The org chart nobody drew

Thanks for coming back for issue eight. This fortnight the same shape turned up in three very different businesses: a multi-store retailer, a large financial services company and a national services business. None of them had a technology problem. Each one had AI assistants, projects and accounts appearing across the business faster than anyone could keep track of, and nobody could say with confidence what each one was for, who looked after it, or what it was allowed to touch.

A few issues back I wrote about the instructions field on a single assistant. This is the layer above that: what happens when a whole organisation starts creating assistants at once.

The pattern: assistants arrive faster than structure

The way it starts makes complete sense. Leadership wants momentum, so the message goes out wide. Everyone gets access. Everyone is encouraged to experiment. Someone builds a project for the managers’ meeting, someone else builds one for rostering, and a third person builds one that does half of both. Some teams are on one platform, some are on another, and a few people are quietly paying for a third.

Six months in, the organisation has the AI equivalent of a few dozen new staff. None of them were interviewed. None of them report to anyone. Several of them do the same job in slightly different ways, and nobody can find the one that was good.

The invisible org chart

The framing I keep coming back to is simple: every assistant is a role on an org chart, whether anyone has drawn that chart or not.

In a coaching session this week I said it to a business owner almost word for word. Think about a Claude or ChatGPT project as an employee. Each employee has a different job, and each one needs different context to do that job. That idea lands quickly with one person. It gets more interesting at the level of a whole business, because an org chart answers three questions that most AI rollouts skip:

• What is the role for? Which piece of work this assistant owns, and which work it doesn’t.

• Who does it report to? Who checks its output, notices when it drifts, and decides when it’s no longer needed.

• What can it reach? Which systems, files and customers it has access to, and which it shouldn’t.

When those three questions get answered, assistants behave like a team. When they don’t, you get headcount nobody can see. It still costs money, it still touches client information, and it still produces work, but it never shows up anywhere a leader would think to look.

Option one: everyone go and build

I came across three versions of the first option this fortnight.

At the retailer, the owner told me they felt they’d put the cart before the horse by onboarding too many people too early. The business was paying for three different AI platforms, and staff were “running off in different directions with all these AI tools.” The owner’s honest read on most of the team was that they were a bit lost and not using it much.

At the financial services company, the instruction from the top had been that everyone needed to go and build their own AI assistants. The person describing it to me put it plainly: “There was no context. There was no framing.” The assistants got built. Their view was that most people weren’t using them well.

At the services business, the first move had been a board-level mandate: everyone uses Claude. Three months later, everyone was moved onto a different platform because it was too expensive. My read is that the tool wasn’t the problem. A seat for everyone, with no particular job attached to any of them, was always going to look expensive.

Same shape each time. The rollout answers “who gets access” and skips “what is each of these for.”

Option two: draw the chart first

The same services business is also the clearest example I’ve come across of the other way.

After the mandate was unwound, they went back to basics. Their leadership group spent an off-site mapping how the work flows before anyone touched AI: who does what, in what order, and where the pain sits. Only then did they redesign specific steps with AI built in. The rule they set themselves was blunt: no AI on a process until people understand the process. They also held off connecting anything to their core systems until it had shown it worked. In their words: “we don’t do any integration until it’s proven to be working and getting the results.”

The result, by their own count, was roughly a three-fold reduction in the number of steps across the first three processes they rebuilt. Every assistant that came out of that work had a role, a process it belonged to and someone who owned the result. It was a much smaller footprint than “everyone go and build”, and it did a lot more.

Same organisation. Same people. Same pressure on the budget. The difference was the order: the work first, then the roles, then the tools.

When IT answers the third question for you

There’s a side effect of skipping the first two questions that I didn’t expect to see so often.

If nobody has decided what each assistant is for or who owns it, the only question left with a clear owner is the third one: what can it reach. And IT will answer that one, because it’s their job to. The answer is usually “nothing.”

The retailer’s owner told me their security team won’t let them install anything. A research organisation I spoke with works under a risk posture where out-of-the-box tools simply aren’t allowed. The financial services company has global restrictions that shape which features anyone can use at all. In every case IT is being sensible. They’re being asked to approve access for roles nobody has defined, so the safest answer is no.

It’s hard to argue for access on behalf of an assistant when you can’t say what job it does. It’s a lot easier when you can point to the process map and say: this one handles quotes, it needs the estimating system, it never sends anything to a client, and this person checks its work.

Why sprawl wins by default

Nobody chooses sprawl. It wins because creating an assistant is free and takes five minutes, while retiring one has no owner. Merging two projects that overlap takes a conversation nobody has scheduled. Adding an automation feels like progress. Switching one off feels like going backwards.

It also wins because the people closest to it are often the least placed to see it. One of the professionals I coach runs a lot of AI assistants in their own practice, and this fortnight they asked Claude to audit the whole setup. The verdict that came back was that the upkeep costs more than the output. One scheduled filing job had run more than 160 times and filed a single transcript. I found that funny, and also completely familiar.

That’s one person, with one setup, who goes looking. Multiply it across a business where nobody has the full list and nobody is looking, and the cost is real but invisible.

What the missing chart costs

The cost never shows up as a line item. It shows up as staff who tried it, got a bit lost, and went back to the old way. As three subscriptions doing the work of one. As client information sitting inside assistants nobody remembers creating. As an IT team saying no because nobody could tell them what they were saying yes to. And as a leadership team that has “rolled out AI” and still can’t name the handful of roles AI now plays in the business.

The organisations getting somewhere aren’t the ones with the most assistants. They’re the ones who could draw the chart if you asked them to.