Build the Infrastructure Before You Deploy

What the Agencies That Didn't Stall Did Differently

Minnesota's Pollution Control Agency didn't start with a tool. They started with a problem: complex air permits took a median of nearly 600 days to process, almost four times the agency's stated goal. In early 2025, the MPCA secured $780,000 from Minnesota's Technology Modernization Fund for AI applied to three specific workflows: emissions modeling data preparation, public comment analysis, and permit risk determination. The dollar amount and the technology are the least interesting parts. Governance requirements were written into the grant application before a single vendor was contacted. The agency had to answer how the tool would be overseen, who would be accountable for its outputs, and what would happen when it was wrong. All before the money was approved.

Governor Walz followed up in 2026 with an executive order directing the MPCA to pilot AI tools, upgrade its permit application tracker for public visibility by April 2026, and submit a joint efficiency report with the Department of Employment and Economic Development by December 2026. The oversight arrived packaged with the technology. That ordering matters.

In the first post, I described what we've been calling the Abundance Paradox: 320 permitting technology tools, most of them working, most with APIs, almost none connected. The reason is always the same. No single entity has the authority to require interoperability. AI will follow the same pattern unless agencies make different decisions before deployment. The agencies that stall are the ones that treated deployment as a technology decision; their tools were usually fine. The tool arrives, no one has been designated to oversee its outputs, no one knows what to do when it's wrong, and it runs for a few months before being quietly q.

Five governance decisions separate the agencies that succeed from the ones that stall. None requires new technology. All require authority. Each one is a locked widget, to use the term from the last post: a governance decision sitting unresolved, invisible to the org chart, waiting for someone to own it.

Start with documented decision rules before deployment. What decisions will AI inform? What decisions will it not inform? Written down, reviewed by legal, and signed by a named official. If the AI flags an application as incomplete, what happens next? Who reviews the flag, and what criteria do they apply? This document makes an AI pilot defensible. Without it, the tool runs on informal consensus, which works until someone disagrees.

Then: a named accountable official for every AI-informed decision. A person who reviewed the AI's recommendation, applied professional judgment, and can explain what they decided and why. This is what makes an administrative record hold up. It also gives staff confidence that they're using the tool rather than being used by it. As the NAEP code of professional ethics states, environmental professionals are personally responsible for the validity of analyses performed under their direction. AI doesn't change that obligation. It makes it more visible.

A single-pilot use case, run end-to-end before scaling. Agencies that succeed don't deploy AI across their workflow at once. They pick one task and run it through a real project. They validate accuracy against the outcome they would have reached without the tool. They document what happened when the AI was wrong. Only then do they expand.

The fourth is easy to defer and expensive to reconstruct later: audit logs that outlast the project. What did the AI recommend? What did the reviewer decide? When? This documentation is legal protection, yes. It's also how an agency learns whether the tool is actually working over time. Agencies that don't build this in from the start tend to lose the data they need to make the case for continued investment during budget season.

And public reporting. If AI is being used in decisions that affect public interest, say so. Where in the process is it used? What oversight applies? A one-paragraph note in the administrative record is enough. The counterintuitive finding from agencies that have been transparent about AI use is that they tend to face fewer legal challenges, not more. Transparency turns AI from a liability into a documented choice.

The federal government has done part of this work. The CEQ Permitting Technology Action Plan directed agencies to begin adopting NEPA data and technology standards.[1] OMB's 2025 memoranda require federal agencies to establish AI governance and procurement frameworks.[2] Federal mandates create conditions, though. They don't create authority inside each agency. That authority has to be built from within, one decision at a time.

If you've read this far, you've seen the whole arc. AI can handle the administrative work that consumes reviewers' time using tools that exist today. The line between where AI helps and where it doesn't is knowable. The right vendor questions protect agencies before they sign. The governance infrastructure that makes it all stick costs less than the tool you're about to buy.

Here is the explicit ask: before your agency deploys anything labeled with AI, designate a human official accountable for AI-informed permitting decisions. Write down the specific decisions AI will inform and the decisions it won't. Run one use case end-to-end before you scale. Those are three decisions. They take an afternoon to make and months to undo if you skip them.

The tools exist. Most of them work. The sequence is always the same: data standards before interoperability, interoperability before automation, governance before all of it. Whether your agency builds the authority structure that lets any of it matter is the question these four posts have been circling.

