Industry Thought

The Fainting Goat Phase Is Over. Now Comes the Hard Part.

The state of AI in emergency management in 2026: from curiosity, to the fainting goat phase, to real pull, and the procurement, evaluation, and governance problems that come with it.

The state of AI in emergency management in 2026: what I'm seeing, hearing, and learning. Notes from a few thousand conversations with practitioners, more than fifty presentations, and more contracts than I ever wanted to read. Nothing here is attributed. The people in these conversations didn't sign up to be quoted, so what follows is the pattern, not the person.

In 2023 and most of 2024, almost nobody in emergency management was using AI. But I was, and we were building. The question in every room was what is this thing. I'd ask who had used AI that day and about a fifth of the hands went up, and then I'd walk them through Netflix and Google Maps and the weather app until most of the room was in. The work that year was awareness: show people what was possible, build fast, let them poke at it. Curiosity, not adoption.

2025 was the year I started calling the fainting goat phase. You've seen the video. You walk toward the goat, it stiffens up, it tips over. That was every agency I approached. Funding got disrupted, national priorities were suddenly in question, and the field had no capacity to take on anything new. Say the words adopt AI and the whole organization locked up and fell over. They weren't being difficult. They needed some semblance of stability before they could consider anything, and the right move was to leave them alone and let them get back up.

2026 is different, and the difference is the subject of this piece. The appetite came back, and it came back as pull rather than push. This year I did the raise-your-hand test in a room of about 150 people and didn't need the walkthrough; two people said they hadn't used AI that day, and they were almost certainly wrong. The people calling me now aren't early adopters who'll tolerate something half-finished because they're curious. They have a problem, they've concluded that waiting it out isn't a plan, and they want to know how to buy. Which is where the new friction lives. The question in the room has moved from what is this thing, to how do I use it, to how do we procure it, and the third question turns out to be the hardest of the three. Procurement, contracting, security review, AI policy, and the vocabulary to evaluate any of it: that is the hard part, and it's where the field is stuck right now.

A note on where I'm standing. I run an AI company in this field, which means I have skin in the game and you should be skeptical of me. I'd rather you be skeptical and informed than trusting and burned. That's the choice the field is facing right now, and most of what I see every week is some version of it.

1. What the modal preparedness office looks like

A county emergency manager opens a demo by saying the office is down to a one-person shop and nobody knows whether the second seat will be filled. Asked what's on the docket, the answer comes in order: get the grant work done, then the training the grant requires, then the actual work work.

That's the office I see most. One or two people. A multi-county coordinator covering three counties, three administrations, three emergency managers, with roughly half a person's time for preparedness. A state preparedness lead back after five years away, finding plans five and six years old and an exercise calendar already locked in for a five-year grant period she didn't write. A single office with seventy plans, most outdated, the first fifteen pages of each repeating the last.

The work plan is written by the grant, not the hazard. The grant deliverable will tell us what we're going to focus on is a sentence I hear in different words every month. The exercises get chosen years in advance to satisfy a funder, the after-action report arrives months after the exercise, and the people who understood the surge dependency have taken new jobs by the time anyone reads it.

Underneath that is a debt the field hasn't paid down. A lot of things were left undone during COVID and never picked back up. And 2026 is the year the federal floor moved. The council the President appointed to review FEMA delivered its final report in May and recommended a model it calls locally executed, state managed, federally supported, which in practice means pushing environmental review, procurement, eligibility decisions, and mitigation dollars down to the states, with most of the big changes waiting on Congress. The preparedness grant cycle itself went to court: new conditions attached to this year's preparedness and counterterrorism grants drew a lawsuit from more than two dozen states, a federal judge had already blocked an earlier round of grant policy changes, and as recently as the end of September a restraining order forced FEMA to pull back award notices and stop payments. Inside the agency, the government's own watchdog found that roughly a fifth of one class of disaster response employees wasn't renewed at the start of the year with no record of anyone assessing the effect, a hiring freeze ran until May, and a new requirement for cabinet-level sign-off slowed grant administration. State agencies responded the way you'd expect: vacancies held open, training and purchasing decisions made in a hurry or not at all. And the federal money that paid for hospital and coalition preparedness is being cut, which forces a choice nobody wants to say out loud: drop exercising and planning, or find a different way to do them.

