Put a Number on It: The Target That Turned Knowledge Sharing Into a KPI

Every knowledge management conversation I am in this year ends in the same request. Somebody wants search across the whole document estate, or a summariser that reads the last five proposals before the team writes the sixth, or a cautious pilot with one of the new language models answering questions from internal material. Deployment tooling is available, pipelines are assembly work, and a demo now costs a week rather than a quarter.

What almost nobody asks first is what fraction of the documents are in the estate at all.

I keep returning to an internal change strategy written by a global professional services organisation about itself. It was not our engagement: it is a firm's own team planning how to make its people share what they produce. Read as change management it is competent, slightly weary work. Read by somebody who will later point a retrieval system at that estate, one clause on one slide matters more than everything else in it.

A value with no denominator

The programme's own diagnosis is unusually honest about why previous attempts went nowhere. At senior level it recorded people saying that knowledge-sharing performance was not measured, and that they therefore did not act on it or give feedback based on it. One rank down the statement is blunter: if it is not measured it will not be done. Elsewhere the diagnosis lists misaligned incentives, and records management not seen as a priority or as mandatory.

None of that is a shortage of belief. Everyone in the document agrees that sharing is good. What the organisation lacked was a denominator. Sharing more is not something you can miss, because there is no arithmetic in which you fall short of it. It costs nothing to endorse and nothing to ignore, which is exactly why it had been endorsed for years and ignored for the same years.

What a number does that a statement cannot

The target, when it appears, is one clause: 70 percent of proposals, shared client documents and the short records describing completed work, uploaded by the close of the year. It is measured per practice. It is read off a reporting system the organisation already operated, not a new dashboard commissioned for the purpose. It has a review slot, the December leadership meeting sitting alongside close of books, and the practice head is the person in the room when the gap is discussed.

Four properties, none of them technical. A denominator, so the percentage counts something. A unit of account, the practice rather than the firm, so nobody hides inside an average. An owner who has to speak. A date when the number stops being a forecast. That combination is the finding, and it is the moment a cultural programme quietly becomes an operating one, with a metric definition, a data path and a review cadence. It is the same machinery RealAI's Platform work stands up around any capability a client wants to claim.

The baseline the number landed on

A target is only interesting against the position it starts from, and the document supplies that: a table with one row per reporting unit and one column per document class, measured over engagements closed in the first half of the year at a stated data cut. The ranges are stark. Proposals ran from 24 percent to 77 percent. Shared client documents ran from zero to 26 percent. The completed-work records ran from 8 percent to 35 percent. Exactly one cell in the whole table sat above the 70 percent line, and it was a proposals cell in a single unit. Two cells read zero: those units had captured none of that class at all in the measured period.

So the 70 percent line sat above every proposals figure in the table except that one, above every shared-client-document figure by at least 44 points, and above every completed-work figure by at least 35 points. Six months to close that, in an organisation whose senior people had told the diagnosis that the behaviour was not measured and therefore not acted on. That is not a criticism; ambitious targets are a legitimate instrument. It is a caution about how such a number reads later. A percentage set that far above the observed baseline is intent that has been given arithmetic, and it stays intent until the first measured report lands.

The baseline also covered only part of the estate: several locations had no measured position at all, and the plan committed to surveying every one of them before the autumn leadership meeting. A firm-wide percentage computed over the units that happen to have data is a different number from one computed over the firm, and the difference usually flatters.

Two yardsticks in one document

Twenty slides after the 70 percent target, the same document states its principal measure differently: whether all engagements completing in the year have project documents. That is a binary test across every engagement. The other is a proportional test across three document classes. They are not the same measure and they cannot both be primary.

This is not carelessness. It is what drafting looks like when two people write two sections. But it is expensive in a predictable way: at the December review, with two available yardsticks, the presentation uses whichever reads better for the practice in question. Nobody has to cheat for that to happen; the ambiguity does the work.

