Twenty-One Ways In

At a glance

---

The question they started with

Habitat Michigan is a state support organization. It runs its own housing programs directly and it backs 43 local affiliates with technical assistance, advocacy, and program resources. Roughly fifteen people carry all of it.

Everyone there already assumed AI could help somewhere. What they needed was a specific answer about where and in what order, because a team that size can run one initiative at a time and cannot afford to spend its first three months on a complicated software build.

An audit produces that answer for one particular organization. It works from these systems, this staffing, and the way the work actually moves through the building.

---

How the audit ran

1. Research. Existing documentation, the org chart, and program descriptions were read first and turned into a structured company profile: what each program does, who owns it, what systems it runs on, and where friction was already known to exist. The point is to arrive at the kickoff already informed, so no one spends an hour explaining what the organization is.

2. Kickoff. A working session with leadership on May 18 to confirm what the research got right, correct what it got wrong, and settle the interview list. The research always misses something, and the kickoff is where that surfaces.

3. Interviews. Six sessions between May 22 and June 4 with five department leads, covering all five programs. Each person was asked to walk through their real work rather than their job description, including the parts everyone knows are inefficient and no one has time to fix. Where a process ran through a spreadsheet or a system, they shared their screen and walked through it step by step.

4. Opportunities. Every friction point raised in an interview was assessed for whether AI could actually address it, then written up with the quote that produced it. Twenty-one made the list. Each was scored twice, once for impact and once for difficulty.

5. Blueprint. The 21 opportunities were sequenced into a phased plan, each one carrying its evidence, a before and after description of the process, what would have to be true for it to work, and the specific next step to test it.

---

What surfaced

Five patterns showed up across departments. Every number below is what staff reported in their own interviews.

The same information gets entered twice

Housing counselors write case notes and action plans in one set of forms, then copy that information by hand into the add log in the client management system, their system of record. They do it because uploaded attachments do not display clearly enough to be usable as client history, so the add log is the only place the story actually reads.

That is 30 to 45 minutes of documentation after every appointment. The program was down to two counselors at the time of the audit, absorbing the caseloads of the people who had left.

“Counselors draft detailed case notes and action plans using multiple forms before copying this information manually into [the client management system’s] add log.” — Housing and Family Services Manager

The exceptions eat the calendar

Down payment assistance closings run smoothly when the lender follows instructions. When they do not, the file goes into a loop of manual review, an email to the affiliate, a wait, and another review of the corrected documents. The errors repeat: wrong sale price, wrong cash to close, missing signatures.

Volume is three or four closings a month, with peaks of eight in three weeks. The distribution of effort is the finding.

“It’s probably even 90/10. The 10% that are problematic take up 90% of my time.” — Intake and Loan Processing

Every program keeps its own version of the truth

Each program tracks through its own spreadsheet, folder, or email thread. Confirming where a client stands means checking several of them and reconciling what they say. Documents land in the wrong place often enough that finding them is itself a task, and in one case a client had been uploading loan documents into the housing counseling folder for weeks.

Staff reported roughly two and a half hours a week just monitoring folders for newly arrived documents, and three to five hours a week on manual cross-checking.

“Clients sometimes don’t know where to put them, and then we’re looking for them and not really knowing where to find them.” — Lending Manager

Compliance checking is done by reading

Closing disclosures, loan estimates, and appraisals are reviewed by hand against HUD and USDA requirements. The rules are concrete and written down, which is what makes this notable. The work itself is a person reading a document against a list, catching an error late, and starting a correction cycle that did not need to happen.

Coordination work has become the job

Scheduling a single intake appointment takes 15 to 60 minutes of back-and-forth, a burden created when a system migration removed the client self-scheduling tool the team used to have. Client-submitted documents arrive as phone photos, often in HEIC format, and turning them into USDA-ready PDFs runs 2 to 4 hours per client. Inbound email consumes 2 to 3 hours a day in one program and, in the words of the lending manager, “a majority of my day.”

None of this is anyone’s actual job. All of it has to happen.

---

Scoring it twice

Impact tells you what to want. Difficulty tells you what you can finish.

Every opportunity was scored on both. For an organization with a technology department, sorting by impact alone is reasonable, because someone’s whole job is turning hard things into shipped things. For a nonprofit with five to fifty staff and no dedicated technical person, sorting by impact alone puts everything that cannot be completed at the top of the list.

Here is how the 21 landed.

Twelve of the 21 are high impact. Only four of those twelve are also low difficulty, and those four are the ones that can be running inside a week, built by the person who has the problem, without a purchase order.

The four at the high impact and high difficulty corner are the ones an organization actually wants. Each of them needs systems connected, data cleaned up, and money approved before any AI is involved. That is worth knowing before you start.

---

The full list

High impact, low difficulty

High impact, medium difficulty

High impact, high difficulty

Medium impact, low difficulty

Medium impact, medium difficulty

---

What each opportunity carries

Each of the 21 opportunities was written up in seven parts:

• The problem, described in the terms the person who has it used
• Evidence, meaning the direct quotes from the interviews that produced it, attributed and timestamped
• The proposed approach, at the level of what it does rather than what to buy
• Before and after, the current process step by step against the proposed one
• Fit assessment, covering why it suits this organization and what the real concerns are
• What would need to be true, the conditions the opportunity depends on. If they don’t hold, it won’t work, and the write-up says so up front.
• The next step, a small, low-risk test to try before committing to the full build.

---

What Habitat Michigan walked away with

A company profile covering every program, system, and owner. A one-page analysis for each department. The prioritized list of 21 opportunities, each traceable to the conversation that produced it. And an AI Strategy Blueprint sequencing all of it into phases, with the specific first test named for each item.

The most useful output was the picture that emerged across all of them: a shared, written, honest account of where the organization’s time actually goes, which had never existed in one place before.

About this audit

I ran this audit from the inside. I'm Director of Operations at Habitat for Humanity of Michigan, and it was the first full AI Strategy Audit I ran, before taking the same process to other organizations. Staff are identified by role, never by name, and client details and case specifics are left out.