The situation in the room
A vendor demonstration ends and someone asks who is accountable if the system disadvantages a protected group. The room promises to circle back, then procurement continues.
The ethical question arrived after momentum, budget, and executive attention had already made the decision feel irreversible. The work looked active from the outside. Inside the organization, the unanswered decision was still controlling what happened next.
Who is affected, and were they in the room?
Name the people the system acts on, not only the internal users operating it. Then ask whether anyone with that lived perspective shaped the decision before it became expensive to change.
Absence is not automatic disqualification. It is a stated limitation. The organization should know when it is proceeding on assumption rather than input.
What happens when it is wrong?
If the framework never makes the room uncomfortable, it has not reached the decision.
Design recourse before complaints arrive. Under pressure, organizations default to defending the existing process rather than examining it.
What data made this possible?
Ask what data trained, tuned, or informs the system, whether the organization has permission for the use, and what the vendor can disclose. Record a refusal to answer as a material fact.
Unknown provenance does not become known because the procurement document uses confident language.
Who benefits, and who carries the burden?
Automation redistributes effort. A simpler customer path can leave frontline staff handling only the most difficult cases. A faster decision can move the work of correction onto the person affected.
Name both sides of the exchange. Aggregate efficiency can conceal concentrated harm.
Could we explain this publicly?
Ask whether a reasonable person would accept the plain account of how the decision works, what it uses, and what happens when it fails.
If the room hopes never to explain the practice, the discomfort belongs in the decision record now.
What good practice looks like
A framework needs owners, deadlines, evidence, and a route to stop. Without those, it is an artifact rather than a control.
Good practice is visible in ordinary work. People know what they may do, what they must check, who can decide, and where an exception goes. Leaders hear the same account from different teams because the framework is shared rather than remembered differently.
The aim is not perfect control. The aim is a system that notices when reality changes and knows who has to respond. That creates movement without pretending uncertainty has disappeared.
Make the answer operational
Start with the smallest version of the decision that still changes behaviour. Put it in plain language and test it with the people who will use it under pressure. Ask them what they would do, what they would record, and who they would call when the ordinary case stops being ordinary. Their hesitation will show where the instruction is still abstract.
Then connect the decision to the systems around it. A rule without training becomes a document people forget. Training without authority becomes advice people cannot follow. Authority without review becomes a permanent answer to a changing question. The three have to move together, with one person accountable for keeping them connected.
Review the result through actual work rather than confidence surveys alone. Look at a sample of decisions, outputs, exceptions, and escalations. Ask whether the intended boundary held and whether people found a hidden route around it. When the evidence changes, revise the operating choice openly and record why. That is not inconsistency. It is governance doing its job.
What to do on Monday
1. Put the seven framework questions against the latest AI purchase. Keep the action small enough to complete, specific enough to observe, and connected to a named decision rather than a general ambition.
2. Name every answer that is still missing. Keep the action small enough to complete, specific enough to observe, and connected to a named decision rather than a general ambition.
3. Assign each gap to one person and one date. Keep the action small enough to complete, specific enough to observe, and connected to a named decision rather than a general ambition.