From assistant licences to governed agents
Year:
2026
Service:
AI Strategy & Adoption
Industry:
Manufacturing
Team:
4 specialists, 12 weeks
Reference scenario: a UK industrial manufacturer moves from under-used assistant licences to role-based enablement, AI literacy measures for its EU plants and three agent workflows that act only behind approval gates. Adoption is measured in outcomes.
Introduction
This scenario follows a UK manufacturer of industrial equipment with plants in England and two in the EU. A year ago it bought assistant licences for 1,400 office and engineering staff. Telemetry shows that about a third use them weekly, mostly to summarise email. Meanwhile engineers paste supplier specifications into personal tools, and the EU plants ask what the AI Act expects of them.
The chief operating officer does not want another training day. The questions are whether the licences change how work gets done, and whether the company is ready to let agents take steps on their own. Keter AI runs a twelve-week adoption programme that answers both: enablement by role, measured in outcomes, and a first set of agent workflows that act only within approved limits.
Challenge
Licences were bought as a product. Adoption has to be run as a change programme.
Usage is not value. Self-reported time savings have not become capacity. Public evidence sets a realistic bar: a UK government trial of Microsoft 365 Copilot with over 20,000 civil servants reported 26 minutes saved per person per day. That has to be earned workflow by workflow.
One course for everyone. A maintenance planner, a quality engineer and a buyer need different skills. A single webinar serves none of them.
AI literacy is a legal duty in the EU plants. Article 4 of the AI Act has applied since 2 February 2025. As reworded in July 2026, it requires deployers to take measures to support staff AI literacy, with no duty to guarantee a specific level. The UK has no AI-specific statute, so its sites follow company policy and existing regulators.
Agents without limits. Teams have begun wiring assistants to the ERP through shared service accounts, with no approval step.
Shadow tools. Personal accounts stay in use alongside the official assistant.
Solution
The programme measures before it teaches, and teaches before it automates.
Measure cohorts, not opinions. Five role cohorts start in waves, each with a control group. We track outcome measures per role, such as time to prepare a work order or to close a non-conformance report, alongside usage telemetry.
Teach by role, record the measures. Each cohort gets a short curriculum, a library of prompts and skills for its own tasks, and a champion on site. Attendance and materials are recorded as the company's Article 4 measures for EU staff.
Agents only for bounded steps. Public research on frontier models shows that the tasks agents complete reliably are far shorter than headline figures suggest. Each workflow is therefore cut into short, checkable steps, with a human approval before anything is written to a system of record.
What is built:
Role curriculum and skill libraries for maintenance planning, quality, procurement, finance and plant management.
Governed connectors - the licensed enterprise assistant (for example Microsoft 365 Copilot or Claude) reaches ERP, quality and maintenance systems through MCP servers with scoped, per-agent permissions, behind SSO and data loss prevention controls.
Three agent workflows - supplier quote comparison, non-conformance triage and maintenance work-order preparation, each with durable workflow state and an approval gate.
Evaluation harness - a task set per workflow that must pass before release and after every change, with tracing in MLflow.
Adoption dashboard - usage and outcome telemetry by cohort.
Acceptable-use policy and a route for bringing shadow tools inside it.
Result
This is a reference scenario, so the figures below are targets and acceptance criteria agreed at the start, not measured results.
Weekly active use - target: 70 percent of licensed staff in the trained cohorts by week twelve, up from about a third.
Outcome per role - target: a measurable reduction in preparation time for each cohort's chosen task against its control group; cohorts that show none are redesigned or stopped.
Approval coverage - acceptance criterion: every agent action that writes to a system of record passes a human approval and is logged.
Agent identity - acceptance criterion: no shared service accounts; each agent has its own scoped permissions.
Literacy record - acceptance criterion: documented Article 4 measures for all EU staff who use AI systems.
At the end the manufacturer owns the curriculum, the skill libraries, the champions network, the three workflows with their evaluation sets, and the dashboard. Its own teams can add a fourth workflow by the same rules.
