The Feeling Was the Mechanism

Almost everything I read from an operations programme is about process. One short proposal I keep going back to, written for a large European banking group about a decade ago, gives one of its seven slides, and a named box on another, to something else entirely. It is about what more than two thousand people were supposed to feel.

The setup is ordinary. A step-level analysis had been run across the whole internal incident handling chain, the path a fault takes from a user through a service desk into internal departments and sometimes out to a supplier and back. On the document's own account the analysis worked. It put a conservative saving of 3.3 million euros a year on the table, conditional on 300,000 steps coming out of the chain, and it named pathologies the document says had stayed invisible until then. The realisation team was staffed, the resources were freed, and the money had not moved.

The diagnosis in the proposal is that what was missing was behavioural rather than analytical, and I think that diagnosis was correct. What I want to examine is the intervention chosen to act on it. One of the two proposed workstreams has, as its stated aim, the provocation of an imagination: that somebody is now watching how each person works. The document names the effect after the watcher in Orwell's novel, describes it as something to be carefully created, and in the same breath frames the watching as help.

What the workstream actually asked for

Strip the framing away and the specification is short. Take the success stories the chain teams are producing. Push them outward through channels that already exist, deliberately rather than incidentally, so that awareness travels faster than the project team can. Attach a message with a specific behavioural expectation in it. Let the arithmetic do the rest: sixty people who know they are being measured, talking inside a population of more than two thousand who now suspect they are.

As communications planning this is competent and cheap. As an intervention it is doing something quite particular, and worth naming plainly. It is not distributing a finding, and it is not teaching a method. It is redistributing a belief about visibility.

That distinction matters because the two produce different behaviour. A finding tells somebody which of their hops was wasteful and gives them a way to remove it. A belief about visibility tells somebody that their hops are counted and leaves the response entirely to them. The first is bounded by what the analysis found. The second is bounded by nothing, which is exactly why it looked like the affordable way to cover a reach gap of that size.

Nothing in it changes the process

Run the workstream against the three pathologies the analysis had already exposed and the mismatch is complete.

Monitors with no asset tag send tickets bouncing between a user and a service desk. That is fixed by tagging the estate. Suppliers using the incident process to handle warranty questions is fixed in a contract and a routing rule. Incidents forwarded to departments with no connection to them is fixed with better classification at intake and a routing table somebody owns. Every one of those is a structural change with a name and an owner, and not one of them moves because two thousand people have a sharper sense of being observed.

So the workstream is aimed at the residual: the discretionary part, where a person who could take four hops takes three because it is being counted. That residual is real. It is also the smaller part of the waste in chains like this one, and the part that regenerates the moment attention moves elsewhere.

The line at the end of the message

The message closes on a threshold: no incident is closed below the average number of steps. I am reading a compressed line in translation, and it can be read two ways. As a rule, it says a closure using fewer steps than average will be looked at. As a statement of coverage, it says no closure escapes comparison with the average any more. Both readings have the same problem in them.

An average is a standard that half the population fails by construction, and it moves as behaviour improves. Set it as the line an incident is measured against and you have built a target that recedes as people approach it, in front of an audience that has just been told somebody is checking. Anyone who has watched a service desk under a moving target knows what arrives next, and it is not a shorter process. It is a shorter record of the process.

The workstream would not have taken a single step out of the chain. It proposed to take away the belief that nobody was counting, and to wait for the step count to fall. That is a cost to write into the business case, not a lever to reach for.

~60

People the chain teams reached directly

2,000+

People directly involved in the incident process

300,000

Steps the business case required to come out of the chain

EUR 3.3M

Conservative annual saving the business case projected

What observation does to the instrument

This is the part I would put in front of anyone proposing the same thing today, and it is not an ethical objection. It is a measurement objection, and it happens to arrive at the same place.

The whole business case rests on a step count. The step count is produced by people logging what they did, in a system, while doing it. Tell those people that the step count is now watched, that a threshold applies to it, and that the watching is meant kindly, and you have introduced a strong incentive acting directly on the recording layer rather than on the work. Steps get batched into one entry. Hand-offs happen by a quick word and appear nowhere. An incident is closed and a fresh one opened rather than carried. Nothing about any of that is dishonest in the mind of the person doing it, and every bit of it makes the chain look shorter while the user waits the same length of time.

That is the trap in an observability programme that reaches for felt observation. The measurement is the asset. Making people feel watched is the fastest available way to degrade it, and the degradation shows up as success. A programme that pulls this lever and then reports the resulting improvement has no way, from inside its own numbers, to tell the two apart.

Which is why I would count it as a cost rather than call it an intervention. If a programme wants the compliance effect anyway, fine, but it should write down that it is buying compliance, say who decided, and hold back a source of evidence the recording layer cannot reach: sampled reality, user-side waiting time, an audit of closures against what actually happened. That held-back evidence is the first artefact a RealAI Consult engagement writes down, because it is the only thing in the file that stays true if the recording layer bends.

The same choice, now made by default

The reason to read a decade-old slide this closely is that the choice on it has stopped being a choice anyone makes deliberately.

The observation this proposal would have had to manufacture, one ordinary internal channel at a time, now arrives for free. Process mining reads the event log directly. Copilots sit inside the work and emit telemetry per interaction. The early and carefully scoped autonomous experiments crossing my desk write a full trace of every action they take on someone's behalf, and that trace, by construction, is also a record of how the person beside them works. Nobody has to leak a success story to produce the feeling. Switching the thing on produces it.

So the question the old document answered in the open, and answered badly, is now answered silently by whoever configures the dashboard. Who can see an individual's trace. Whether it aggregates before it is displayed. Whether it enters a performance conversation, a staffing decision, or an evaluation set used to judge the assistant rather than the person. Whether the people generating the trace were told, in specific terms, what is recorded and what it will be used for.

None of that is exotic governance work: a paragraph of policy, a permission model and an aggregation rule, decided before the first dashboard rather than after the first grievance. The European AI legislation that reached political agreement has workplace uses squarely in its sights, and works councils in several markets will get to this before any regulator does. The programmes I would not want to be running are the ones where the answer to who decided this is that nobody did.

The honest version of what that proposal recommended is that the feeling of being watched was the mechanism, and the improvement was hoped to follow. Compliance comes first and it comes quickly. Improvement is a separate piece of work, on the routing table and the asset tags and the supplier contract, and it does not arrive because anybody felt anything.

Drawn from a short performance-based proposal to a large European banking group on the benefit realisation of a completed step-level analysis of its internal incident handling chain. Figures and the message wording are as stated in that document, in translation. It produced a diagnosis and two recommended interventions, not delivered results: no behaviour change is recorded in it. Reading the second intervention as a cost to the measurement it depends on is ours.