How we deliver and govern AI
An AI feature is a piece of operations software with a probabilistic part in the middle. We treat it that way: a metric agreed before build, an eval set it must pass, a cost budget it cannot exceed, a gateway that keeps credentials and spend under control, a tenancy you own, and a person who signs off.
The seven rules
A metric is agreed and baselined before the first line of code.
- Every feature has an eval set of real cases
Run before release and after every model change
Cost per unit is budgeted; the feature falls back to a manual queue rather than overspend.
All model calls go through Bizmap's provider-agnostic gateway with budgets and attribution per client and project; models are swappable.
Inference runs in your tenancy or on the device; nothing trains a third-party model.
Every feature names the role that approves and documents what happens when the model is wrong.
- Unmeasured claims are labelled "expected"
With the method shown
How the work runs
Applied AI runs a short route of its own: a fit check, a pilot with a person at the checkpoint, then scale. The same gate language as an implementation.
- Fit check
- Pilot with a person at the checkpoint
- Scale
What this looks like in a project
- Week 1 to 2
fit check, baseline, data audit, eval set drafted.
- Weeks 3 onwards
build in two-week sprints; the eval set runs in CI.
- Before release
eval pass, cost review, approver training, fallback rehearsal.
- After release
monthly report on the metric, the cost and the exception rate; model changes go through the eval set first.
Our own use of AI
Bizmap engineers use coding agents through the same gateway, with budgets and attribution per project. It changes what a client gets: a first working demo in days, and tests and documentation that used to be cut from scope. It does not change who is responsible: architecture, data-model decisions, security review and every merged change stay with a named senior engineer. Client code never trains a third-party model.
Technology
- Bizmap AI gateway (provider-agnostic)