcome.
How are changes approved in a regulated setting?
Through the controls agreed for the system, alongside your own change control. Each change is tested against the reference cases before it goes live, and it can be withdrawn when the evidence no longer supports it.
What if a system is no longer worth running?
We review it with you, then improve it, move it to other work or retire it.
Can our own teams run it?
Yes, by agreement: on your own with our support, or jointly with us. Either way, the record of how your operation decides belongs to you.