Siobhan Hargrove runs the public engagement office at a city transportation authority, a staff of seven, a $2.4 billion capital program, and a mandate to develop an AI governance policy before the authority could deploy the automated bus scheduling tool that operations had been waiting on for eight months. The City Council had passed a resolution requiring public consultation on any AI system that would affect public transit service. Siobhan had run public comment processes before: the usual approach was a 30-day comment window on the city website, a public hearing on a Tuesday evening in City Hall, and a summary of comments that the agency published before making its decision. She knew, from years of experience, that this process consistently reached the same population: organized advocacy groups, business interests, and a small number of highly engaged individual residents. It rarely reached transit-dependent riders, non-English speakers, or people who work night shifts and cannot attend Tuesday evening hearings. She decided to design the consultation for the people who were missing, not for the people who always showed up. It took more time. The result was better policy.

That decision is the whole lesson in miniature. A consultation process is a design, and every design has a default population it serves. The 30-day website window and the Tuesday evening hearing are not neutral instruments; they are instruments tuned to residents with reliable internet access, flexible evenings, and the vocabulary to write a comment that agency staff will treat as substantive. Nothing about them is improper. They simply produce a public record that answers the question "what do the people who already participate think" when the question the authority actually needed answered was "what will this system do to the people who depend on the service".

Why Public Consultation on AI Is Different

Public consultation on AI policy is not the same as public consultation on a zoning variance or a transit fare increase. The topic is technically complex, the terminology is inaccessible to most residents, and the potential effects, on employment, on service equity, on privacy, on the fairness of government decisions, are diffuse and difficult to describe in plain terms. A fare increase is a number a rider can react to immediately. A scheduling algorithm is a set of design choices whose consequences arrive months later, unevenly distributed, and attributed by most people to the agency rather than to the tool.

This creates two distinct risks. The first is that consultation processes attract only technically sophisticated participants, which means the public record reflects the concerns of a narrow slice of the affected population. The second is that consultation processes generate responses that are primarily rhetorical rather than substantive: participants voicing general support or general opposition without being able to engage with the specific design and governance choices that actually determine whether an AI system will be fair, accurate, and accountable.

Effective public consultation on AI policy addresses both risks by doing two things simultaneously: making the technical content accessible to non-experts, and actively reaching the populations who are most likely to be affected by AI decisions but least likely to participate in standard consultation processes. Neither alone is enough. Accessible materials distributed only through the usual channels reach the usual people in plainer language. Aggressive outreach to underrepresented communities, paired with a technical document nobody outside the agency can read, produces attendance without input.

Designing for the People Who Are Missing

Siobhan's method starts with an inventory rather than a schedule. Before choosing any channel, she asks who the AI system will affect and how, then asks which of those groups reliably appears in the authority's existing public record. The gap between the two lists is the consultation design brief. For the bus scheduling tool the affected population was transit-dependent riders on lower-frequency routes, shift workers whose trips fall outside peak service, riders with disabilities who depend on predictable timing, and the operators whose assignments the tool would reshape. The existing record contained almost none of them.

Accessibility belongs in the same inventory rather than in a compliance annex at the end. Materials have to exist in the languages spoken by significant portions of the community and in formats that people using assistive technology can actually read, and venues have to be reachable by the service the consultation is about. Disability rights organizations are worth engaging early for a second reason as well: they can tell you which design choices in the system itself will land hardest on riders who depend on predictable timing, which is exactly the technical input a general session is unlikely to produce. Treat their participation as expertise, not representation.

Naming the gap changes what you build. If the missing group works nights, the session happens in the morning at a transit hub rather than in the evening at City Hall. If the missing group speaks Spanish at home, a session runs entirely in Spanish rather than through consecutive interpretation of an English meeting. If the missing group cannot spare an evening at all, the channel has to be asynchronous and short. Each of those is more expensive than the default, and each one buys input the default cannot produce at any price, because the people it reaches were never going to be in the room.

The Four Participatory Methods That Work

Deliberative panels bring together a demographically representative sample of residents, typically 15 to 30 people, selected by stratified random sampling from the affected population, for a structured multi-session engagement. Participants receive accessible materials explaining the AI system and its potential effects before the sessions. Trained facilitators guide discussion of specific policy questions. The panel produces a written report of its findings and recommendations. Deliberative panels are more expensive than standard comment processes, and Siobhan's cost approximately $180,000 to design and run, but they produce substantive policy input from a representative sample, including populations that standard processes do not reach.

