Agents are for the work the rules cannot name.

Start with the good news the hype skips over: most work is deterministic and that is a gift. Clear steps, written once, that run the same way every time, that you can test, trust and then stop thinking about. That is a script. For the overwhelming majority of what an organisation does, it is exactly the right thing to build. The frontier worth getting excited about is the smaller slice of work that genuinely resists being written down. That is the interesting part: what becomes possible once you let a system find its own way through it.

Agentic AI is real and it is also the most over-sold idea of the year. Holding both of those at once is the whole skill. What tells them apart is a single property: whether the system is deterministic or stochastic.

01Deterministic or stochastic, the distinction that decides it.

Strip away the marketing and a few quite different things wear the same label. Rule-based automation runs fixed steps that never change. Intelligent automation adds a model and a human checkpoint, so it suggests and a person confirms. Agentic AI takes the checkpoint off the routine path and hands the system a goal, a set of tools and the latitude to decide the order for itself.

What separates them is determinism, not the model. A script is deterministic: give it an input and you know the output and you can prove it before it ever runs. An agent is stochastic: give it a goal and it works out the route at runtime, shaped by what each step returns, so two runs may not take the same path. That one property is the whole trade and everything genuinely useful about agentic AI flows from it.

02Where an agent genuinely earns its place.

An agent earns its place where the work is genuinely open: where the inputs are not known in advance, where the right next step depends on what the last one turned up and where the cost of being a little wrong along the way is small enough to absorb while the system finds the shape of the problem. Think of triage that weighs a dozen signals at once, research that follows a thread wherever it leads or reconciliation across systems that were never designed to speak to each other. Work like that resists a decision tree and it is precisely the kind an agent can carry when nothing else can.

The other side of that is just as clarifying and just as useful. A refund that follows fixed rules does not want an agent; it wants a script and it will be cheaper, faster and easier to stand behind as one. Knowing which kind of work you are holding is most of the craft and getting it right is what lets you spend the agent where it genuinely earns its keep rather than everywhere the word sounds like progress.

03The craft that makes autonomy dependable.

Here is the part I find most satisfying, because it is where the real engineering lives. An agent you cannot fully predict is not something to be feared; it is something to be instrumented. You may not be able to say in advance exactly what it will do, but you can make sure you can always reconstruct what it did: a record of every decision and the reason behind it, boundaries it is not permitted to cross and a person kept on the consequential calls by design rather than as an afterthought. Do that and the very unpredictability that made the agent powerful becomes something you can rely on.

This is not theory for us. We run multi-agent systems in regulated production today, with a separation of duties between models for quality control, guardrails on what each one is allowed to do, human-in-the-loop gates at the points that carry consequence and logging complete enough to answer "why did it do that" months after the fact. The autonomy is real and so is the accountability; they are designed together, not traded off against each other.

The EU AI Act asks for a version of exactly this from high-risk systems: keep records under Article 12 and keep a person in genuine oversight under Article 14.¹ The timeline for those high-risk obligations has since moved, but the point holds with or without the regulation. These are simply what a serious operator wants for a system whose behaviour cannot be fully predicted in advance. In this respect, the Act is good engineering written into law.

Instrument it well and the unpredictability that made the agent powerful becomes something you can trust.

04Where the advantage actually comes from.

There is a craft to starting well and it has less to do with the model than with judgement. It is choosing the one slice of work that genuinely needs an agent rather than the flashiest, then instrumenting it so it earns your trust on the numbers that actually matter, which are rarely the ones a demo shows you. Get that part right and the rest tends to follow.

The advantage, when it comes, does not go to whoever deploys the most agents. It goes to the team that can tell, for each piece of work, whether it is deterministic or not, builds each one the honest way and can account for the difference. That is a more interesting place to be than the front of the hype cycle.