Proactive without deciding for you

In everything I have described so far, there was one thing the user still did by hand: phrasing the request.

Here is the last case, the one where the agent no longer waits.

Recurrence, not the calendar

The example always given is Monday morning. It is misleading, because it suggests that proactivity is a matter of calendar.

It is a matter of observed regularity, and regularity takes several forms. Temporal, when a validation comes back at the start of the week. Sequential, when a check systematically follows another task. Conditional, when a follow-up is triggered beyond a certain lead time.

What matters is that the detected recurrence is displayed with what it is based on. Eleven occurrences over twelve weeks, thirty-seven consecutive sequences. No black box.

• Supplier order validationtemporal recurrence

  Done at the start of the week, 11 times in the last 12 weeks.

• Receipt variance checksequential recurrence

  Systematically chained after every batch validation, with no exception recorded.

• Carrier follow-upconditional recurrence

  Triggered as soon as an announced lead time exceeds 48 hours, 9 times out of 10.

What triggers proactivity is the detected recurrence, and it stays displayed with what justifies it.

On opening, the work is prepared

The user opens the application without asking for anything. Fourteen supplier orders are already gathered, with the reason for the proposal displayed.

Nothing is committed. It is exactly the principle of the second case again: the agent produces a readable intention, and waits for an explicit gesture.

• Profile Purchasing manager
• No request made

Single screen — conversational agent

1. System response

   14 supplier orders are waiting for validation. I have gathered them, you have nothing to look for.

   why this proposal

   You have been doing this validation at the start of the week for eleven weeks.

   • Open the selection
   • Not now

The agent removes the last thing that was still done by hand: phrasing the request. It prepares, it does not execute.

The real trap: automation by fatigue

This is where the model plays out, and it is the part I find the most interesting.

An agent that merely speeds up the routine builds a perverse mechanism. By dint of preparing well, the user stops looking. The confirmation still exists, but it no longer controls anything: it has become a reflex.

So the agent does not put forward the twelve compliant orders. It surfaces the two that deviate: a unit price eighteen per cent above the last contract, an announced lead time of twenty-one days against five usually.

A well-executed routine does not deserve your attention. The exception does.

Single screen — conversational agent

1. System response

   Of the 14 orders, 12 are in line with what you usually validate. Two deviate from it — I suggest starting with those.

   • Novaline SASUnit price at €15.30 against €13.00 in the last contract, 18% above.
   • Kervad IndustriesAnnounced lead time of 21 days, against an average of 5 days over the last six orders.

   The other twelve stay grouped, ready to validate in a single gesture whenever you want.

This is where the difference between assisting and replacing plays out. The agent does not speed up repetition, it surfaces what deviates from it — and puts the decision back where judgment has value.

The agent does not speed up repetition. It surfaces what deviates from it, and puts the decision back where judgment has value.

What the system learns belongs to you

Modelling an employee’s habits means processing behavioural data at work. In France, that is not a subject settled after design: declared purpose, retention period, informing and consulting staff representatives.

The constraint is healthy, because it imposes what had to be done anyway. The model is the user’s own, it is shown to them, and they can switch it off routine by routine.

• Validation at the start of the weekBased on 11 occurrences observed over 12 weeksForget
• Variance check after every batchBased on 37 consecutive sequencesForget
• Carrier follow-up beyond 48hDisabled by the user on 9 SeptForget

The agent helps you decide. It does not decide for you.

Modelling an employee’s habits means processing behavioural data at work. The constraint does not come afterwards: the model is the user’s own, it is shown to them, and they can switch it off.

The agent helps you decide. It does not decide for you.

What the series establishes

At the end of these four cases, the synthesis fits in one sentence: the screen is no longer frozen, rights and the style guide are.

No longer frozen: navigation, the form of the rendering, the scope of a task, the screens delivered per cycle, the filter nobody had thought to provide, and the initiative of the request.

Stays frozen: the actions the profile allows, the client’s design system, the data model, end-to-end traceability, confirmation before writing, and what the agent learns — which the user can view and erase.

What is no longer frozen

• Navigation and menus
• The form of the rendering
• The scope of a task
• The screens delivered per cycle
• The filter nobody had thought to provide
• The initiative of the request

What stays frozen

• The actions the profile allows
• The client’s design system
• The data model
• End-to-end traceability
• Confirmation before writing
• What the agent learns, and what the user can erase

Data stays alive all the way to the screen.

What I do not know yet

Three questions remain open, and I would rather ask them than hide them.

The determinism of reading: two close wordings can produce two different aggregations, and nothing tells the user.

The displacement of complexity: it leaves the screen to join the semantic layer. The system has to know what a “product to reorder” is in this client’s vocabulary. That is not less work, it is different work — and it looks a lot like business modelling.

The imposed habit: if someone validates every Monday, is that a choice or a constraint of the process? A system that learns habits also learns the workarounds, then institutionalises them.

Data stays alive all the way to the screen. Where do you put the boundary between a system that anticipates and a system that replaces you?