A little over a year ago we made a decision that most firms our size would call either brave or reckless. Rather than give everyone an AI assistant and see what happened, we put AI agents inside the business itself – into sales operations, market research, reporting, platform monitoring and the running of our expert network.

Today a small team of accountable people works alongside around a dozen agents in production. We have spent more than 500 hours building the tooling, orchestration and governance that holds it together.

This is not a showcase. We advise organisations on AI operating models, and we think the most useful thing we can do is show our working: what held up, what we got wrong, and the AI rollout lessons we would now give anyone starting out.

Why we put AI at the heart of the business, not the desktop

We are a network business. For 13 years our value has come from the judgement of experienced practitioners, assembled around a client’s problem. That model has a weakness every network firm will recognise: the knowledge lives with the people. When an engagement ends, much of what it taught us walks out of the door with the practitioners who did the work.

We had the usual symptoms. Insight from one engagement rarely reached the next. Several critical activities depended on a single person. Engagement with our wider network dropped away after the first month or two because nobody had the time to keep the conversations going.

Desktop AI tools made individuals faster. They did not make the firm any smarter. Each person’s prompts, drafts and discoveries stayed on their own screen. That is how most organisations use AI today, including many of the largest consultancies: personal productivity tools where the learning stays with the individual.

We wanted the opposite – a system where, when one piece of work produced an insight, every future piece of work could use it.

“Desktop AI made our people faster. It didn’t make the firm any smarter.”

3 decisions we would make again

1. We built one system, not a set of tools

The first and most important decision was to treat AI as part of the operating model rather than as software for individuals. Our agents share context. They work from the same records, follow the same standards, and hand work to each other and to people through the same routes.

That sounds like an architectural nicety. In practice it changed the economics of ordinary work. Research done for one proposal is available for the next. A pattern spotted in conversations with our network shows up in sales preparation. A lesson captured at the end of an engagement is there when the next one starts, rather than in someone’s notebook.

It also forced a discipline we would otherwise have avoided. You cannot share context across a system without agreeing what the records are, who maintains them and what “good” looks like. That work was tedious. It is also the reason everything else functions. If we had skipped it, we would have built a dozen clever agents each working from its own version of the truth.

2. We made every handoff explicit

Work moves between people and agents through a fixed sequence: a heads-up, an assignment, an explicit accept or decline, completion, and review.

The step that mattered most was the explicit accept or decline – including for agents. An agent that cannot do something properly says so, rather than producing something plausible and wrong. A person who is overloaded can decline rather than letting a task quietly die in an inbox.

Before this, the failure we feared most was not bad output. It was silent dropping: work that everyone assumed someone else had picked up. Making acceptance explicit removed most of it. It also gave us a clean record of where work actually goes, which has turned out to be the best evidence we have about how the business really runs – far better than any process map we drew in advance.

3. We wrote down what agents may do – and told them to escalate, not improvise

Every agent works within defined authority. Agents draft, recommend and execute inside that boundary. Beyond it, the rule is short: escalate, don’t improvise.

We expected this to slow things down. It did the reverse. Because the boundary is clear, nobody needs to check everything an agent does inside it. Review effort goes where the risk is, rather than being spread thinly across every output. And because agents escalate rather than guess, the escalations themselves tell us where the boundary should move next. Governance stopped being a fixed document and became something that improves with use.

2 decisions we reversed

1. We let agents “own” things

In our early organisation designs, some capabilities showed an agent as the owner. It looked efficient on paper. A research agent owned market intelligence; a reporting agent owned the dashboards.

It did not survive contact with a real problem. When something went wrong – a stale figure in a report, a briefing that missed a relevant development – nobody felt it was theirs. The agent had done what it was asked. The people around it assumed the agent was responsible. The gap between those 2 positions is exactly where quality goes to die.

We reversed it completely. A named human now owns every outcome. Agents are responsible for execution; people are accountable for results. Every capability in our operating model has a person’s name against it, with agents listed as support. That single change did more for quality than any improvement to the agents themselves.

2. We counted agents as if they were progress

For a while we talked about the number of agents we had in the way firms talk about headcount. More agents felt like more capacity.

It was the wrong measure, and it quietly encouraged the wrong behaviour – building another agent felt like an achievement whether or not it changed anything. We stopped treating agent count as a capacity or value metric. What we now try to measure is what the system does for the work: time to the first useful insight, rework avoided, continuity between phases, and how much of our senior people’s time goes on judgement rather than administration.

Those measures are harder to collect. They are also the only ones a client, or a board, would care about.

A smaller adjustment worth sharing

We switched agent initiative on by default – agents could raise new topics, spot risks and suggest work. Within weeks, people were drowning in well-meant suggestions. We kept initiative on but gave it a budget: a small number of new topics per person per day, tunable by the person receiving them. Initiative is valuable. Unlimited initiative is noise.

We made a similar adjustment to governance. We started with more review forums than a small team could sensibly sustain, and we now consolidate any forum that doesn’t produce distinct decisions.

What we now tell clients on day 1

When we sit down with a leadership team about to put AI agents into how their organisation runs, these are the things we say first – because we learnt them the expensive way.

1. Start with the work, not the tool. Pick a workflow that matters and understand how it really runs before deciding where AI belongs in it.

2. Name the human owner before you build anything. If you can’t say who is accountable for the outcome, you’re not ready to automate any part of it.

3. Write the authority down. What can the agent do on its own, what must it check, and what must it hand back? Ambiguity here becomes either paralysis or risk.

4. Design the handoff before the automation. Most failures happen between steps, not inside them. Make acceptance explicit.

5. Measure the work, not the agents. Decide in advance what better looks like in business terms, and baseline it before you start.

6. Expect to reverse something. You will get at least one design decision wrong. Build in the review that lets you notice early.

What we haven’t solved yet

Honesty cuts both ways, so here is what remains open.

We believe a blended team of experienced practitioners and governed AI produces better outcomes for clients: more senior attention per pound spent, faster iteration, better continuity and more capability left behind. We see it internally. Until we have measured it consistently in client engagements, those benefits remain hypotheses rather than claims – and we treat them that way. Our first engagements on this model are instrumented as both client delivery and evidence-gathering, and we will publish what we find.

We are also still working out how to keep shared records clean as the system grows. Agents are only as good as the context they share, and data quality decays quietly unless someone owns it.

Common questions

What’s the difference between AI on the desktop and AI at the heart of a business?

Desktop AI means individuals using assistants for research, drafting and search; the learning stays with each person. AI at the heart means agents working inside a governed operating system that shares context, captures lessons and reuses them across all work, so the organisation itself gets smarter.

Who should be accountable for an AI agent’s work?

A named person. Agents can be responsible for executing tasks within defined authority, but accountability for outcomes should always sit with a human who has the authority to act when something goes wrong.

How should you measure an AI rollout?

Not by the number of agents or users. Measure changes in the work: time to first useful insight, rework, cycle time, continuity between phases and how much senior time goes on judgement rather than administration – each against a baseline taken before you start.

How many agents should an organisation start with?

As few as it takes to change one workflow that matters. Agent count is not a measure of progress; a single well-governed agent in the right place is worth more than a dozen working from separate versions of the truth.

What this looks like in practice

We call ourselves Client Zero: every method we recommend for AI operating models, we run on our own business first. When we work with a leadership team on scaling AI, we start where we started – with the work, the owners, the authority and the evidence – and we bring the scars as well as the successes. Explore our AI Value & Scale practice →