AI and organisation design

Agents change the work. Almost nobody changes the organisation

Most AI programmes fail as organisation problems, not as technology problems. The tools work. The structure around them was designed for a company where people did every step.

Decision rights still sit where they sat. Spans are still sized for supervising people. Nobody has decided who is answerable when an AI agent produces the output, so in practice nobody is.

The pilot impresses, and the operating model absorbs it without changing.

Four questions this actually raises, and where we land on them

Where does accountability sit when an AI agent produces the output?Never with the AI agent, and it cannot sit with whoever bought the tool. It has to sit with a named person who can be asked why, and who has the authority to stop it. In a regulated setting that is not a nicety; it is the thing an auditor asks first.

What is a span of control when half the team is not human?Supervising AI agents is a different job from managing people, and the 2 do not add up. A manager with 6 people and 12 agents does not have a span of 18. Layers usually do come out, but rarely where the slide says they will.

Which roles change shape, which get faster, and which stop being roles?Most roles change shape rather than disappear, and the ones that do disappear are seldom the ones everybody expects. The only honest way to answer is to look at tasks rather than job titles, which is slower to say and faster to get right.

When does headcount stop being the unit of design?Once a meaningful share of the work is produced rather than staffed, sizing by full-time equivalent stops describing capacity at all. At that point you design around work volume and decision points, and headcount becomes an output rather than an input.

The constraint nobody has priced

An AI agent takes work off the people who write the thing. It takes nothing off the people who have to check it and put their name to it. Writing capacity scales. Checking does not.

That is not a philosophical point, it is arithmetic, and it has a crossing point you can put a number on. 4 steps below, one commercial team, and a slider each. Move them to your own numbers and the last step works out where it breaks.

How much work comes in

Quotes that go out in a week today 40 quotes a week

40 quotes a week. Every one of them is written by somebody, checked by somebody else, and signed by a person who carries it.

Who does the work on one quote

Share of one quote an AI agent drafts 70% by AI

An AI agent drafts 70% of a quote. The employee still does the other 30%, which is the part that has to be right for this customer.

AI agent 70% employee 30%

How long one quote used to take

Minutes an employee spent on a whole quote before AI 60 minutes

A quote used to cost an employee 60 minutes. Their share now costs 18 minutes. Nobody goes home earlier, so the same team turns out 133 quotes a week instead of 40.

60 min before

18 min now

What it takes to check it and sign it

Minutes for a second pair of eyes, plus the signature 15 minutes

People who check and sign 2 Hours a week each of them has for it 8 h

15 minutes each, and 2 people with 8 hours a week for it. That is 16 hours of checking, and it does not grow when the writing speeds up.

→

What that adds up to

Your checkers run out at about 65 quotes a week. The AI takes this team to 133.

The writing never became a problem. That team spends about 40 hours a week either way. Checking goes from 10 hours to 33 against the 16 they have, and that lands on 2 people. 3 doors out, and all 3 are organisation design.

01 Put more people on itYou would need about 5 people checking and signing, up from 2. That is the management layer you had just congratulated yourself on removing.

02 Check a sampleThey can properly get through about 50% of what comes out. Approved then means sampled, and someone has to say so out loud.

03 Need fewer signaturesMove the threshold so roughly 50% of quotes never need a second pair of eyes, because they were never the kind that warranted one.

A mechanism, not a measurement, and a generous one. It assumes checking an AI-written quote takes exactly as long as checking a colleague’s, which is optimistic: a person’s mistake tends to look wrong and a machine’s tends to look right. It also takes one kind of work. Run it on the kind an AI can help with least, a tender or a claim, and the crossing arrives sooner.

Doors 2 and 3 produce the same arithmetic and mean opposite things. In one, what a signature is worth quietly erodes while everyone keeps signing. In the other, you decide on purpose that a class of work never needed one, and you move the threshold to say so. Which of the three you take is a design decision, and it is the sort that repays an afternoon of senior attention rather than a year of pilots.

One thing this is not. We do not measure any of this for you, and there is no time and motion study waiting behind it. Half an hour with the person who signs gets these numbers close enough to argue about, which is all they need to be. The number is not the deliverable. Which door you take is, and what follows from it is decision rights, sizing and where the work sits, which is exactly the level we work at.

What we do about it

The same 4 things, pointed at this question. A Verdict on a target operating model that assumes AI agents but was drawn as if it did not. A Design that works out where the work should sit once part of it is produced rather than staffed. A Brief when there is an AI ambition, a board expectation and no project yet. It is the same craft; AI agents are a design variable, not a separate discipline.

What we do not do

We do not implement AI, select tools, build AI agents or run your data programme. There are good firms for all of that and we will name them. We design the organisation the AI agents will work inside, which is the part almost nobody is doing and the part that decides whether any of the rest pays back.

And we use it ourselves

Analysis, synthesis and first drafts are faster than they were, which is part of why a resolved design costs what it does here and takes weeks. What it does not do is form the judgement or carry the consequence. That still needs a senior person who will put their name to an answer.

See the 4 ways in →