UM Ulrich Monthe
AI Product Consultant · Civic & Social Sector

Most organizations should build less AI than they think. I help you find the part worth building.

I work with foundations, nonprofits, and civic-tech organizations under pressure to "do something about AI" — setting direction with an explicit build / don’t-build decision, shipping pilots safe enough for the communities you serve, and building teams that run them without me. Currently Product Director at the Black Wealth Data Center, a Bloomberg Philanthropies company.

Book a strategy sprint call See the proof →
594,053
IRS filings parsed into a working tool, not a slide
118
foundation pages published, each citing its source line
4
defects my own checks caught — written up in full

Real verdicts from build / no-build reviews

Don't build A chatbot answering questions people already ask general AI
Build Citation-locked answers over your own validated data
Don't build An "AI strategy" deck that never says when to stop
Build A check that blocks made-up answers before launch
Buy, don't build General-purpose AI assistants for staff workflows
Build A system that says “I don’t know” rather than guessing
Don't build A public model with no answer for "what if it's wrong?"
Services

Three stages. Any of them can stop the work.

Start small. The sprint gives you a defensible answer in weeks — and often that answer is "don't build." If it is "build," the pilot ships under evaluation that can fail it. If the pilot holds, I embed until your team runs it without me. Knowing when to stop is the product — anyone can write a plan.

01 · Start here

AI Strategy Sprint

2–4 weeks · fixed fee
Wk 1 workflow interviews, not workshops
Wk 2 scoring every idea against one test
Wk 3 verdicts, roadmap, board memo

A structured answer to "what should we do about AI?" that you can defend to your board and your funders — including the parts where the answer is no.

We stop if

Your users can already get the answer from a general AI tool. Most ideas do not survive this. That is the point of asking before you spend a budget.

You walk away with
  • A scored list of every idea — the verdict, the reason, and what would change it
  • For each idea that survives: what you expect to happen, and the result that would prove you wrong
  • A board memo that holds up when a trustee pushes back

See how I score and refuse →Funder Standing publishes the conditions it requires, and shows which foundations fail them.

For: leadership teams facing an AI mandate they didn't ask for.

→ Feeds stage 02: the top-ranked survivor becomes the pilot's scope.
02 · Deliver

Responsible-AI Pilot

8–12 weeks · scoped engagement
Wk 1–2 the tests it must pass, written before any code
Wk 3–8 build against it
Wk 9–12 someone tries hard to break it, then the release decision

A working generative-AI pilot with guardrails from day one — because your users are people who cannot afford a fabricated answer. The evaluation comes first and has the authority to stop the release.

Nothing ships unless

There is a set of tests it has to pass, written before the code and able to block the release. If we cannot describe what a wrong answer looks like, we do not launch — we go back and work that out first.

You walk away with
  • Every answer shows its source, and you can click through to the original record
  • The system says “I don’t have enough to answer that” instead of guessing — designed in from the start, not added after a bad incident
  • Tests that can stop a release, rather than a review written once it is already live

Read what my own checks caught →Four defects in a live product, three invisible in every summary statistic — and the check still outstanding.

For: orgs whose sprint said "build" — or who have a stalled AI project that needs governance.

→ Feeds stage 03: the habit of testing before release is what your team keeps.
03 · Embed

Fractional AI Product Lead

1–2 days/week · monthly retainer
Mo 1 watching how your team actually works, role by role
Mo 2–3 your staff run the tests, I review
Ongoing a written review cadence, and my exit

Senior AI product leadership at a fraction of a full-time hire — for organizations that cannot recruit this role at sector salaries. The measure of success is how quickly I become unnecessary.

This ends when

Your team ships without me. I write the exit criteria into the engagement at the start. If a standing dependency is what you want, this is the wrong arrangement.

You walk away with
  • An AI roadmap that survives contact with your teams
  • Staff who run the tests and the releases without me in the room
  • A written review cadence your board can hold you to

See what gets transferred →The policy generator shows its work: every clause, and the rule that put it there.

For: foundations and nonprofits who need the capability, not a dependency.

How I work

Four rules, learned the hard way.

Differentiation decides it.

If users can already get the answer from a general AI tool, we don't build it. I've halted funded builds on exactly that test — saying no is the strategy.

