Performance augmentation

The request names a solution. The diagnosis names the cause.

“Train all thirty hotels” on data nobody trusts

A request arrives to roll out service-excellence training across thirty properties before peak season. Guests seem less satisfied, staff turnover is high, and the regional managers say outright that the brand-standard audit scores are gamed and not to be trusted. There is, in their own words, no data anyone would rely on.

What it did with nothing solid to stand on

The interesting part is not that the tool found an answer. It is what it did when there was almost nothing to find one with, which is the situation a real analysis is usually in.

It refused to convert anecdote into a finding. The environmental causes (a central-reservations system double-booking at peak, below-market pay at the high-turnover sites, unstable floor management, the franchise structure that resists a mandated rollout) went in as hypotheses, each with its reason for being only that: reported by operations but not yet verified against system data; named by HR but not yet checked against local wage figures. And it put the disagreement between HR, operations and the VP to use rather than resolving it, because a three-way blame is itself a signal about where authority and information sit.

It also named the missing measurement as a cause in its own right. The audit scores are gamed and distrusted, so the group cannot locate the gap, track a change, or make the case for any fix, which is underperformance sustained, not a data inconvenience.

So the honest first deliverable is not a training plan and not a verdict. It is a measurement the group does not yet have, and a short diagnostic at the best and worst properties before a single euro is committed. That is a smaller, truer answer than the one that was asked for, and it is available in an afternoon rather than a quarter. And because a run costs nothing, the same analysis is worth making on the request that turns out to be justified: a yes with evidence attached is worth as much as a no.

The same instrument, on either side of the same reflex

AI does not replace the professional’s judgement; it strengthens what they already do: augmentation, in the sense of the series this number grows out of (Arets, 2023). Two audiences meet the same reflex from opposite sides, and the same tool serves both.

From the request to the reports, in five moves

The tool guides a conversation, not a form. The hotel request from page one, carried through the five moves, and what the person sees on the screen at each.

How sure it is, and how it shows a source

Two mechanisms separate this from a system that always sounds certain. The first is that it rates its own diagnosis. The second is how it puts a source in front of you.

The second is the source. Take a claim an analysis might produce: orders above the approval threshold wait an average of eleven hours. One tool ends that sentence with a reference; Source: Order policy, section 4. Another ends it with the line it actually read: “Orders exceeding €25,000 require regional approval; approval is available between 09:00 and 13:00 on working days.” The first tells you where it says it got the sentence; the second lets you see whether the source says it, and here it does not, because the policy sets the threshold and the window and says nothing about eleven hours.

Everything stands ready. The person decides how and what

Eight tasks in one analysis. The assistant prepares all of them and presents the result; the line between machine and person is not about current capability and will not move as models improve, because it is drawn on who holds the consequence.

The reason for that line

The line is not drawn out of caution, and it is not a concession to today’s technology that will move as models improve. It is drawn on who bears the consequence of being wrong.

One distinction sharpens where it falls. An assistant can be active: it answers what it is asked, retrieves, computes, drafts on demand. It can also be proactive: noticing that a question is still unanswered, that the measure changed definition in April, that this organization established the same cause eighteen months ago. Proactive is where most of the value sits, and also where the temptation lies, because an assistant that anticipates well starts to look like one that could decide. It cannot, for the reason below.

An analysis that lands on the wrong cause costs an organization a program, a quarter and some of its patience with the function that proposed it. The decision that follows is management’s, or the team’s, but in either case a person’s, which is Deming’s point about where responsibility for the system sits, and it does not move because an instrument got better (Deming, 2000). Everything in the right-hand column of the table has that property: the choice of measure, the judgement that a condition is binding, the decision to report a finding a director will not want, the signature. Those are not the hard parts because they are computationally hard. They are hard because being wrong about them has a cost that falls on a person.

One diagnosis, four reports for four rooms

The largest thing it delivers is not a saving; it is a grounded diagnosis. It holds the whole method, will not run past an unanswered question, and states its confidence in what it has produced. That diagnosis comes out in four forms, one for each room it has to enter.

At a four and at an eight, the same instrument

The hotel group scored a 4 because there was almost nothing to stand on. Pointed at a global specialty-chemicals group (17,800 people, 31 sites, two years of records) the same instrument lands at an 8. The distance between those two numbers is what the rating is for.

The scale, as the tool states it, runs in five steps: 2, a provisional framing (the problem framed, little or no data yet; 4, a working hypothesis) the most plausible explanation from what is available; 6, indicative (the direction clear, worth acting on with caution; 8, substantiated) strong evidence across multiple sources, one focused step from 10; and 10, triangulated (hard numbers, qualitative evidence and independent sources confirming each other). An 8 is strong enough to act on. What the two steps toward a 9 buy is precision, not direction.

The eight, in detail and what it ruled out

A culture program was on the table because people were afraid to challenge their managers. The evidence read differently: the fear was real, and it was downstream of structures nobody had put beside each other.

The consequence structure rewards production over risk prevention: plant-manager pay weights EBITDA at 35 percent and output at 25, with safety at 10; of 26 leaders promoted in three years, 22 had top-half financial results and 9 had top-half safety-climate scores. Corrective-action management has effectively broken down: open actions grew from 4,800 to 13,700, average closure from 41 to 96 days, and 15 of the 19 serious events involved a known abnormality. Near-miss reporting fell by more than a third while employees still identified 92 percent of deliberately planted anomalies. People see the risks. Reporting them stopped paying.

