Deploying reliable AI agents

The move from assistant to agent is the point at which AI stops being an IT question. Once a system does not merely prepare work but performs it — raises an order, produces a report, answers a request — accountability, approval paths and job profiles all change. Introduce agents as a technical project and you get resistance. Lead them as an organisational one and you get results.

Autonomy in levels

We work with four levels because they make approval logic manageable. Each level is tied to the risk profile of the use case, not to what is technically possible.

• Assist — the agent prepares, a person decides and acts.
• Act on approval — the agent proposes a specific action, a person releases it.
• Act independently — the agent acts within defined boundaries, with a log and sampled review.
• Orchestrate — one agent coordinates several others and escalates to a person at the edge of its remit.

What makes an agent production-ready

• A tightly drawn remit. Agents with vague scope can be neither audited nor improved.
• Defined tools and data spaces. The agent inherits the rights of a role, not those of an administrator.
• An escalation rule. Every agent needs an answer to what it does when it is unsure.
• Evaluation before and after rollout. Test sets, regression checks, sampling in operation.
• Cost and usage transparency. Consumption per use case, visible to the business owner.
• A named owner. Not the team, not IT — a person.

The organisation behind it

Once agents work alongside people every day, team composition becomes a leadership question. We use the human-agent ratio as a steering measure: the ratio of employees to agents in productive use within a unit. In a group-wide programme we support, the target is 1:6 by the end of 2027, starting in HR, finance and procurement. The measure is not an end in itself. It forces the question of which work is genuinely being redistributed, and it makes progress comparable across units.

Two roles have proven necessary:

• Forward deploying AI engineer — based in the business rather than central IT, understands the process, and builds and supports the application where it is used.
• AI business partner — the counterpart inside a division or group function, accountable for portfolio, prioritisation and local adoption.

Alongside these sit the questions that decide success in large organisations: leading hybrid teams, adjusting objectives and targets, involving employee representatives early and honestly, and communicating to people what will change about their work — before it changes.