There is plenty of discussion about AI adoption in business. How do we get people to use it? How do we train teams? How do we choose the right tools? These questions matter. But they are not enough to make the work itself run better.
The question I want to ask is more practical: how do we design systems in which AI understands enough about the work to become part of it?
Understanding, here, does not mean giving a model human intuition. It means providing usable context, identifiable sources, explicit rules and clear limits on what it can do.
Start with the work, before choosing the tool.
An order is more than a document. It belongs to a customer relationship, a pricing agreement, an availability constraint and a promised date. An email can contain a request, but also a reservation that someone familiar with the account immediately understands.
The first design task is to observe what teams already do. What information do they consult? What decisions do they make? Which exceptions lead them to call a colleague? That knowledge belongs in the system, rather than staying hidden in working habits.
Turn context into usable reference points.
Context does not mean putting every company file into a tool. We need to identify which source is authoritative, when it was updated, who may access it and which decision it supports.
In an order-processing scenario, that could mean finding the correct customer account, matching a reference to its catalogue and consulting the applicable terms. A plausible answer is not enough: the team needs to be able to trace it back to supporting information.
From real work to useful action
01ObserveThe tasks, decisions and exceptions.
02ConnectThe right sources, rules and permissions.
03DelegateA defined action that can be checked and reversed.
Business context guides AI. People remain in control of decisions.
Define what we are actually delegating.
Consider an illustrative scenario: a purchase order arrives by email. The system can prepare a proposal from the document. It can also flag an ambiguous reference or a pricing discrepancy. That does not mean it should immediately create an order in the ERP.
Preparing, recommending, approving and executing are different responsibilities. Separating them lets us delegate more to automation while keeping track of decisions. Each action needs defined permissions, checks and a recovery path. The team sets the goals and needs to be able to revise or withdraw the mandate.
Help: gather information and prepare a proposal.
Act: perform an authorized action within an agreed scope.
Ask: make ambiguity visible and involve the right person.
Do nothing: suspend the action when its conditions are not met.
The change belongs to the team, too.
A system can be technically convincing and still be difficult to use if people do not understand what has changed. Why did this suggestion appear? What happens after approval? How can an error be corrected? Who takes over the record?
Supporting teams means answering these questions with real cases and clear responsibilities. Their corrections should also inform a review of the rules and the workflow. This improvement should not be confused with guaranteed, unsupervised machine learning.
The useful measure: are we working better than before?
The number of requests made to an AI tool does not tell us whether work is improving. It is more useful to observe a specific process: less re-entry, an easier handoff, clearer exceptions or a shorter processing time, when those outcomes are measured.
We can start small: one document type, one team, a few representative examples and criteria agreed before the pilot. This is how the conversation moves from a broad promise to everyday work.
A tool that serves human work. Not the other way around.
A useful first workshop
✓ Choose a recurring process and observe how it works today.
✓ Gather representative records, including exceptions.
✓ Identify authoritative sources, access permissions and responsible people.
✓ Define what the system prepares, what it executes and what needs approval.
✓ Compare the before and after workflow with the team doing the work.