The fix costs an afternoon. Pick one primary measure, write its denominator down in a sentence, and store the query that produces it beside the definition, so the number has one path from source system to slide. Everything else is a secondary indicator, reported but not argued over.

The review date arrived before the definition

The planning section closes with an open-questions register: eight questions, each with a named owner, six of them carrying a due date. One asks how the audit tool and the KPIs would actually be measured and published in the reporting system. Another asks how the measure would connect to the appraisal process and the performance scorecard.

Both were still open in the same month the target was already scheduled into leadership communications. That sequence survives if the definition lands before the first published report, but it is the most common failure I see in capability programmes: the number gets announced on the communications timeline while the measurement path gets built on the timeline of whoever owns the reporting system. When those diverge, the first published figure is whatever the tool happened to be able to produce, and that accidental definition becomes the definition forever.

The machinery a number drags behind it

The plan behind the target runs to thirty-five activities across three pages, with a column each for the accountable person, the due date, completion, notes, and whether extra resource or funding is needed. Five activities have the accountable column left blank. One is ticked complete. The document is a draft circulated for discussion inside the team that wrote it.

Those activities are the real cost of the number: animations contrasting good and poor practice, posters and message templates, a campaign site, three newsletters, cascade emails drafted for every level of the hierarchy, training courses updated so the process and the tools are taught consistently, briefings for managers on reviewing the behaviour in a year-end conversation. Teams that write the target and skip that list get the number without the participation, then conclude that targets do not work.

What happens when a consequence attaches

The communications sequence escalates deliberately. Encouragement and role reminders first. Then a message that not sharing is no longer acceptable, with good and poor practice described explicitly. Then confirmation that KPI reporting is live. Then the year-end position of each practice against the target. Then, in the new year, briefings for leaders on handling people who still have not complied, and the behaviour written into year-end reviews and objective setting.

That is how you make a number move, and the plan is honest about being a lever rather than a nudge. It is also the moment the number stops being a neutral observation. Any measure with a consequence attached becomes a managed measure. Upload rates will rise, because uploading is the cheapest way to satisfy the person asking. Whether what gets uploaded is worth retrieving is a separate question, and the plan carries no measure for it. A participation target says how much of the expected material arrived. It says nothing about whether those documents are final versions, correctly attributed, or readable by anyone who does not already know what they are looking at.

70%

Firm-wide capture target set per practice

1

Baseline cells above that target, out of the whole table

0%

Two units' capture rate on one document class

35

Work-plan activities behind the single number

Why this comes before anything model-shaped

Suppose the search project, or the careful language-model pilot, had gone ahead against that estate at the measured baseline. Proposals would have been reasonably represented in some units and thin in others. Client deliverables would have been absent entirely in two of them. The records describing completed work, the exact material anybody wants when writing a new proposal, would have been present for fewer than one engagement in five across most units.

A retrieval system does not warn you about that. Ask it what the firm has done in a sector and it answers confidently from whatever it holds, so the units that captured nothing simply appear not to have worked there. The failure is silent, systematic, and indistinguishable from a correct answer unless you already know the coverage per class per unit.

Which is why coverage belongs to the operating model rather than to the data team, and why Synapsa reports participation per unit before it reports anything else. A number, per unit, per class, with an owner and a review date. This document got that right without anybody in it thinking about machine learning at all.

The corpus completeness target has to exist before the model target. One is a management decision that costs an afternoon; the other is a programme that quietly fails without it.

The awkward part is that the number is the easy half. Writing 70 percent takes a minute. The definition, the query, the report, the owner, the review and the thirty-five activities behind it are the actual programme, and they are what gets cut when the timeline tightens. Cut them and the number survives on the slide with nothing behind it, and somebody eventually concludes that knowledge sharing cannot be measured.

Observations are drawn from an internal knowledge and information management change strategy authored by a global professional services organisation about its own operations, at a draft circulated for discussion within its own team. It is a plan and a set of targets, not a delivered outcome, and it is not our work: nothing here is a result we produced. The baseline arithmetic and the reading of the two competing year-end measures are ours.