Community listening sessions bring agency staff into communities rather than asking communities to come to City Hall. Sessions are held in community centers, libraries, churches, and transit hubs. They are held at multiple times, including evenings and weekends. Interpretation services are provided in the languages spoken in the communities where the sessions occur. The sessions use plain-language materials prepared specifically for non-expert audiences. Siobhan's team held nine listening sessions across the city, including two conducted entirely in Spanish and one at a transit hub at 6 AM to reach morning shift workers. The listening sessions produced 847 individual comments across all sessions, compared to 94 comments received through the website, roughly nine times the volume from a channel the authority had never used before.

Online comment platforms with accessibility features provide a digital participation channel that reaches residents who cannot attend in-person sessions. Effective AI consultation platforms include plain-language explainers about the AI system and its specific policy choices; specific, answerable questions rather than open invitations to comment on AI in general; and multilingual support. Response rates increase significantly when questions are specific and when the platform makes clear that responses will be read and summarized by agency staff. The platform is a complement to in-person work rather than a replacement for it, since the residents least likely to attend a hearing are frequently the residents least likely to have a stable connection and an hour to spend on a government website.

Stakeholder working groups convene specific constituencies, including disability rights organizations, labor unions, civil liberties groups, business associations, and neighborhood groups, for structured engagement on the specific technical and policy design choices the agency is making. Working groups are more efficient than broad public sessions for engaging organizations that have the technical capacity to review documentation and provide detailed policy analysis. They are not a substitute for broader public engagement, but they are a valuable complement. The failure mode to watch is substitution by convenience: organized groups are easier to convene, easier to schedule, and easier to summarize, so a program under time pressure drifts toward them and calls the result public input.

Running It Under Delivery Pressure

None of this happens in a vacuum. Operations had been waiting eight months for the scheduling tool, the engagement office had seven people, and every week Siobhan added to the consultation was a week the authority did not get the operational benefit it had already justified internally. That pressure is real, it is legitimate, and pretending otherwise is how engagement offices lose credibility with the rest of the agency. The answer is not to argue that participation matters more than delivery. It is to be specific about what the extra time buys and to spend it on the parts that produce input the agency cannot get any other way.

Being specific means sequencing. The channels that take longest to stand up, representative recruitment and multi-session facilitation, have to start earliest, which means the decision to run them is made before the requirements are drafted rather than after. The channels that are fast, a platform with published questions and a set of community sessions, can run in parallel. And the analysis window has to be planned as work rather than absorbed by staff on top of their existing load, because a consultation that generates hundreds of comments and then sits unanalyzed for a month has produced a backlog rather than a policy input.

The other half is honesty about the timeline in public. Publish the consultation period and the decision date together, say what happens to comments received late, and if the schedule slips, say so and say why. Residents who gave up a morning to attend a session at a transit hub are entitled to know when the decision lands. An agency that goes quiet after collecting input teaches the population it worked hardest to reach that participating produced nothing, which is a more expensive outcome than never having asked.

Asking Questions People Can Actually Answer

The quality of the input a consultation produces is set mostly by the quality of the questions it asks. "What do you think about AI?" produces sentiment. A question tied to a real design choice produces policy. Siobhan's platform asked whether the scheduling system should be allowed to reduce service on routes with lower average ridership, even where those routes serve transit-dependent communities. That question is answerable by someone who has never heard the word algorithm, it names a trade-off the authority genuinely had to make, and the answers to it can be carried directly into the specification the vendor builds against.

Writing questions at that level is harder than it sounds, because it requires the agency to have identified its own real choices before consultation opens. A team that has not yet decided what is negotiable will write vague questions to preserve flexibility, and vague questions return vague answers that the team is then free to interpret however it prefers. The discipline is to publish the specific decisions that are still open, say which ones are not open and why, and accept that residents will sometimes attack the boundary you drew. That objection is itself useful input about a choice the agency made before anyone was asked.

Comment Analysis at Scale

A public consultation process that generates 847 individual comments faces an analysis challenge that traditional manual review cannot easily address. AI tools can assist with comment analysis, categorizing themes, identifying the most frequently raised concerns, and flagging comments that contain factual claims requiring staff response, while preserving the human judgment required to accurately interpret public input.

