McKinsey published How can the public sector meet the AI moment? in July 2026, and within a few weeks it had become the slide every vendor in this market puts in front of an executive director. We are not going to pretend we are above that — we build AI software for housing authorities, and the report is good for us. But most of the citing stops at the scary maturity score, which is the least useful number in the document.
Here is what the report actually says, and what it means when you are the person who has to sign the contract, answer to the board, and keep 4,000 vouchers moving while the transformation happens.
The headline number is not the finding
The public sector scored 28 out of 100 on McKinsey's AI maturity index. The cross-industry average is 33. Technology, unsurprisingly, sits at the top around 44.
That gap is smaller than the panic suggests. Government is not 10 years behind private industry on AI. It is a handful of points behind an average that includes a lot of companies who also have a chatbot nobody uses. The interesting question is not why the score is low. It is why so many public-sector AI projects that get funded never reach a resident.
The report's answer: fragmented data, workflow integration that never happens, model risk nobody owns, and operating costs that surprise everyone in year two. Every one of those is an execution problem. None of them is solved by picking a better model.
The finding that should change your procurement
This is the number to put in your board packet:
About 70% of programs that redesign an entire service or operational domain reach production, compared with about 30% of projects organized around an individual use case.
Read that twice, because it inverts the way almost every PHA AI project is scoped.
The instinct — and it is a reasonable, risk-managed instinct — is to start small. Pick one narrow thing. Put a chatbot on the website. Automate the after-hours voicemail. Try an AI tool on one form. Prove it works, then expand.
McKinsey's data says the narrow start is the thing most likely to strand. Not because small is bad, but because a single use case has no path to production. It sits next to the real workflow instead of inside it. Staff keep doing the old process because the old process is the one that produces a HUD-compliant outcome. The pilot gets a nice writeup and quietly stops being used in month five.
The programs that make it are the ones scoped to a domain: the whole intake-to-outcome path, with the handoffs, the exceptions, and the system of record included from day one.
What "a domain" means at a housing authority
McKinsey writes for agencies in general. Translated to a PHA, your domains are not hard to name. They are the things that consume your specialists' weeks:
- Annual recertification — notice, outreach, document collection, income verification, EIV reconciliation, the interim requests that arrive mid-process, and a 50058-ready packet at the end.
- Lease-up / RFTA — packet intake from a landlord who has never done this before, blocker detection, chasing the missing W-9 and the unsigned page, rent reasonableness and affordability, inspection scheduling, HAP execution before the voucher expires.
- The phone line, which is not a domain of its own so much as the surface where the other two leak. When a recert stalls, it becomes a call. When a lease-up packet is incomplete, it becomes three calls.
A chatbot that answers "what are your office hours" touches none of these. An AI that reads a paystub but hands you a number with no link back to the source document touches one step of one of them and creates a new verification burden.
The unit of work that actually reaches production is: this case moves from open to a decision your staff can sign, and everything that has to happen in between happens.
The 60% nobody budgets for
The report puts roughly 60% of AI's value in the workflow redesign itself, not in the model. And it offers a ratio that we think is the single most honest line in the document: for every $1 spent on technology, expect something like $5 in adoption, training, and change management.
Most PHA AI proposals we see — including, historically, some of ours — invert this. They are 90% software cost and a line item that says "training included."
If you take one procurement action from this report, make it this: ask every vendor to show you their number for the non-software work. Who configures this to your admin plan? Who sits with your recert specialists in week two when the process feels wrong? Who rewrites the resident-facing language after the first 200 calls show that nobody understands the phrase "interim reexamination"? If the answer is a PDF and a webinar, the $5 is coming out of your staff's evenings.
The staff-trust number is the real risk
Buried further down: about one in five public-sector employees expect AI to significantly change their daily work, and only 31% trust their employer to deploy it responsibly. Across all industries that trust number is 71%.
That 40-point gap is the thing that kills housing authority AI projects, and it does not show up in any RFP scoring matrix.
Your recertification specialists have watched "efficiency" arrive before. They know that the last three systems added steps. And in a public agency, the word "automation" carries a specific implication about headcount that no amount of reassuring email undoes.
The design response — not the communications response — is to make the tool structurally incapable of taking the decision away. That means:
- The AI prepares; a person approves. Eligibility, rent, HAP, and inspection outcomes are staff decisions, always, by design and not by policy.
- Every value the AI produces links back to the source document it came from, so a specialist can check it in two seconds instead of re-doing it in twenty minutes.
- Nothing gets sent to a resident or a landlord that staff cannot see, and nothing is finalized without a human in the loop.
When that is how the system is built, the pitch to staff stops being "trust us" and becomes "look at the packet — it's the work you would have done, already assembled, and you still decide." That is a much easier conversation, and it is the one that determines whether anyone uses the thing in month six.
"Our data isn't ready" is real, and it is not a reason to wait
Every PHA says this, and every PHA is right. The tenant file is partly in the PMS, partly in a shared drive, partly in a filing cabinet, and partly in the head of someone who has been there 19 years. Imports are messy. Phone numbers are stale. Half the applicant records have no date of birth.
We know exactly how real this is. On one authority's imported applicant data, we found zero out of 112 applicants had a usable date of birth or SSN on file — which broke the identity-verification design we had built on the assumption that of course that data exists. We rewrote it to verify against what PHAs actually have: the caller's name and the phone number on their file.
That is the lesson, and it is the opposite of "clean your data first." A five-year data-cleanup program is how you never start. The right move is to choose tools that are built to operate on the data you actually have, and that improve the file as a byproduct of doing the work. Every verified call updates a phone number. Every document collected fills a gap. The cleanup is the workflow, not a prerequisite for it.
The scoreboard McKinsey recommends
The report's fourth recommendation is human oversight for consequential decisions, which in HUD-regulated work is not optional anyway. But its framing of success is worth copying directly into how you evaluate any vendor:
Leaders measure success by whether residents can tell the difference — through tangible progress such as shorter wait times.
So the metrics that count for a PHA are not "calls handled" or "documents processed." They are:
- Days from recert notice to a signed 50058.
- Days from RFTA received to HAP executed, and how many vouchers expire unused.
- Percentage of calls a resident gets an answer on the first try, at 8pm, in their language.
- Hours per week your specialists spend chasing paper instead of making determinations.
If a vendor cannot tell you which of those numbers their product moves, and cannot show you the before-and-after at an authority roughly your size, the report has a word for what you are buying. It is "pilot."
What we would do in the next 90 days
Not a transformation program. Three things:
- Pick one domain — recertification or lease-up, whichever has the worse backlog — and map its actual current path, including the exceptions. Not the flowchart in the admin plan; the real one, with the sticky notes.
- Count the leaks into the phone line. Pull a month of call logs and classify them by which stalled case caused them. This number is usually the most persuasive thing in the entire business case, and you already own the data.
- Run a real caseload through one tool, not a demo dataset. Ask for a pilot scoped to a whole journey with your files, your admin plan, and your staff doing the approving. If it works, the metric moves in weeks and you have the evidence for a full procurement. If it does not, you found out for the price of a small purchase.
That third one is what we do, and we will say plainly that it is a sales pitch. It is also, per McKinsey's own data, the scoping that reaches production about 70% of the time instead of 30%.