What Happens When the Machine Learns the Room
A recent AI experiment is teaching us something important about process, power, and the organizations AI is already moving into, if we are willing to look closely.
Housing and community development has always had a document problem. Not a shortage of documents. Quite the opposite.
We have requests for proposals written before anyone has decided what they actually need. Scopes of work that describe the solution before the problem has been fully understood. Grant applications built around timelines that quietly assume site control, environmental review, Section 106, engineering, procurement, underwriting, and half a dozen other things will somehow line themselves up after the award arrives.
Then the project gets funded. Everybody celebrates. And somebody asks the question that should probably have been asked six months earlier: Are we actually ready to do this?
That is the predevelopment trap. The money can be real. The need can be real. The project can be worthwhile. And the system surrounding it can still be nowhere near ready to move.
A bad RFP is a particularly clean example. If the organization has not clearly defined the problem, the authority, the information it already has, the decisions that still need to be made, or what success actually looks like, a consultant can only do so much with the assignment. The procurement may be compliant. The deliverables may arrive on time. Everyone can technically do exactly what the scope requested and still discover that the scope requested the wrong thing.
Generative AI makes that problem more interesting because it removes so much of the friction that once exposed it. Give AI a weak premise and it can produce a remarkably strong-looking RFP. Give it an incomplete development concept and it can produce the implementation plan. Give it a grant agreement and a pile of meeting notes and it can build the dashboard, the project schedule, the responsibility matrix, the risk register, the board memo, and the weekly status report.
Some of that is genuinely useful. We use these tools precisely because they can help surface dependencies, compare requirements, organize complicated information, and find the question buried underneath forty pages of process. But there is a difference between making complexity visible and making bad thinking presentable.
AI cannot tell us that an organization has commissioned the wrong study unless someone gives it permission to question why the study exists. It cannot fix an RFP whose central assumption nobody is willing to challenge. It cannot complete an environmental review by adding “environmental review” to a project-management dashboard. And, it cannot turn an award into a buildable housing project simply because the timeline now fits nicely on one page.
That is what has been bothering me. We talk a great deal about what artificial intelligence can do inside organizations. We spend less time talking about what those organizations are teaching the technology when it gets there.
Because AI does not enter an empty room. It enters a room with procurement habits, reporting structures, grant requirements, old decisions, missing decisions, incentives, personalities, authority problems, and occasionally an RFP that should never have made it out of draft.
Read our most recent Field Notes on what happens when AI enters organizations with their own habits, power structures, and bad assumptions. AI is not the dysfunction. It is the accelerant. The real question is what we are asking it to accelerate.
