After a year of stewards in name only, a company funds the role properly: 18 stewards at 20% of their time, written into objectives, with authority to refuse an access request. Two quarters later definitions are current and the analytics team is complaining loudly. What did the organisation buy, what is it paying, and when does that bill arrive?

What is gained

Three things that the nominal version never produced. Definitions that are current, because re-attestation now competes for protected time rather than for goodwill. Refusals that hold, because a steward with the role in their objectives can say no to a director and point at the mandate. And a named route for a dispute, which converts a standing argument between two functions into a decision with a date on it.

The measurable change is usually the share of catalogue entries with a reviewed description in the last 180 days, which moves from the 20% to 30% range that nominal programmes reach to something above 80% within two quarters.

What is paid

3.6 FTE of business capacity, not platform capacity. This is the expensive kind: it is drawn from the people who understand the domain, who are the same people the domain needs for its own work. A platform engineer is replaceable from the market in a quarter; the person who knows why the claims system has two date fields is not.

Second, serialisation. A steward with authority to refuse is now a synchronous dependency in the access path. If 18 people each handle 6 requests a week at a 2-day turnaround, median queue time looks fine and the tail does not: one steward on leave takes their domain's median from 2 days to 9.

Third, capture. A steward drawn from finance will, over a year, resolve ambiguity in favour of finance's definition, because that is the definition they are measured against.

When the bill arrives

Not in the first quarter, when enthusiasm covers the cost. It arrives at the first quarter-end crunch, when every steward is pulled back to close the books and the access queue triples in a fortnight. It arrives again at the first reorganisation, because the 20% allocation was a manager's promise, not a headcount line, and it does not survive a new manager.

How to keep the option to reverse

• Put the steward on the exception path only. Default-approve access to assets labelled Internal or below with a recorded purpose and a 30-day expiry; route only Restricted assets to a human. This typically removes 80% to 90% of the request volume.
• Publish queue time as a platform SLO, for example 2 business days at p90, and treat a breach as a resourcing signal rather than a steward failing. Without this the queue is invisible until analysts route around it.
• Rotate the domains a steward covers annually, which costs some depth and blunts capture.
• Keep the fallback written down: if the 20% is withdrawn, which controls degrade first and what replaces them.

When this is the wrong answer

Below roughly 15 data practitioners, a funded stewardship programme is heavier than the problem. One accountable owner per domain recorded in the catalogue, plus a weekly triage of anything contested, gets the same decisions without the queue. The programme starts paying when the number of people who can create a number exceeds the number of people who can be in one room.