I've said before that states and locals hold the ultimate responsibility here, and I still think that's right. But for a one-person office whose year is written by the grant, a grant that is simultaneously being restructured, conditioned, litigated, and delayed is not a budget. It's a weather forecast.

Emergency management is invisible unless something goes wrong. The public thinks we're in an ambulance. We're closer to insurance, and the budget shows it. So when a practitioner reaches for AI, it's rarely enthusiasm. It's we're underfunded, do this faster, better, cheaper. That's a real need. It's also not an answer to why the function isn't valued, and AI won't fix that part.

2. The first question is always the same

Before I've shown anything, the first question on most calls is how is this different from the Copilot we already have? That question is the tell. General-purpose AI is already in the building. It's on the government machine because it came with the license. It's our only approved AI, as one program lead put it, busy being angry at it because it couldn't build the pivot table she needed.

The pattern after that is consistent. Someone points the bundled assistant at the document repository, it looks good on the surface, and some weeks later they realize it isn't doing what they thought. It's pulling the old version. It's confidently summarizing a plan it only partly read. Then they go shopping for something purpose-built, usually on a personal device, because the network blocks the vendor's website.

Here is the mechanism, because practitioners deserve to understand it. Every model carries a bucket. The system instructions go in the bucket. Your documents go in the bucket. Every turn of the conversation goes in the bucket. The model runs laps between you and a server, and as the bucket fills, water spills out. It never tells you what fell out. Hand it a thousand-page hazard mitigation plan, which is about average, and it says I've got you while it has already dropped most of it. When the session gets long it offers to save a memory, which is a summary of a conversation it already lost the details of. Most people think the thousand-page plan is in there for good. It isn't.

And it won't say I don't know. By design, it's been rewarded in training to answer. So it flatters. Anyone who has used a consumer chatbot has been told they're the smartest person in the room. That's charming on an apple pie recipe and dangerous on an evacuation annex.

The lesson I've drawn: access is settled. The question is no longer whether practitioners will use AI. It's whether what they use is grounded in anything. Bans don't change the first fact; they only move the use to personal phones, along with whatever sensitive content goes with it. The policy that assumes use and governs it will beat the policy that pretends otherwise.

3. The real game isn't EM-shaped content

Strip the marketing off most AI products in this space and you find a model pointed at an emergency management template. Plan-shaped output. Scenario-shaped output. Inject-shaped, AAR-shaped. It looks like the field. That's the surface game, it's mostly solved, and on its own it's close to worthless.

The underlying game is four things, and they're harder than the surface game by an order of magnitude. Complete context: the whole of what the organization knows, held without the bucket spilling, not whatever fit in the last session. Retrieval: getting the right thing back, which is a relevance problem before it's an accuracy problem; you can accurately retrieve something that has nothing to do with the question. Evaluation against a fixed corpus of rules and behaviors: doctrine, regulation, the organization's own commitments, and the patterns it has actually shown in exercise after exercise, so that an output can be checked against something that doesn't move. And surfacing when the language model is wrong or ungrounded. Ungrounded doesn't mean useless. In practice, the made-up thing might be exactly the counterfactual you needed. But you have to know it was made up, and you have to be the one who decides what to do with it.

Here's why that separation matters. Language models are good at language, framing, synthesis, and opinionated judgment. They are not good at keeping the truth straight. Ask one to draft a hurricane scenario for a hospital and it will sound exactly right: the cadence of incident command, the right acronyms in the right places. Then ask whether the population figure matches the county, which CMS regulations bind that facility, and what happens to fuel reserves at hour 72 of generator load. It will answer all three confidently. Two of those answers will be wrong, and you won't be able to tell which two.

So the model writes the story, and something that isn't a model has to own the truth and check the story against it. We built the company around that split: a deterministic engine of rules and hazard data underneath, language models on top, and an explicit flag whenever the story leaves the ground. Most vendors in this space didn't build that, and you can tell really fast. A chatbot writing stories feels one way. A system that handles context structurally feels another. The only similarity between us and some of the products I get compared to is that we both write words using AI.

The practical test for a buyer is one question. What happens when the model is wrong? If the answer is some version of it isn't, you've learned what you needed to know.