Council on Environmental Quality. Permitting Technology Action Plan. https://www.whitehouse.gov/ceq/

OMB Memoranda M-25-21 and M-25-22. Office of Management and Budget, April 2025. https://www.whitehouse.gov/omb/information-regulatory-affairs/artificial-intelligence/

What to Ask Before You Buy

Four Questions for Any AI Vendor in Permitting

I've watched this happen more than once: a vendor demos something impressive, a director signs, and six months later, the tool sits half-used because nobody thought through oversight before procurement. In April 2025, the Office of Management and Budget issued two memoranda (M-25-21 and M-25-22) that changed how federal agencies buy AI.[1] Documented risk assessments, human oversight, performance tracking, and vendor accountability are all built into contracts awarded after September 2025. The reason these became mandates is that agencies kept buying tools they couldn't oversee.

State environmental agencies aren't bound by OMB memoranda. The failure mode is the same. Standard IT procurement evaluates functionality, security, and cost. AI procurement adds two questions that standard RFPs don't ask: can you explain the output, and who is accountable when it's wrong?

Canada's federal government built a tool that forces those questions to be asked before procurement starts. The Algorithmic Impact Assessment (AIA) is a mandatory, 65-question, open-source, publicly available risk questionnaire that assigns an impact level and determines the required safeguards before any automated decision system is deployed.[2] It makes vendors document their own risk model before a contract discussion begins. Scholars have drawn an explicit comparison between the AIA process and an EIS: both give the agency and the public a structured opportunity to evaluate a proposed system before deployment. The AIA takes what I described in the last post as a locked widget and forces someone to unlock it on paper before the money moves.

Georgia has moved in the same direction at the state level. Its GS-25-002 guidelines, effective since July 2025, require vendors to conduct and share an Algorithmic Impact Assessment, provide documentation on AI logic and model transparency, and, where feasible, complete a pilot phase before full deployment.[3] Georgia's framework is live. It sits in front of every AI contract in the state.

Most state environmental agencies do not have either Canada's mandatory questionnaire or Georgia's vendor requirements. If your agency doesn't have a formal framework yet, four questions cover most of what you need. They're the minimum, not the ceiling. How a vendor responds will tell you more about how their tool was built than any demo will.

The first: Can you explain how it reached that output? Not "is it explainable in principle." Show me, with a specific output from your system, how it got there. If a vendor says "the model is proprietary," that's an answer worth taking seriously. It means the administrative record for any AI-assisted decision will have a gap that your agency, not the vendor, will have to defend.

Second: Who is accountable when it's wrong? Contracts typically disclaim vendor liability for outputs. That's expected. "The agency is accountable" isn't a process, though. Name the official. Document what review they applied and what criteria governed their decision to accept or reject the AI's recommendation. Without that documentation, you have a liability exposure without a safeguard.

Third: Does it integrate with existing workflows, or does it create a parallel one? Across the 320 tools EPIC has tracked, the pattern is consistent: they work in isolation. An AI tool that requires staff to leave their existing system, run a process in a separate interface, and manually re-enter results is a new source of transcription errors. Ask the vendor to show you the handoff point between their system and yours.

Fourth: What does the vendor do with your data? Permitting data includes sensitive project information, tribal consultation records, and proprietary applicant submissions. Where does it go when it enters the vendor's system? Is it used to train the model? Is it retained after the contract ends? Can you get it back? Agencies that can't answer these questions after signing are in a difficult position when a project applicant or a tribal nation asks.

The litigation dimension makes these questions more than bureaucratic prudence. As I noted in the previous post, Beveridge and Diamond published an analysis in late 2025 warning that project opponents will exploit AI use in permitting if the documentation is thin.[4] Courts haven't established how they'll give deference to AI-assisted agency decisions, and the early case law will likely come from the agencies that documented the least. A well-documented AI process protects the agency. A poorly documented one invites challenge.

Minnesota's Pollution Control Agency offers one model for incorporating these requirements into procurement from the start. The MPCA's $780,000 Technology Modernization Fund grant for AI in permitting required governance justification before funding was approved.[5] The procurement filter became a governance mechanism in its own right. The next post describes what that governance infrastructure looks like in practice.

Before the next vendor demo, send these four questions in advance. A vendor who won't answer them beforehand will give you less useful information during the demo. A vendor who has thought seriously about responsible deployment will have answers ready. The difference between those two responses tells you more than any slide deck.

The final post covers what happens after you buy the right tool: the governance infrastructure that determines whether it sticks.

