The question we hear most often now is not which model to choose or how to comply with the EU AI Act. It is simpler and more anxious than either: where should AI go in our organisation, and should it go anywhere at all?
It is the right question, and most organisations answer it at the wrong level. They decide by department or by job title: automate the claims team, give the analysts a copilot, replace first-line support. The evidence says that decision belongs one level down, at the level of the task.
Jobs are the wrong unit
A job is a bundle of tasks. Some are routine and some need judgement. Some go to a customer and some never leave the building. Labour economists have analysed work this way for more than twenty years (Autor, Levy and Murnane, 2003), and it is how researchers measure exposure to AI today. When OpenAI and University of Pennsylvania researchers applied the method to large language models, they estimated that around 80% of US workers could have at least a tenth of their tasks affected, and around 19% at least half (Eloundou et al., 2023).
Read that the other way round and it says most jobs are partly exposed and very few are wholly exposed. A plan made at the level of the job title will automate tasks that should have stayed human and miss tasks that were ready, usually inside the same role.
Same people, same tool, opposite results
Two of the best field studies we have show why the task matters more than the tool.
In a study of 758 consultants at Boston Consulting Group, those given GPT-4 for tasks inside the AI's capability completed 12.2% more tasks, worked 25.1% faster and produced work rated more than 40% higher in quality. On a task deliberately chosen to sit just outside that capability, the same consultants using AI were 19 percentage points less likely to reach the correct answer than colleagues working without it (Dell'Acqua et al., 2023). The researchers call the boundary a "jagged frontier": tasks that look equally difficult to a person can fall on opposite sides of it.
In a study of more than 5,000 customer support agents, an AI assistant raised issues resolved per hour by 14% on average. For novice and less-skilled agents the gain was 34%. For the most experienced it was minimal (Brynjolfsson, Li and Raymond, 2025).
Put those together. A decision taken for "the consultants" or "the support team" gets the gains and the losses at once, and the average hides both. The second study carries a harder lesson too. The biggest gains went to the people who were still learning, which makes the junior work, where people build their skill, the most valuable place to use AI and the most dangerous place to remove people from.
The tool is the same on both sides of the frontier. The task is what changes.
Three ways an AI plan fails without a map
The tool in search of a problem. Bought on the strength of a demonstration and piloted with success measured as how many people logged in. Six months later nobody can say what changed, and the team is more sceptical than before it started.
The broken process, automated. A process that lives in one person's head, or runs on three spreadsheets that disagree, does not improve when you automate it. You get the same mess, faster, and harder to see.
The apprenticeship, removed. The work AI does best (research, first drafts, reconciliation) is often the work junior staff learn the job on. Automate all of it and in a few years nobody in the building is qualified to check the machine. This is the failure that does not show up in the first year's figures.
A map you can start this week
You do not need a consultant to begin, and for one team you can do it yourself.
If you cannot say where the saved hours will go, the business case is not finished.
What the map tells you
One of three things. Usually it gives a short list of two or three places to start, smaller than the vendor proposed and far more likely to stick. Sometimes it says "not yet", with the conditions that would change the answer: cleaner data, a documented process, a team with room to take on something new. Occasionally it says no, because the volume is too low or the cost of a mistake is too high to justify the set-up.
All three are good outcomes. The bad outcome is the one you get without a map: a purchase made first and justified afterwards. If you want to see what a finished map looks like, our workforce mapping page walks through an illustrative one for a 30-person property management firm.
When to bring someone in
A single team is a manageable exercise to do yourself. It gets harder when the map crosses several functions, when a restructure is already on the table, when the work is regulated, or when the people being mapped have reason not to trust an exercise run by their own management. That last one is more common than boards admit, and it is where an outside view earns its fee.
Workforce mapping is one of our seven consultancy services. Scope, timetable and cost are agreed in writing before an engagement begins.