4. Where the time actually goes

A state champion finishes IT review through our public trust center and budgets three years at list price. The next email is from the contracting office, promising a year-long process. The three-year budget will now buy one year, a year from now.

Procurement is where AI hope goes to die. My life has become contracts and insurance. The timeline runs from weeks to eighteen months. I've watched a county take three months to land a four-figure subscription because its contracting office required its own commercial agreement. I've watched a large city spend three months approving a purchase that cost about as much as a streaming service. One state has been ready to buy for two years. Another killed a deal on a single sentence claiming the state owns everything the product generates, and afterward asked me what had killed it. One deal died after fifteen months because the risk manager quit, that signature was required for approval, and the position sat vacant. Fifteen months of work, undone by an org chart.

The clauses repeat. Uncapped liability and indemnification. Five years of record retention after termination. Multiple governments have required my AI to carry automobile insurance. Three-hundred-question security questionnaires. Terms of service that must be frozen, when software terms are living documents, so every frozen copy becomes a branch to maintain: one agency wants data deleted five days after cancellation, another is fine with a year. A request for proposals worth eight figures required six people in seats, which is to say a headcount contract that cannot buy software at all. Nobody has figured out how to bill for an agent in six-minute increments.

None of this is malice. Government contract templates were written for labor, not software, and long before anything called AI. They're written for the buyer, not bilaterally, and that made sense when the vendor was a staffing firm. Emergency managers don't come from legal or procurement, yet contracts, mutual aid agreements, and vendor terms are what actually get you the stuff in a disaster. That's a gap worth exercising, by the way.

Some encouraging things. Security review is now universal and nearly identical from the smallest county to the largest enterprise: models, data flows, retention, hosting. Transparency clears it. A public trust center with requestable architecture documents does more than any sales deck. A few states have built formal AI safety reviews, and passing one is becoming a portable credential: one state's committee reviewed us unprompted for four months and then authorized any state agency to adopt without further AI review. That's governance working. Meanwhile other states have a review requirement and no submission process. You have to find someone to do it.

What I've concluded: the binding constraint on AI in emergency management in 2026 is not model capability and not practitioner appetite. It's contract language. Roughly 70,000 jurisdictions each have their own secret thing, and the vehicles are being rewritten mid-deal. The fix is boring: standard clauses for AI software, state-level review gates with a front door, and vendors willing to publish enough that review has something to review.

5. What people ask me about accuracy, and what I tell them

On a demo, a prospect asks what's its accuracy and waits for one number. I say around ninety percent, depending on what you're asking it to do. Wait, it's not just one number? And I watch some of the trust drain out of the room while I explain the trick.

Accuracy isn't a number. It's like saying Justin is accurate. It moves with the task, the corpus, the question, and how tired the model's context is. A system can be ninety-five percent accurate at retrieving the right section of a plan and seventy percent accurate at summarizing what the section says. Borrow the intelligence community's language instead: confidence, never truth. This is thirty percent likely to be true. Make your choice.

The ceiling is real. No lab, with billions of dollars, has pushed a language model past the low nineties on the benchmarks it publishes, and hallucination is a mathematical property of the tool, not an engineering bug. The best I can do is say when the system is making something up. So when I mean ninety-three percent on a specific measurement, I say ninety-three percent. When a vendor leaves the percentage out, they are telling you it's a hundred. Trust me, bro.

A 95 percent accurate AI in marketing is survivable. A 95 percent accurate AI in public health preparedness or emergency management carries 5 percent of possible harm, and that 5 percent is not survivable. But it's not one bar. A 90 percent accurate AI drafting an after-action report is a productivity win; errors get caught in review. The same model generating a wireless emergency alert during an incident is a liability event waiting to happen. One alert in ten wrong, pushed to hundreds of thousands of phones, with no recall. Emergency managers already think this way about every other piece of equipment: a HazMat detector that's 90 percent accurate is fine for perimeter screening and dangerous for clearance. AI is no different. The accuracy bar has to match the consequence of the decision it's informing, which means the field doesn't need a benchmark. It needs a benchmark family.