The appropriate role for AI in comment analysis is as a first-pass organizing tool, not as the analytical authority. Human staff must review the categorizations the AI produces, correct misclassifications, for example comments about privacy concerns categorized as scheduling concerns, and make the substantive judgments about which comments raise issues that require policy response. The categories themselves are a policy artifact: a comment that does not fit any category the agency defined in advance is exactly the comment most likely to contain something the agency had not considered, and an automated pass will tend to file it under the nearest available label rather than flag it as novel.

Two practical safeguards keep the tool honest. Read a sample of raw comments directly, chosen at random rather than from the categories, so staff retain a feel for what people actually said. And never let comment volume by category become the decision rule, because a single well-organized campaign can generate hundreds of near-identical submissions while a concern raised once by a rider who cannot get to work is the one that identifies a real defect. Counting comments measures mobilization. Reading them measures the policy.

Analysis is also where consultations quietly fail on resourcing. A channel mix that produced 847 comments from sessions plus 94 from the website has produced several weeks of reading for a team that still has its ordinary work, and the deadline for the published response does not move because the volume was higher than expected. Plan the analysis as staffed work with a named owner before the first session, decide in advance how a comment becomes a documented concern, and keep the raw record intact so a later reviewer can check the summary against it. The credibility of everything downstream rests on that record.

Transparency in Policymaking

Transparency in AI policymaking means that the public can understand what decisions the agency is making about AI, why it is making those decisions, and how public input affected the outcome. This is not the same as technical transparency, publishing source code or model weights, which is rarely meaningful to the public. It means publishing accessible descriptions of what the AI system does, what data it uses, what decisions it informs, and what oversight mechanisms apply.

Before a public consultation begins, the agency should publish a consultation document that includes: a plain-language description of the AI system being considered; the specific policy questions the agency is seeking input on; a description of who will be affected by the system and how; a summary of the risks and benefits the agency has identified; and the timeline for the consultation and the decision process. This document should be available in all languages spoken by significant portions of the community and in accessible formats.

After the consultation concludes, the agency should publish a summary of comments received organized by theme, a description of how the comments informed the agency's decision, and, where the decision does not reflect significant concerns raised in the public comment, a substantive explanation of why. The summary of public input that the agency publishes before making its final decision must accurately represent the full range of comments received, not just those that align with the agency's preferred approach, and not just those from the most organized commenters.

The accuracy of that summary is a legal exposure as well as an ethical one. OMB Circular A-4, the federal guidance on regulatory analysis, requires federal agencies to demonstrate that they considered public comments, and many state and local administrative law frameworks impose similar requirements. Publishing a selective summary that does not reflect the content of opposition or concern expressed in the public record is an administrative law risk as well as a transparency failure. Check what your own jurisdiction requires before you design the record, because the published summary is the artifact a reviewing body will read years later.

What Consultation Does Not Do

The most damaging idea in this field is that running a process settles the question. It does not. A comment period that was open, a listening session that was held, and a response document that was published are evidence that the agency asked. They are not evidence that the agency listened, and they do not by themselves make a contested system legitimate. Legitimacy comes from what visibly changed as a result, and residents assess it exactly that way: they look for the concern they raised and ask whether anything in the final design moved.

The same caution applies to duties. Where a resolution, statute or local ordinance requires consultation, whether a particular process satisfies that requirement is a question for your counsel and for whoever has authority to interpret the instrument, not something an engagement office can conclude on its own by pointing at an attendance sheet. Treat any procedural window, notice requirement or deadline your agency adopts as the agency's own commitment unless the source of the obligation says otherwise, and confirm the legal ones rather than assuming the customary ones carry legal weight.

There is a third limit worth stating plainly. Consultation surfaces the concerns of the people you reached, through the channels you built, about the questions you asked. It does not surface concerns nobody thought to raise, and it does not substitute for testing the system against the populations it will affect. A well-run consultation and a disparate impact analysis answer different questions, and an agency that treats the first as covering the second will be surprised later by a pattern that no resident could have predicted from a plain-language explainer.

