You are setting data quality thresholds for a dataset with five consuming teams. How do you decide the numbers?

Not by the data team, and not one number for everyone

Fitness for purpose differs by consumer. Ninety-five percent completeness on an attribute is fine for a marketing segmentation and unacceptable for a field feeding a regulatory return.

Ask each consumer what breaks at what level, in their own terms. The conversation produces better numbers than any standard, and it surfaces consumers whose expectations were never stated.

Set to the strictest requirement, and say who set it

The dataset's guarantee is the strictest consumer's need, because a weaker guarantee fails them. Record which consumer drove each threshold — it makes the number defensible and it makes renegotiation possible if that consumer goes away.

Where one consumer's requirement is dramatically stricter than the rest and expensive to meet, that is a signal they may need their own derived dataset rather than the shared one.

Two thresholds per check, not one

A warning level notifies the owner; a failure level stops downstream processing. The gap prevents halting over a marginal fluctuation while still catching a genuine collapse.

Decide the failure behaviour in advance

Halting protects downstream correctness and makes yesterday's dashboard stale. Continuing with a quality flag keeps things moving and requires every consumer to check the flag, which they will not.

Choose per dataset, before the first failure. Deciding during an incident, with five consumers on a call, is how the threshold ends up being lowered instead of the data being fixed.

The thing that makes it a contract

The thresholds are published, monitored, and breaches are notified to consumers automatically. A quality standard the consumers cannot see the current state of is a policy rather than a guarantee.