OMB Memorandum M-25-21, "Accelerating Federal Use of AI through Unified Standards." Office of Management and Budget, April 2025. https://www.whitehouse.gov/omb/information-regulatory-affairs/artificial-intelligence/

Government of Canada. Algorithmic Impact Assessment tool. https://www.canada.ca/en/government/system/digital-government/digital-government-innovations/responsible-use-ai/algorithmic-impact-assessment.html

Georgia Department of Administrative Services. GS-25-002 Artificial Intelligence Procurement Guidelines. Effective July 2025. https://doas.georgia.gov

Beveridge & Diamond. "AI in Environmental Permitting: Anticipating the Next Wave of Legal Challenges." 2025. https://www.bdlaw.com/publications/

Minnesota Pollution Control Agency. Technology Modernization Fund AI in Permitting Grant. 2025. https://www.pca.state.mn.us

Where AI Fits, and Where It Doesn't

Drawing the Line Between Administrative Support and Professional Judgment

At a 2025 PNNL workshop on AI in environmental review, Paces CEO James McWalter demonstrated a tool that compresses two to twelve weeks of environmental report drafting into under twenty minutes for a large power project. An audience member responded directly: "I'm concerned about the commoditization of permits to just become checkboxes." McWalter's stated goal was full automation. The audience member's concern was that full automation guts the purpose of public consultation.

Both of them were right about something. This post is about figuring out where that line sits, and why it matters that agencies draw it before vendors draw it for them.

AI is reliable for tasks with clear structure and correctable outputs. Comment clustering, where natural language processing categorizes thousands of submissions by topic and flags duplicates, lets a reviewer start with a map instead of a pile. Document comparison flags changes between EIS drafts, identifies unresolved citations, and surfaces sections where data points conflict. Completeness checking screens applications against known requirements before a human opens the file. In Cook County, Illinois, Esri's aerial imagery comparison overlays new satellite photos against permit databases to detect unpermitted structures. Detection is automated. The enforcement decision requires a human.

These are all cases where being wrong is correctable before a decision becomes final. The reviewer still reads, still decides, still signs. AI reorganized the work. It didn't replace the judgment.

The cases where AI shouldn't have the last word are fundamentally different. Final permit determinations that weigh public interest against project benefits require a type of reasoning that resists pattern-matching: interpreting ambiguous regulatory language where precedent is contested, weighing community input that is itself part of the analysis rather than data to be processed, producing outputs that anchor an administrative record subject to judicial review. The consequences of being wrong in these situations compound rather than self-correct.

Lauren Schramm, a consultant at ESA, presented a session at NAEP 2026 called "Beyond the Hype: The Dark Side of AI." It was the sharpest talk I attended. Her target was carelessness, not AI. Environmental review, she said, is not a pattern-recognition problem. AI assumes transferable patterns and generalizable logic. NEPA requires site-specific analysis, procedural rigor, and legal defensibility. The mismatch produces what Schramm calls "AI Slop": text that reads as complete but contains no actual analysis. Her example was a generic wetlands paragraph that referenced the importance of wetlands four times in different phrasings and included a mitigation sentence with no quantification, resource type, impact category, or proposed measure. The recursive phrasing was the obvious tell. The empty mitigation sentence was the functional one. Agencies are already receiving documents like this in their review pipelines, she reported, with fabricated citations and incorrect regulatory references.

Schramm offered a triage framework organized around four questions: Is it necessary? Is it proportional? Is it transparent? Is it reviewable? Meaningful AI use (formatting, summarizing large datasets with human review, data exploration, and efficiency on low-risk tasks) passes all four. Frivolous use (drafting full NEPA sections, generic impact analysis, automating conclusions) fails on reviewability and, more often, on proportionality. The decision threshold: if you cannot explain how the output was generated, verify the sources independently, or defend it in a public record, do not use it.

That threshold isn't arbitrary. Schramm grounded it in the NAEP code of professional ethics, which states that environmental professionals will be personally responsible for the validity of all data collected, analyses performed, or plans developed under their direction. The accountability doesn't transfer to the tool. Most practitioners are already bound by professional ethics that give them the decision rule they need. Triage is the operational manifestation of that commitment.

There's a role shift here that I think the field hasn't fully reckoned with. When AI handles the drafting, the practitioner's job changes. You stop producing narrative and start validating assumptions and stress-testing analysis. Agencies and consulting firms that recognize this shift early will train their staff differently from those that don't. The entry-level question I raised in the previous post gets sharper here: the professional judgment that distinguishes a competent environmental reviewer from a competent AI prompt operator develops through years of doing the analytical work that AI is now absorbing. How you build that judgment when the drafting is automated is a question the field hasn't answered yet.

Beveridge and Diamond published a useful framing in late 2025: project opponents will exploit AI use in litigation if it isn't well-documented.[1] Courts have not yet established how they'll give deference to AI-assisted agency decisions. Early adopters who blur the line between administrative support and professional judgment will make it harder for every agency that follows.

For any AI application you're evaluating: AI holds up when the task is pattern recognition on large, structured datasets, and a human reviews the output before it matters. It doesn't hold up when the task requires judgment about competing public interests, cumulative impacts, or contested regulatory language. If you can place your use case clearly on one side of that line, you have a defensible position. If you can't, that's a locked widget: a governance decision that hasn't been made yet, and no amount of software will make it for you.

