The first agent is fun. The twentieth is a governance problem.
Getting started with AI agents is easy, and you should. But without structure they grow arms and legs in every direction. Here is what I see happen, and what holds it together.
Week 1: one agent, one task. Everyone knows what it does.
Month 3: a few more. Still easy to follow.
Month 9: who does what, and with what access?
With governance: every agent has a place in a process with an owner.
The good start
I like starting with an agent here and there. One that summarises meetings. One that sorts the inbox. One that writes the first draft of a memo. It is low-hanging fruit, the wins come fast, and people get curious. It is a good start, and I recommend it.
Then it starts to grow
Over time everyone builds their own agents. Instructions get copied and tweaked. An agent gets a bit more access, because that was easiest. Two agents solve the same task in different ways. Nobody did anything wrong, but nobody has the overview any more.
The problem is not the agents. It is that they automate tasks, not processes. Without a map of the processes, every agent becomes a local fix, and speed makes what is unclear happen faster, not clearer.
What holds the structure
The answer is not fewer agents. It is a shared map of the processes, and clear rules for who owns what.
- A process landscape that shows which processes you have, and where the agents belong.
- Owners who decide how each process should work.
- Traceable decisions that show who claimed what, on what basis, and what has actually been decided.
- Agents with roles that only read from allowed sources, only take allowed actions, and where everything is logged.
- People who decide, while the agents capture, structure and retrieve.
| Claim | Claimed by | Basis | Status |
|---|---|---|---|
| The invoice flow takes too long | Finance lead | Interview | Claimed |
| Approval moves to department heads | Working group | Workshop | Proposed |
| New invoice flow goes live | Leadership team | Decision meeting | Decided |
Agreeing on a wording is not a decision.
For technical leads
Landscapes and capabilities are modelled in ArchiMate, detailed processes in BPMN 2.0 and decision rules in DMN, open standards other tools and vendors can read. Automated flows run on a process engine, not in hidden scripts, so every step is visible and can be audited.
How we start
We do not start with tools. We help leadership tell the why, in five parts that you own yourselves:
- NowWhere are we today?
- GoalWhere are we going?
- BenefitWhat do we get out of it?
- Gap and planWhat is missing, and in what order?
- MandateWho decides, and what is settled?
Then we take the agents you already have, and give them a place on the map.