I was on a panel at the Parcel+Post Expo in Amsterdam last year, and a line from one of the other panellists has not left me since. He said the future IT department will be the HR department of agents.

Sit with that for a second. Not the team that keeps the laptops running. The team that hires, onboards, supervises, corrects and, when needed, lets go of a workforce that happens to be software.

It sounds like a good line for a conference stage. It is actually a quiet warning about how unready most organisations are for what is already arriving.

In the launch issue I argued that consulting without outcomes is over. Agentic AI is the reason it is over, not the reason it is fashionable to say so. (If you missed it: "escape the deck" means escaping the deck that ends at recommendations and walks away. Three ways out, depending on where you sit: leave consulting to operate and build, stay and rebuild how your firm delivers, or, as a client, stop buying decks and demand outcomes.)

Here is the mechanism. AI does not just change the slides a consultant produces. It changes whether the deck is worth buying at all. When analysis-for-hire can be generated in an afternoon, the value of analysis-for-hire collapses. What is left, the only thing left, is a delivered, governed outcome. Someone who will stand behind the result, not the recommendation, and more and more, someone who is paid for the result, not the advice.

That is why "the future IT department is the HR department of agents" matters to every reader of this newsletter. If you are a buyer in government or enterprise, you are about to be sold a lot of agents. If you run a consulting practice, you are about to be asked to deploy them. And if you are trying to escape into an operating role, this is exactly the work that role now involves.

Because here is the part nobody demos: the gap between a slick AI pilot and a scaled operation is never only about more technology. It is operating model design. Think of the pilot as a bright kid in nursery. It looks brilliant, because one adult is watching its every move and clearing every obstacle before it hits one. Rolling it out is expecting that same kid to turn up at university on Monday and thrive on its own. It does not work like that for children, and it does not work like that for AI. The pilot works because one clever person is babysitting it. The scaled version has to work when nobody is.

The babysitting is not a figure of speech. “Botsitting” is feeding agents the context they are missing, checking their work, and cleaning up the confident, wrong answers they leave behind. That cost is invisible in the pilot, because the clever person absorbs it. It is very visible at scale, when there are a hundred of them and nobody spare to watch. If your scaling plan quietly assumes that hour away, it is not a plan.

Run the four-question test, on something real, this fortnight.

Pick one thing: an AI pilot you are running, or a consulting recommendation currently in flight that ends at "you should adopt agents". Put it through the four questions above.

Write your answers down in plain sentences, no slides allowed! If you can answer all four cleanly, you have a genuine scaling plan and you should move faster. If you stall on even one, you have found exactly where the gap is, and that is useful to know before you have committed budget, headcount and credibility to it.

A pilot that cannot survive these questions was never going to survive contact with real operations.

The OECD published its first Digital Government Outlook this month, and one pairing of numbers says everything about this moment. AI is now used in at least one area of government in 97% of OECD countries. But only 28% of those countries report doing any assessment, financial or otherwise, of whether their AI use cases actually deliver.

So AI is almost everywhere in government. The outcomes are not. Adoption is nearly universal and proof is rare, which is close to a working definition of theatre.

That 28% is the whole argument in one figure. The countries that win the next phase will not be the ones with the boldest strategy. They will be the ones who can answer the four questions above. That is the focus now: from AI ambition to operating reality.

A new section, and the idea is simple. Each issue I will share one piece of advice from someone else that I think is genuinely worth your time.

This fortnight it is Ethan Mollick, the Wharton professor behind Co-Intelligence and the One Useful Thing newsletter. His advice for anyone trying to work out what AI means for their job is disarmingly simple: give it ten focused hours.

Not a demo, not a webinar. Ten real hours using one frontier model on your actual work, the emails, the analysis, the plan you were writing anyway. His argument, and I've seen this in the work I do, is that nobody can tell you where AI helps your job until you have felt where it is brilliant and where it falls over. Ten hours is the entry fee to an informed opinion.

Try it before the next issue. One model, your real work, ten hours. Then hit reply and tell me what surprised you.

Five that shaped this issue, all worth your time:

Where is your own pilot stuck? Or, more usefully: which of the four questions can you not yet answer for the AI work currently on your desk? Whether you are buying it, building it, or trying to escape into the team that owns it, I want to hear which one trips you up. I read every reply.

Keep building, Alex

PS. If you are mapping a move out of consulting and into operating or building, I made the Consulting to Tech Transition Map for exactly that. It is free on the site. Next issue lands in a fortnight.