Alerts are where this argument gets tested, so let me take the strongest version of the other side. Some will say AI-generated alerts only need to be better than the typically bad ones humans write, and on the surface that's true. I've read the typical human alert: acronyms nobody outside the agency knows, no clear protective action, a reading level that loses half the audience. The research backs that up. A review of 6,080 wireless emergency alerts sent between 2012 and 2022 found that only 8.5 percent included all five things a warning needs: the hazard, the location, the source, protective action guidance, and the time. Source and timing were the most often missing (Olson, Sutton, Cain, and Waugh, 2024). A model fixes most of that on the first pass.

The comparison hides three things. First, it compares rates, and harm is a count. Harm is volume times error rate times what each error costs. AI lowers the cost of producing an alert, so volume goes up: more platforms, more languages, more updates, finer zones. Say a human writes twenty messages across an incident and gets six wrong, and a novice with a generator writes two hundred and gets twenty wrong. The rate improved and the N of harm went up. Those numbers are hypothetical, because nobody has published a real one. Second, the errors stop being independent. A human's mistake is one message on one channel. A generator's mistake is one wrong zone pushed to every platform in both languages at once. The error is cheaper and it travels further, and the human review that would catch it doesn't scale with the volume, so the effective error rate drifts toward the raw model rate as output rises. Third, better than humans at what? Human alerts fail mostly on completeness, and that failure is well measured. Models fail on facts: wrong zone, wrong time, wrong contact, a website address with an extra letter in it, and nobody has published how often that happens, for humans or machines. The improvement is on the axis we've measured. The new risk is on the axis we haven't. A single comparison hides the trade. None of that counts alert fatigue, where volume itself erodes the public's response to the next message, correct or not.

I'm not saying don't use AI for alerts. We built an evidence-based alert template tool with researchers and risk communicators and gave it away. I'm saying evaluate that use case on its own terms. That means a benchmark for alerting, stratified by failure mode, that measures total harm across an incident rather than a per-message rate, against a human baseline that someone has actually measured, because right now nobody has measured factual error rates on either side. And it means friction between generation and transmission that grows with the volume: a hold, an approval, a named human. It's also worth noticing where the measured improvement in alert quality has come from: structure. In controlled experiments, messages written to a standardized template outperformed typical messages on understanding, belief, and confidence about what to do, before any model was involved (Sutton and colleagues, 2025). Pre-scripted templates built in advance with experts reviewing are one thing. Live generation at incident tempo is another, and the second hasn't earned its place yet.

Any vendor can claim accuracy. Any vendor can claim safety. Any vendor can make up great-sounding terms, like closed circuit AI, that mean little and differentiate less. Drill into closed circuit and it usually means not connected to the internet, while the data still flows to a model provider. I use fourteen models and I say so. I've read terms of service in this space that still permit training on customer data, and I've watched a competitor change its terms in the middle of a trial without showing what changed. I've watched an emergency operations assistant agree to role-play a pirate and invent answers within seconds. Vendors who sell safe, accurate, closed circuit with no qualifier aren't earning trust. They're renting it, usually from a relationship.

There's a trap here for anyone honest, and I've fallen into it. The more I explain where the AI is wrong, the less some buyers trust me. There is research on this: the person who explains the magic trick loses credibility even when he built the trick. More than once I've thought I should have just shut up. I'm simultaneously selling AI to these agencies and telling them the most dangerous thing they're about to do is believe it's perfect. I don't want to rent their trust. I want an informed purchaser. That costs me deals, and I'd make the same call again.

6. The pull toward response

Mid-demo, a hospital preparedness coordinator cuts in: "You keep saying practice, practice, practice. From what I see, this is very accurate on our real data. Why couldn't this run during a real activation?" I tell her it's a philosophical line. She shrugs. All she wants is for the system to whisper that could be wrong when someone on the call says something the record contradicts.

The pull toward response is universal. Partners tell me there is infinitely more interest in anything related to live response than to practice. Search interest in testing is falling while training rises, and customers ask for just-in-time reps. Investors and advisors tell me I have a response platform and don't know it. Academics ask whether the data will include real events or only exercises. A response demo is thrilling. A preparedness demo is a person explaining their planning process. The money and the visibility are on the response side, which is exactly the side where a probabilistic tool is least appropriate.

