The RFI Is Due Friday: Where AI Actually Helps an EDO (and Where It Makes Things Up)

The RFI lands Monday morning. It's due Friday. It asks for labor shed data, utility capacity letters, site control documentation, incentive summaries, and a workforce narrative — and your economic development organization has a staff of three, one of whom is also running the annual dinner.

So somebody opens a chatbot.

I understand the impulse. I use these tools every day, in a practice where the deliverable is exactly this kind of work product. I've also watched them produce, with total confidence, a utility capacity figure that did not exist. Both of those things are true, and the distance between them is the entire subject of this post.

“A number without a source is a liability wearing a tie.”

AI is genuinely useful to an EDO. It is also genuinely dangerous in the specific ways that matter most in site selection, because our industry runs on verified specifics — acreage, megawatts, gallons, deadlines — and a language model's core talent is producing plausible text whether or not the facts underneath it are real. Here is what I've learned from using it in live work: where it earns its keep, where it makes things up, and the rules that keep it on the right side of that line.

Where It Actually Helps

Five jobs where AI consistently saves time without putting credibility at risk, as long as a human owns the output.

The first draft. The blank page is the real time sink in an RFI response. The data exists — in last year's response, in the property fact sheet, in the utility's letter — but assembling it into the format this prospect wants takes hours nobody has. Fed the right source documents, AI turns bullets and prior responses into a structured first draft in minutes. Not a submittable draft. A draft. That distinction is the whole ballgame, and I'll come back to it.

Prospect and company research. Before a call with a company or a consultant, somebody has to build the picture: footprint, recent expansions, leadership changes, WARN notices, news. That used to be an afternoon of tabs. It's now twenty minutes of synthesis — with sources attached that a human still has to spot-check. The research didn't get worse. It got faster, which means it actually gets done.

Meeting notes and follow-through. I record calls and run them through AI to produce summaries, decision logs, and action items. Anyone who has watched a follow-up email die in a notebook knows what this is worth. The meeting is the cheap part; the follow-through is where projects are won, and AI is very good at making sure nothing said on a call evaporates.

The surveillance work. RFP scanning, grant deadlines, competitor community announcements, state program changes — the monitoring a two-person shop knows it should do and never staffs. Automated digests won't win you a project, but they'll stop you from losing one because you found out about it on day eleven of a ten-day window.

Data synthesis. First passes over labor data, job posting intelligence, and cluster information — summarizing what's in a dataset, surfacing the questions worth asking of it. It's a competent junior analyst for the first 60 percent of the work. The last 40 percent — the part where you decide what it means — is still yours.

Where It Makes Things Up

Four places where the confidence of the output has nothing to do with whether it is true.

Site and utility specifics. Ask a model about a site's available power or water capacity and it will answer. Fluently. With numbers. Those numbers will be invented, and they will be formatted exactly like real ones. In most industries, a hallucinated fact is an embarrassment. In an RFI response, it's a credibility event — the consultant on the other side checks one figure, finds it fictional, and discounts everything else in the document. One invented megawatt can do more damage than a missed deadline.

Incentive programs. Models carry a memory of programs as they existed at some point in the past, blended with programs that sound like they should exist. Names get invented. Deadlines go stale. Administering agencies get swapped. Incentive details are among the highest-verification items in our work, and they are precisely where a model's confidence is least earned.

Statistics that trace to nothing. Workforce numbers are the worst offenders — plausible, precise, and unsourced. My rule is simple: if a figure can't be clicked through to its origin, it doesn't go in the document. A number without a source is a liability wearing a tie.

Judgment. AI has never stood next to a substation in February. It has never read a room at a zoning hearing, never watched a mayor's expression change when a project codename leaks, never figured out which of the four finalists already won an RFP before the interviews started. The relational and political layer of site selection — the layer where projects actually live and die — is not a text problem, and it doesn't yield to one.

The Rules That Make It Work

AI drafts. Humans verify. Every factual claim in AI output gets checked against a source or gets cut. No exceptions for deadline pressure, because deadline pressure is exactly when the invented utility letter would have gone out.

Feed it documents, not vibes. The output is only as good as what you give it. Your own fact sheets, prior responses, utility letters, and templates produce work grounded in your community's reality. An empty prompt produces the average of the internet, and your community is not the average of the internet.

Keep confidential work confidential. Project codenames exist in this industry for a reason. Client material, prospect identities, and negotiation positions do not belong in tools you don't control. This is not paranoia; it's the same discipline you already apply to email.

The prospect never talks to the robot. Drafting help is invisible and fine. Judgment, negotiation, and relationships stay human and stay yours. The day a consultant realizes your “team” is a language model with a logo is the day your credibility spends a year in recovery.

The Bottom Line

Treat AI like a brilliant intern: fast, tireless, broadly read, and completely unafraid of being wrong. You would never let that intern send an RFI response unreviewed. You'd be foolish not to let them build the first draft.

That's the actual opportunity for small EDOs. The staffing math in this profession has never worked — too few people, too many deadlines, too much surveillance work. AI doesn't fix that by being smart. It fixes it by making the first 60 percent of everything cheap, so your scarce hours go to verification, relationships, and judgment — the work that was always the job.

The EDOs that win the next decade won't be the ones with the most AI. They'll be the ones that pair fast drafts with slow verification, and never confuse the two.