Less budget defence than people expect, and far more deciding.
A knowledge graph initiative fails for governance reasons far more often than technical ones. The governance failure is almost always the same. Nobody with standing was available to settle what terms mean. A sponsor who secures funding and then disappears has solved the easy problem. The sponsor's real job is making decisions reachable.
That reframing changes who the right sponsor is. It is the person who can convene the people who hold the meaning and require an answer from them, which is rarely the same as the person who owns the data platform budget.
Four things, and only the first is money.
| Commitment | What it costs | What happens without it |
|---|---|---|
| Funding | Platform and a small team | Nothing starts |
| Expert time | Hours from the people who hold the meaning | The model gets authored by people guessing, and experts reject it later |
| Decision authority | Your own time, on escalations | Contested terms stay contested and the model stalls at the interesting parts |
| A standing owner | A named role beyond the project | The asset decays within a year of launch |
The third is the one sponsors underestimate. Most of the terms worth governing are contested precisely because two reasonable teams disagree, and somebody above both has to be willing to rule. Delegating that back down produces a model that covers everything uncontroversial and stops exactly where the value was.
Start from a live failure. A capability story will lose.
The capability pitch describes what a knowledge graph is and what it enables. It is accurate and it loses, because it asks a committee to fund an abstraction. The failure pitch takes one argument the business had in the last quarter, shows what it cost, and proposes fixing that specific thing.
The fourth item is the test. If you cannot name who decides each term before starting, the initiative is not ready, and discovering that during the pitch is far cheaper than discovering it in month four.
One governed term, consumed by one real system, with the decision recorded.
That sounds unambitious and it is the right target. The thing being proven is not that the technology works. It is that a decision can travel from the person who holds it to the system that needs it, with a record attached, and most organizations have never actually done that end to end.
Once that path exists, adding terms is incremental and the argument shifts from whether to how fast. Attempting a comprehensive model in the first ninety days produces something unused, and it burns the expert goodwill you need for everything after.
How to evaluate a knowledge graph platform covers the tooling decision, which should come after the scope and not before it.
Four patterns, all of them visible at kickoff if somebody asks.
All four are predictable at kickoff, and a sponsor who asks about each one in the first meeting will catch most of them. The fourth is the most common and the least visible, because the initiative looks successful right up until it stops being true.
Whoever can convene the people who hold the meaning and require an answer from them. That is a narrower test than seniority and it rules out some obvious candidates.
The workable arrangement in most organizations is a day-to-day sponsor with the experts, plus a named escalation point who can settle a cross-unit dispute inside a week. One without the other produces either a stalled model or an absentee one.
What should you watch after launch?Three numbers, none of which is coverage.
Time to decide. How long a proposed term waits for its approver. Rising latency is the earliest warning, because people stop proposing long before anything looks broken.
Systems resolving against it. If consuming systems keep local copies, the asset is decorative regardless of how complete it is.
Unresolved terms. What systems needed and could not find. This should grow, because growth means the thing is being used against more of the business.
How to measure whether your AI answers from governed data covers the instrumentation.
An hour a fortnight for the first quarter, and almost none after that if the first quarter goes well.
The time is not spent reviewing the model. It is spent on two things: unblocking a decision that two teams could not settle, and visibly caring that the work is happening. The second matters more than it should. Expert time gets diarised when the experts believe the sponsor is paying attention, and quietly deprioritised when they do not.
After the first quarter the pattern should invert. If you are still ruling on terms every fortnight in month six, the decision rights were drawn wrong and the fix is to push approval authority down to named domain owners, never to keep absorbing it.
Sponsoring this is mostly about making decisions reachable. Fund it, commit expert time in the diary, stay available for escalations, and name a standing owner before launch.
TopQuadrant has watched this succeed and fail across a lot of enterprises, and the pattern is consistent. The initiatives that work start from one live argument and one named decider. Our knowledge graph platform is built around that loop: decide, record who decided, make it resolvable. The technology is the straightforward part.