My position hasn't changed, but I want to state it precisely, because the imprecise version gets quoted back to me. The design goal is to create the most space possible between a language model generating something and a human being harmed. Preparedness gives you weeks, months, or years to catch the wrong answer. Response gives you minutes. A language model under duress is all possible answers to all possible questions at all possible times, and the moment a person most needs to scrutinize the output is the moment they're least likely to. I think that's a human problem more than an AI problem: under pressure, people franchise out their decision-making. I won't build a tool that loads a gun and points it at an emergency manager.

So: AI hasn't earned response yet. The word yet is doing honest work in that sentence. There is real value in information work around a response, synthesis and drafting and tracking and retrieval, and that is different from a tool near a time-compressed decision; we should stop letting the first launder credibility for the second. Guided help, where the playbook is and who to call, is different from a recommendation. Just-in-time rehearsal that lets a team test the difference between two decisions is different from telling them which to make. Business continuity, where nobody is dying, is a different consequence space than an evacuation. Deterministic rules engines and physics models, which have math behind them, are a different class of tool than a chatbot. If an independent evaluation ever shows a system performing better than the human alternative on a specific response task, with the mechanisms to prove it, entering that task would still be principled. Until then, the line holds, and I'm against my own product's expansion in that direction more often than my advisors would like.

7. The AAR button

Users ask for it on nearly every call. One after-action report, please, from the exercise we just ran. I keep saying no, and I owe people a better explanation than principle.

Three reasons. First, the easy button is the problem. An AI-written report lets an organization certify that learning happened without any learning happening, faster than before. Second, a vendor-written final report is discoverable. When something goes wrong after a disaster, I don't want the authoritative record of your readiness to be a document a machine wrote and nobody argued with. Third, the public corpus of after-action reports is poor evidence. A national review of thousands of AAR-shaped documents found almost none usable as evidence of anything; peer-reviewed work has found most of them generic and many missing whole capabilities. Public AARs contain the three things that were safe to say. Point an AI at that library and you get the safe findings, faster.

What we do instead is the verb. The system captures the exercise at full fidelity: who said what, in which breakout, against which objective. It evaluates by conversation, with the facilitator defining what good looks like and approving or rejecting every candidate finding. And it does one thing I've come to think is the most valuable feature in the product: it compares the recorded exercise to the sanitized report that followed it and asks, why does this authoritative document say nothing like what was said in the room? We've sanitized so much along the way that it's fair to ask what we're improving from. The data that matters is the conversation that never made it into the document.

Two more things I've learned from exercise people this year. A veteran designer told me clients say they hate the evaluation paperwork, and what they mean is they don't want to plan the exercise carefully against a risk assessment and their plans. He compared it to sandpapering before you paint. And a learning specialist designing a competency assessment kept repeating that situational awareness and decision rationale are not quantitative unless you assign a scale and an evaluator who knows why the participant decided what they decided. Both are right. A tabletop tests a plan. It doesn't test time to decision or individual competency, and most cyber tabletops I see are drills. Europe is well ahead of us on actual simulation. The United States is still write a document and talk at each other.

8. What's working

A state preparedness lead listens to a demo and reframes it for me: so you're not an AI company. Right. I'm a context-engineering company with AI on top. The AI is not the problem anymore. The problem is where all the information is, and it isn't where we need it.

Here's what I see working, across the small number of organizations that are getting this right.

They define the problem first. We want to get into AI isn't a problem. We have seventy plans and can't tell which assumptions still hold is. When the problem is named, the tool follows. When it isn't, I tell people to use a spreadsheet. I now decline most inbound demos, because the organization either hasn't recognized a problem or isn't allowed to use AI, and a demo was always a vanity metric.

They treat plans as documented assumptions. A plan is a set of claims about reality frozen at the moment of signature, and reality drifts. The useful work is finding the delta between what the doctrine says, what the plan says, and what the organization actually does. Doctrine says seventy-two hours of fuel. Three exercises in a row say a hundred and twenty. I want the delta, and I want it surfaced continuously, not rediscovered annually.

They gather context at scale before they design. You could never get a thousand people into a planning meeting. You can interview them with an agent, collect their documents, and look for dependencies, consensus, and contradictions before the first draft exists. Then run a half-hour staff-meeting exercise against the gaps, and let the discussion drive the plan revision. Analysis, then practice, then findings, then revision. The old order was committee, then document, then shelf.

They understand the bar depends on the stake. The same ninety percent that's a productivity win on a draft is a liability event on a public alert. They ask for a confidence, not a verdict.

