The ideas on this site in eight diagrams. If you only have five minutes, scroll through these. Each links to the piece it comes from. Open any diagram full size, and download it as SVG or PNG to use with a credit to David Reynolds · ai.valuecreator.io.
The saving from automating junior work shows up this year. The cost shows up years later, as a missing layer of experienced people, somewhere else in the business. No single report shows both, which is why it's easy to miss. It is avoidable: decide which work people keep because it teaches, and the line doesn't have to bend.
Most automation decisions sort tasks by what the output is worth. Sort them by what doing the task taught as well, and every junior task lands in one of four places. Automate what taught nothing. Where the task was how people learned, let the software draft and keep a person deciding, or keep a share of the volume for people. The right-hand column is where automating cuts into how people learn.
Once a person is past about seventy per cent busy, the exceptions waiting for them pile up faster than they can clear them, and most plans staff supervisors towards a hundred per cent busy. The more reliable the automation, the worse people get at stepping in when it fails. Most plans assume one person can run far more agents than they can.
Along the bottom, how much your agents do without asking. Up the side, how much evidence you hold that they should. Most businesses are piloting, which is a fine place to start. The risk sits with businesses that have granted autonomy without the evidence behind it. To reach "earned", move up before you move across: get the evidence, then grant the autonomy.
Rent, license, host or run the models, step by step, and swap them when something better passes your tests. Own the layer underneath: one connected picture of your customers, your work and your decisions, a register of every agent with a named person against it, and a test set made from your own cases. The decision a customer relies on stays with a person.
Take a customer's contract renewal that runs seventy days from start to finish, with only about four of them spent on the work. The rest is waiting on figures before and going round again after. Automating the work saves days; removing the waiting saves most of the seventy.
One workflow at a time, through five stages, with a decision to carry on or stop between each, taken on measured evidence against criteria written down first. AI makes every stage faster. It doesn't remove a gate.
There are four ways to get a model: rent it from the lab, license it through your cloud provider, host an open-weight model on machines you control, or run one on hardware you own. Choose step by step: the best rented model for rare, hard judgement, a small hosted model for high-volume work that can be checked, and your own hardware where nothing may leave. Whichever you choose, every model plugs into the same layer underneath, which is what makes changing one cheap.
