There is a job inside most agentic AI programmes that nobody wrote down, nobody costed, and nobody owns. It is called botsitting. Nobody seems to know who first coined the term, but it is exactly what it sounds like: babysitting, for agents.

Botsitting is the invisible work of keeping agents in line. Working out the context they should have been given. Correcting their bad manners, which is the polite word for hallucinations. Cleaning up after them when they act on something they should have escalated. None of it appears in the statement of work. All of it appears in someone’s week.

There are only two honest ways to read that problem. The first is that agentic AI is not worth the effort, so we should move on to the next interesting thing. The second is that we are not doing it well yet, and we should get better at it. I am firmly in the second camp, and the rest of this is about what getting better actually involves.

Getting an output from a pilot is roughly twenty per cent of the work. The other eighty per cent is the part that does not demo well: harness engineering, data and integration, change management, governance, monitoring, and continuous improvement. Boring. Next.

That ratio is the whole problem, because the twenty per cent is what gets funded and celebrated. Gartner forecasts that over forty per cent of agentic AI projects will be cancelled by the end of 2027, and gives three reasons: escalating costs, unclear business value, and inadequate risk controls. Look closely at that list. Not one of them is a model problem. Every one of them lives in the eighty per cent.

The failure lives around the agent, in the work everyone skips. Pilots are not dying on capability. They are dying on discipline, or the lack of it.

Most of the market is doing implementation and calling it transformation. The distinction matters more than it sounds.

Implementation is the twenty per cent. It has an expiry date, it gets you an agent live, and what you end up with is a faster version of the way you already work. That is a real result and I am not dismissing it. It is just not the thing most boards think they approved.

Transformation is the eighty per cent. It redesigns the work around what agents can now do, augments the people doing it, and brings them with you rather than around them. Redesigning the work is the single biggest driver of value I see, and it is the step most organisations skip, because it is slow, it is political, and it does not fit in a quarter.

When people talk about getting more out of a model, they almost always mean the first of these three.

An agent is a model plus a harness. The model is the intelligence. The harness is everything else. Most teams ship the model and forget the harness, or they build the harness once and never set up the governance to maintain it. That is the moment the pilot stalls, and it usually happens a few weeks after the demo everyone loved.

If you take one operational thing from this, take this: the metrics most teams have pointed at their agents are infrastructure metrics, and they cannot tell you whether the agent is making good decisions.

Amongst the many measures you could track, four matter most, and they are the ones I would put in front of a business owner rather than an engineer.

And underneath all four, one question of ownership: who is accountable for what the agent did at two o’clock on a Saturday morning. If you cannot name the human, you are not in production. You are in a long pilot with good lighting.

Building agentic solutions is not a project with an end date. It is a transformation process that keeps evolving, and it needs a disciplined, governed, structured, long-term mindset, ideally backed by some hard-earned lessons from things that did not work.

We are building our own harness internally, for our own agents, across our own teams. Not as a demo. As the way we intend to work. That is a deliberate choice, because I do not think you can credibly sell a transformation you have not run on yourself.

Does that sound like hard work? It is. No dramas. Your competitors will figure it out.