They push the button. The first user in every new account gets a note from me that says, more or less, you are not operating a nuclear plant, load a document, you'll be fine. People who have survived a full procurement cycle then stall at the first screen, afraid to break something. Reluctance to be wrong inside software is universal, but it runs deeper in a profession where being wrong has consequences. Adoption isn't a procurement problem. It's a first-click problem.

And a few quiet signals. A financial-services enterprise cleared our product faster than most governments. States are competitive in a specific way: nobody wants to be first, nobody wants to be last, everyone wants to be second, which means one honest public write-up moves ten agencies. And honesty about limits is itself a differentiator, because almost nobody else is offering it.

9. The market, and the room

Two hundred people, senior leaders on stage, and the words trust, human in the loop, and performance standards repeated all day without anyone defining them. That was my experience at the AIDE Summit this year, and I want to be careful about how I say this. The people behind that effort are doing essential work. They mapped the market when nobody else had, they are leading the charge for responsible adoption, and the field is better for it. I think they could have gone further, is all. I had hoped for much deeper dives, and instead much of the conversation circled the same undefined notions that get repeated in every room I'm in. By the afternoon I was frustrated.

So when I got on stage, I decided to speak as honestly as I could, and to accept that I may never be invited to speak again.

The best count we have says there are about 1,179 AI-enabled products for emergency management from 717 companies. Most are general-purpose tools pointed at the field. One in five depends on another product to function. Coverage is heaviest in response and thinnest in preparedness, which mirrors where the funding has always gone. And multi-organization coordination, the defining feature of how emergency management actually operates, is the thing the market has built the least support for. Lock-in is a better business model than integration. A thousand rational commercial decisions made the field harder to coordinate.

2026 is also the year the field got crowded. More entrants are showing up in AI software for emergency management every month, and I think that's great. I advise some of them, and I encourage others to get in, because the problems here are bigger than any one company. The risk that comes with a crowd is slop: thin products with confident marketing that pull attention and budget away from the work that matters, and make it harder for a buyer to tell the difference. Which is why we need evaluation standards now more than ever, so customers can compare tools fairly instead of on charisma. I'll admit the other side of it too. Slop is a great differentiator. When most of what's on the market is a wrapper and a slogan, a product that shows its work doesn't have to shout.

I'd rather say this before than after: parts of this field are going to get burned by what they're adopting right now. Not because the technology is all bad (though some really is), and not because anyone acted in bad faith. Because we are making high-consequence purchases on charisma, and some fraction of those decisions will fail in public at the worst possible moment. When that happens, the damage won't stay with the jurisdiction that bought the tool. It will land on the whole idea of AI in emergency management and set the careful work back by years.

We vendors were asked to agree on performance standards so agencies could compare products. I said what I think: you've pointed at five competing companies to set standards for processes the field hasn't proven in twenty years. We don't have a shared vocabulary for accuracy, relevance, or trust. Accuracy in a process that hasn't been shown to achieve anything doesn't matter. If what we're building is for increasing the production of paper, we're useless. AI adoption is not the problem. What are you trying to solve?

Two other things I'm watching. Practitioners can now build software without an engineering team, and some are shipping emergency management tools that way. You can build really fast and go to market. The question is whether you should. We're vibing our way to more theater, and the field sees AI as one thing, so one headline about one bad tool will undo years of trust-building for everyone. I've spent three years getting this field comfortable with AI, and I'm aware that the next setback won't care whose product caused it.

10. Governance is being written in real time, without us

A national disaster-relief organization sends over its new AI policy as a condition of a free pilot. It reads like a prohibition. It was written without talking to anyone who builds this technology or anyone who uses it in the field. After a redline, they scrap it and start over. I've seen the same document, more or less, at a large technology company and at several agencies.

There is no AI governance in emergency management comparable to what healthcare has. Most existing policies are poor, and the honest version of the problem is that we don't yet know what we're governing. IT's incentive is to take on no risk it can't mitigate, and the easiest way to do that is to say no. So staff read the policy as a ban, keep using the tool on their phones, and a procurement question I now ask early is simply whether the organization has a policy that prohibits AI, because the answer decides whether we're having a real conversation.