Guardrails before features.

Source citation, refusing rather than guessing, and an automatic check before release all ship in the first version — not as a follow-up. In this sector, a wrong answer harms real people.

Adoption is the product.

A tool nobody uses is shelfware. I watch how people actually work before designing anything, train them by role, and leave champions behind so the system outlives my involvement.

Leave capability, not dependence.

Success is your organization running its AI systems independently — with the governance, skills, and momentum to keep evolving them after I'm gone.

Proof

Each service exists because I've done it inside a real organization.

BWDC · Bloomberg Philanthropies
↳ maps to 01 · Sprint

Setting AI strategy under a funder mandate — including what not to build

Bloomberg Philanthropies' Greenwood Initiative asked a hard question: what should this organization build with AI — and what shouldn't it? I led the sprint that answered it, framing direction as three testable hypotheses with differentiation as an explicit build / don’t-build decision. Some funded ideas didn't survive the gate. That was the point.

BWDC · Generative AI
↳ maps to 02 · Pilot

A governed generative-AI pilot for a vulnerable-population context

Shipped a generative-AI system under the strictest conditions I could design: every answer had to cite its source (Census ACS, SBO, Federal Reserve) or it was not returned at all; the system refused to answer outside what it could verify, rather than guessing; and a second model scored every answer before release, blocking any that failed — what the field calls LLM-as-judge evaluation. Release was closed and controlled. No citation, no answer. Restraint was the deliverable — in this context, shipping fast and wrong isn't an option.

BWDC · Explore Data
↳ maps to 01 + 02

Rebuilding a public data platform with the community it serves

Led the end-to-end rebuild of Explore Data grounded in participatory research with 55 community participants — then shipped NLP search, a guided query builder, AI-generated chart summaries, and first-ever mobile access. The platform reached 360K+ unique users in 14 months, peaking at 71K in a single month.

The part most consultants would hide: the data also showed a huge audience engaging shallowly and a small core going deep. Instead of chasing consumer retention, I followed the power users — repositioning the platform toward B2B enterprise intelligence. I instrument what I build, and I act on what it says, even when it argues against my own thesis.
BWDC · Internal Systems
↳ maps to 03 · Fractional

Organization-wide AI adoption that runs without me

Led a six-month implementation of an AI-enabled knowledge and workflow system — Notion AI, custom agents, automation — by watching real workflows, training people by role, and earning buy-in against real resistance. Thirty staff across five teams now operate it independently. The real test of adoption is what survives after the lead steps away.

Live now: EIN Due-Diligence Checker · Nonprofit AI Policy Generator — small, real, useful civic tools, built fast and shipped in public.
The hardest part of AI isn't the model — it's people, process, and trust. The organizations that get this right treat judgment as the scarce resource, not technology.
— Ulrich Monthe
About

A decade in the social sector. Not a tourist.

I'm a product and AI leader based in Washington, D.C. For over ten years I've built technology exclusively for organizations working on economic mobility, racial wealth equity, and global development — I know how funder mandates, board dynamics, and mission constraints actually shape what you can ship.

My work pairs hands-on generative-AI delivery — RAG, LLM evaluation, agentic workflows, responsible-AI guardrails — with the change management that makes new systems stick. I'm also building learnSignal in public: a scenario-based judgment-training platform for AI product managers, my working lab for prompt engineering, evals, and RAG under real conditions.

Fluent in English and French.

  • Black Wealth Data CenterProduct Director · Bloomberg Philanthropies
  • DAI Sustainable Business GroupDirector of Global Product
  • GlobalGivingScaled Atlas to 10M orgs, 180 countries
  • World Bank · IFCFirst gender assessment tool, 15 value chains
  • CatchafireNonprofit marketplace product
Work with me

Under pressure to "do AI"? Start with what not to build.

A 30-minute call is enough to tell you whether a strategy sprint would pay for itself. If it wouldn't, I'll say so — that's the whole method.

Or just book the call

Forty-five minutes, no pitch. Bring the situation you're actually stuck on. If a sprint wouldn't pay for itself, I'll tell you on the call.

Book 45 minutes →
Prefer email? umonthe1@gmail.com ·
Connect on LinkedIn →