And one finding carries the label established rather than hypothesis, triangulated across simulations, incident reviews, assessment records and the interviews: knowledge is not a material cause. Training completion runs at 99.3 percent, practical assessments at 94.7, and in simulated abnormal conditions 91 to 97 percent of operators act correctly. The proposed workshops address a knowledge framing of a problem that is not a knowledge problem: the argument of this whole series, in the tool’s own labels, in the client’s own records.

The comparison cuts both ways. One plant under the same corporate rules keeps justified stop-losses out of shift results and closes high-risk actions in 19 days against a company median of 83: near-misses up 28 percent, serious events down 43. And a fourteen-month pilot of exactly the proposed workshops lifted survey scores nine to fourteen points while shutdowns rose 7 percent: they listen better, said an employee, and the operating constraints have not really changed. The learning measure moved and the business measure did not: number 11’s distinction, in the client’s own words.

What the room sees, from the deck the tool drafted

The Slides version of the same diagnosis, as it left the tool. Two slides carry the whole argument, and they are worth describing, because the room decides on them.

The person presenting edits what does not fit, and signs. That is the division of labor in one scene: the assistant drafted the deck and defended every figure on it; the human decided what the board would see.

Grounded in sources, or not at all

From here on, the paper is for the person who signs. One design decision governs everything else, and it is not a technical preference. It is whether the assistant answers out of a model’s absorbed training, or out of documents it retrieves and can point at, sentence by sentence.

The first produces fluent, confident, plausible text with no way for anybody to check it. Applied to a method, where the whole value is that a finding can be traced back to a record, that is not a weaker version of the same product. It is a different product, and an actively dangerous one, because a fast wrong diagnosis travels further inside an organization than a slow right one.

So the assistant answers only from a defined corpus: the method as this series sets it out, the reference library behind it, and the organization’s own material where it has been made available. It may reason, and it may not invent, and where the corpus has nothing to say, the honest output is that it has nothing to say.

A method is a structure, not a pile of documents

What sits behind the assistant matters more than the model in front of it, and it is not a folder of white papers. A method is a structure of dependencies, which is why the tool classifies a cause by factor, physical resources, structure and process, information, and by level, rather than describing it in prose. Those are Gilbert’s conditions and Rummler’s levels, made operational.

The third question cannot be answered before the second, because it needs the measure. A factor rules certain remedies out and leaves others standing: Gilbert’s structure, a set of relations rather than a list (Gilbert, 1978). A cause at one level cannot be reached by an intervention at another, which is Rummler and Brache’s point about levels (Rummler & Brache, 2013). Those are relations between objects, and a system that stores only the text has thrown them away, which is why an assistant built on a document search can quote this series accurately and still let a user skip the step that made the quotation mean something.

Governance, in three decisions

Anything of this kind used in professional decisions now sits inside a governance frame, and in Europe inside a regulatory one still being phased in. The paperwork will be substantial. The substance is three decisions.

What this does not claim

Four limits, and the tool states the first itself: it delivers diagnostic advice, not a guarantee of results.

• It replaces neither the L&D team nor the consultant. It makes the diagnosis routine instead of weeks of specialist work. The method is fifty years old; what is new is the access, not the analysis, and a person still has to establish which cause binds and own the decision.
• It is only as good as the data it is given. With the figures in place the reports are ready in an afternoon; without them the assistant asks for more, and where they cannot be got it rates its own confidence down rather than guessing. A high score on thin data is not on offer.
• It is far too early for results claims. Assistants of this kind have existed a handful of years. There is no synthesized evidence about what they do to the quality of professional analysis, and this paper makes none, what it sets out are design commitments, checkable today, and an account of our own use, which is experience and not evidence.
• The part that matters most cannot be automated. Establishing which cause is binding requires somebody to watch the work and ask the person doing it. That does not transfer, and no amount of retrieval reaches it.

Six lines, before you buy or build anything

Thirty minutes, on any assistant being proposed to you: ours included. Line three is the one that separates a product from a demonstration.

Two problems, one method, sixteen papers

Two models carry the method. Gilbert’s six conditions ask what kind of cause; Rummler and Brache’s three levels ask where it sits. Both are sound and this series uses both, and it leads with Gilbert, because his unit is the accomplishment, which is what an organization already counts, and because his six conditions are the categories an organization can actually change.

Where the method comes from

Two further pieces close the set and carry no number, because each points at all the numbered papers. Before the guide sits Arets L&D Business Models, the destination piece for the people who receive the requests: four models a learning function can run, and the route from HR-driven to business-driven. Last sits Solutions that cannot work, the field guide, ordered by the sentence that lands on your desk: seventeen forms a performance problem arrives in, each with an address in one of the numbered papers. Three orders in all: the destination you choose, the question you are on, and the form that landed on your desk.

Sixteen papers, six parts

Number 01 is the map and number 02 is the foundation. From number 05 on, every number is depth on exactly one step of the method. The filled block is the paper you are holding; the two unnumbered pieces close the set.

Where every claim in this paper comes from

This number deliberately cites no research about assistants of this kind. The published work is a few years old, it is not yet synthesized, and the figures in it belong to product versions that no longer exist. Number 02 sets out the evidence rule this series follows and why; the three books named earlier are listed in full there.