Liability is the issue underneath, and it's the one nobody wants to know about. A veteran editor in this field told me it's the biggest problem precisely because nobody wants to find out the answer. A former government communications official described the scenario to me plainly: one catastrophic event, an emergency manager who says I used this platform, and a legislative hearing that subpoenas every chat log. Lawyers inside the government tell me the laws aren't adaptable enough for the technology. Everyone is waiting for the first case.

Meanwhile a federal team that quietly dropped a conversation with me two years ago, because they weren't allowed to use AI products, came back this year to say the restrictions have been lifted by a new memo. Policy moves in jumps. The emergency management function is almost never in the room when those jumps get written, and it is the function that will inherit the consequences. Participation, not abstinence.

11. What I'm learning about my own position

I wear a shirt that says it's the AI's fault. It's intentionally ironic, because you can never say the AI did it. Accountability can't be delegated to a tool.

The thing I keep relearning is that this is harder than selling technology. I understand selling to government; it's the AI part the field isn't ready for. Showing most emergency managers an agentic workflow is like showing a caveman a lighter. The field is about two years behind what's possible, and most of it still thinks a language model is a word generator or a fancy search engine. That's not an insult. It's the state of the evidence and the state of the training. A workforce survey of what practitioners actually know, and what they do with these tools, is only now being drafted. We're building the field's capacity to adopt AI into a discipline that hasn't finished defining itself.

I've also learned to say no more. We don't build process when it isn't earned. I don't build a feature because a prospect's logo wants it. I've declined to put the product in rooms that weren't ready, and I've declined revenue tied to headcount, because the whole point is a change of labor, not a faster typewriter.

And I've learned that the sequence matters more than the tool. Data first: get the organization's own records out of the filing cabinet and the shared drive and into a form a machine can hold without losing it. Decision support second. Anything that deserves the word intelligence third. The organizations that get the sequence wrong buy AI they can't use, and the output hides the gaps instead of revealing them.

12. What I think it adds up to

  • The modal preparedness office is one or two people whose year is written by the grant, running on pre-COVID plans, with after-action reports that arrive months late and are read by nobody.
  • Shadow use is the default. General-purpose assistants are everywhere because they came with the license, and bans move the use to personal devices.
  • The binding constraint on adoption in 2026 is contract language and procurement, not model capability or practitioner appetite.
  • Security review has become universal and uniform. Transparency clears it. Secrecy doesn't.
  • The field has no shared vocabulary for accuracy, relevance, or trust, so unqualified marketing words are heard as a hundred percent.
  • The real game isn't EM-shaped content generation. It's complete context, retrieval, evaluation against a fixed corpus of rules and behaviors, and surfacing when the model is wrong. Most vendors are playing the surface game.
  • Silent context loss on long documents is the most common failure mode, and almost nobody recognizes it when it happens.
  • The pull toward response comes from money and visibility, not from evidence that it's safe there.
  • The after-action corpus is sanitized receipts. AI pointed at it reproduces the safe findings faster. The valuable data is the conversation that never made it into the document.
  • Plans should be treated as documented assumptions, and the job is finding the delta continuously.
  • Scale is the new harm vector: a ninety-percent-reliable generator in a novice's hands multiplies errors, not just output.
  • Honesty about limits is a differentiator because almost nobody offers it, and it costs the vendor credibility in the demo.
  • Governance is being written state by state, sometimes with no front door, and the emergency management function is rarely in the room.
  • The evidence gap is double. No one has proven the underlying practices produce outcomes, and no one has measured what the workforce does with AI.
  • The field treats AI as a monolith, so one badly built tool is a shared risk for everyone.

The bottom line

Start with the problem. See if technology helps. Check whether you have the governance to do it safely. Find out whether the tool actually needs AI in it. Then decide if the risk is worth it. AI is not magic; it's a utility with a bill that scales with use. If pen and paper fixes it, you don't need it. If you don't understand the problem, stop.

I sell AI, and I'm telling you not to buy it unless it's necessary to do the work. In a field with no standard, where the cost of being wrong is measured in people, that's the only honest position available. Not everyone around you is going through this process, and eventually, even if you opt out entirely, you'll be an island surrounded by others using it.

The question isn't whether AI arrives in emergency management. It already has. The question is whether we'll know what we handed it, and whether it worked. Better compared to what? Let's go find out.