The next post covers what to ask a vendor before you sign anything.

Beveridge & Diamond. "AI in Environmental Permitting: Anticipating the Next Wave of Legal Challenges." 2025. https://www.bdlaw.com/publications/

The Work Before the Work

What AI Actually Does for Permitting

“Two thousand six hundred public comments on a single federal project. The standard process: four people, four weeks. The first week wasn't reading comments. It was proposing category names before anyone had substantively reviewed a single submission. With AI support, the same task took four hours of data management and another hour or two to run the system. Same output. The weeks vanished.”

That example comes from Jacobs Engineering's AI-Engage platform, presented at a PNNL workshop on AI tools for environmental review. Most conversations about AI in permitting get the bottleneck wrong. The delay is rarely the environmental analysis itself. It includes completeness checks, comment triage, citation cross-referencing, document version tracking, and drafting project descriptions. Administrative work consumes the people who are supposed to do the analysis. By the time a reviewer reaches the substance of a permit application, they've already burned their capacity on paperwork.

I spent a week at the National Association of Environmental Professionals conference in Anchorage, where practitioners gathered to talk about where the field is headed. Presenters from different agencies and firms, in separate sessions on separate days, arrived independently at the same observation. Eric Beightel, an ESA consultant who works with state DOTs on categorical exclusions, put it plainly: there are no standardized systems to track CEs across agencies, workflows vary by office, and definitions that should be shared aren't. A governance failure made worse by the absence of basic workflow tools. Wes Furlong from the Native American Rights Fund, speaking in a completely different room, described the same structural pattern from the tribal consultation side: agencies that lack basic infrastructure to track who needs to be consulted, when, and about what. Between sessions, practitioners put it in more personal terms. They spend their days on logistics that shouldn't require a professional degree. With federal staffing reduced, AI has become a practical necessity for many teams, not a luxury.

The tools that address this aren't experimental. They're deployed.

At Idaho National Laboratory, a developer built a tool called AI Scope Assist in about 1.5 weeks. It drafts project descriptions for categorical exclusion documents. The CX process at INL wasn't slow because the environmental determination was complicated. It was slow because writing the project description was the most time-consuming step, and the task followed predictable patterns. Scope Assist plugs into the existing workflow. No system-switching. No new platform to learn.

In Los Angeles County, Archistar's eCheck platform uses generative AI and computer vision to review building plans for code compliance.[1] The pilot launched in July 2025 as part of the wildfire rebuild effort, after the LA fires turned the permitting backlog into a public emergency. Municipalities using eCheck report a 55% reduction in permit review cycles and 81% fewer resubmissions. The platform checks plans against local codes before a human reviewer opens the file. When the plan passes, it passes faster. When it doesn't, the reviewer starts with a specific list of issues instead of reading the full submission cold.

None of these tools requires professional environmental judgment. Comment triage doesn't involve deciding whether a comment has merit. Completeness checking doesn't involve evaluating the substance of an application. Drafting the project description doesn't involve making a determination. These are administrative tasks that consume analytical professionals' time. AI works here because the tasks are structured, repetitive, and can be corrected before a decision becomes final.

EPIC has tracked 320 permitting technology tools across every stage of the environmental review process.[2] They solve individual tasks well. Most have APIs. Almost none connect to each other. We've started calling this the Abundance Paradox: the tools exist, but the governance decisions that would make them useful don't. Getting from isolated tools to connected systems requires a sequence that most agencies skip: data standards before interoperability, interoperability before automation. That governance problem is the subject of the rest of this series. It starts here, though, with the recognition that the work consuming most reviewers' time isn't the work that requires their expertise.

There's an uncomfortable question here. At the PNNL workshop, someone asked what happens to entry-level staff when AI handles the tasks they used to cut their teeth on. The answer was that they're learning to use AI tools interactively, developing a different kind of fluency with the work. That's a real tradeoff, and it mirrors every previous wave of workflow tools, from word processors to GIS. I'll come back to this in the next post, because the workforce question gets sharper once you see where AI stops being useful.

If you recognize your own workflow in any of this, here's the practical step: identify the single task that consumes the most time before you get to the actual analysis. That's your pilot candidate: the most annoying task, even if it isn't the most impressive to automate. Share this post with whoever controls your software budget. Ask them to read the next three.

The next post gets into the harder question. Knowing where AI saves time is easy. Knowing where to trust it is where agencies need a framework.

Archistar eCheck platform. https://archistar.ai

EPIC. "2026 Permitting Technology Landscape Report." https://policyinnovation.org/insights/the-2026-permitting-technology-landscape-report