AI Strategy 7 - The three As

There are three ways a business can work with a machine:

1. assignment: the machine does a defined piece of work and gives the result back;
2. augmentation: the machine works alongside a person, who remains involved throughout; and
3. agency: the machine acts, and keeps acting, without a person checking each action.

Take the example from the previous two chapters. Sales rose by five units a month. The recorded explanation was that a new advertising campaign had made the difference, and the decision was to keep running it. The business could use a machine to help with this in any of the three ways. It could give the machine the sales and campaign figures and ask it to produce a weekly report for someone to read. It could have the machine work alongside the person deciding whether the campaign really had increased sales, answering questions as they went. Or it could let the machine compare the figures with the target recorded when the decision was made and change the campaign budget itself, week after week, without anyone reviewing the report first.

The main difference is how much freedom the machine has to act. That freedom increases as you move down the list. With assignment, the machine makes no decisions. The task has clear limits, it ends, and a person can accept or reject the result before anything further happens. With augmentation, the machine helps with work that a person is still doing. It can influence the result, but the person remains responsible for carrying the work through. With agency, the machine acts without someone checking every step. It responds to a trigger, causes an outcome, and waits for the next trigger. Anyone who reviews its work is reviewing it after the outcome has already happened.

Risk increases with autonomy because autonomy determines how far a mistake can travel before someone notices. With assignment, an error in the weekly report is usually caught by the person who reads it. The cost is the time needed to correct the report. With augmentation, the mistake may be hidden in the machine’s reasoning rather than in its final answer, so it is harder to spot. But a person is still involved and can question the answer. With agency, the machine may already have changed the campaign budget before anyone reviews its decision. That change may affect the market, a customer, or a business account. It may happen again next week because the machine is still acting on its own. The task has not changed. The business has simply stopped watching as closely.

This is why the business needs to examine the work before handing it over, and why the amount of checking depends on the mode. It needs to establish what the work involves, what it affects, what commitments it creates, and what could happen if it goes wrong. The amount of checking should match the possible harm. A mistake in a report that someone reads may cost an hour to fix. A mistake made by a system acting on customers without supervision may cost far more.

The machine will not ask these questions for itself. The same tool can draft a report when someone asks, help a person think through a problem, or run automatically each time a new email arrives. The tool has not changed. What changes is the arrangement around it. That arrangement determines the mode, so the business needs to state clearly which mode it is using. If it does not, the person using the tool that day will decide without necessarily realising it.

That kind of checking depends on understanding the work properly. You cannot assess the risk in a task until you know what the task actually involves, and its name is not enough. “Produce the weekly sales report” describes the result you want. It does not tell us which figures the report should use, which figures are estimates, where the campaign budget is recorded, whether everyone who receives the report is allowed to see the customer records behind it, or which numbers already reflect somebody else’s judgement.

At this level, a requirement is really a statement about data. Every task reads facts, produces facts, or does both. To understand the task, the business needs to know which facts it depends on, where they came from, what form they are in, where they are kept, and who can access them. These are the four things recorded by the map in chapter three. Once they are known, the business can assess the risk at each level of autonomy and choose a mode based on evidence. Without them, it is deciding how much freedom to give a machine over work it cannot properly describe. This is where the four rungs of the ladder stop being an audit exercise and start deciding what can safely be handed over.

The next chapter, on delegation, starts with the mode already chosen. It asks what should go to the machine, what should stay with the person, which test should be applied to a particular piece of work, and why some work should remain with a person whatever the machine can do. Choosing the mode only after the work has been handed over is not a decision the business made. It is simply a description of what happened.