Anti-Patterns

• The process as absolution. Running a comment period, holding a session, and publishing a response document, then treating the box as checked. The process is evidence that the agency asked. It is not evidence that the agency listened, and it does not confer legitimacy on a decision that was made before the consultation opened.
• Consulting after the decision. Opening a consultation once the vendor is selected, the design is fixed, and the launch date is set. Residents recognize this immediately, and the credibility cost extends to every future engagement the agency runs. If nothing about the answer could change the outcome, the honest move is to announce rather than to consult.
• Designing for the people who always show up. The Tuesday evening hearing and the website comment window feel neutral and are not. They produce a record dominated by organized groups and highly engaged residents, and an agency reading that record will conclude the affected population feels something it does not feel.
• Asking questions nobody can answer. "What are your thoughts on artificial intelligence in transit?" returns sentiment, not policy input. The vague question is usually a symptom: the agency has not yet decided which choices are actually open, so it avoids naming them.
• The curated summary. Publishing a comment summary that foregrounds supportive input and compresses opposition into a sentence. This is a transparency failure and an administrative law risk, and it is discovered by exactly the people whose comments were compressed.
• Comment counting as the decision rule. Weighing themes by volume rewards organized campaigns and buries the single comment that identifies a genuine defect. Volume measures mobilization; content measures the policy.
• Automated categorization without human reading. Letting a tool assign every comment to a predefined theme and reporting the totals. The comment that fits no category is the one most likely to contain something new, and an automated pass will file it under the nearest label rather than flag it.

Practice Prompts

• Map the gap. For an AI system your agency is considering, list the populations it will affect and how. Then list the groups that actually appear in the public records of your recent consultations. Write down the difference and design one channel specifically for the largest missing group, including time, place, language and format.
• Rewrite the questions. Take a consultation document your agency has published and replace every general question with a specific one tied to a real design choice the agency has not yet made. If you cannot find a real open choice, that is the finding, and it should go to whoever set the timeline.
• Price the options. Cost out a deliberative panel, a set of community listening sessions, and an online platform for one upcoming system. Compare each against the population it would reach and the type of input it would produce, and write the memo recommending a mix rather than a single channel.
• Draft the response document in advance. Before the consultation opens, write the outline of the document you will publish afterward, including the section that explains what you decided not to change and why. Notice which parts you are unwilling to commit to in advance, and ask why.
• Audit an old summary. Take a published comment summary from a previous process and compare it against the raw comments. Count how many distinct concerns appear in the record and how many are represented in the summary. Report the gap to your engagement lead as a process finding, not as a criticism of the drafter.

Reflection

Think about the last consultation your agency ran. Who showed up, and who was affected by the decision but absent from the record? What would it have cost to reach one of the missing groups, and who would have had to approve that cost? Were the questions you asked answerable by a resident with no technical background, and did any of the answers change the outcome? If a rider asked you today what their comment changed, could you point to a specific line in the final policy? And if the answer is no, is that because nobody raised anything worth acting on, or because the process was never designed to let them?

Glossary

• Deliberative panel. A demographically representative sample of residents, selected by stratified random sampling, convened over multiple structured sessions with accessible briefing materials and trained facilitation, producing a written report of findings and recommendations.
• Stratified random sampling. Selection that draws randomly within defined population strata so the resulting group reflects the composition of the affected population rather than the composition of the people who volunteer.
• Community listening session. An engagement held in a community setting at a time the affected population can attend, in the languages spoken there, using plain-language materials, rather than at a central government location on a fixed evening.
• Stakeholder working group. A structured engagement with specific organized constituencies that have the technical capacity to review documentation and provide detailed analysis, used as a complement to broad public engagement rather than a substitute for it.
• Consultation document. The material published before a consultation opens, describing the system in plain language, naming the specific policy questions at issue, identifying who is affected, summarizing risks and benefits, and stating the timeline for decision.
• Response document. The material published after a consultation closes, summarizing comments by theme, explaining how they informed the decision, and giving substantive reasons where significant concerns did not change the outcome.

Related Lessons

• Community Engagement in Government AI extends the outreach practice here into a continuing relationship rather than a single consultation.
• AI and Equity: Reaching All Communities covers the equity analysis that a consultation surfaces but cannot itself perform.
• Building Trust Through Transparency addresses what the agency publishes and why it matters to the people reading it.
• Public Reporting and Algorithmic Transparency takes the disclosure obligations further, into what an agency reports about systems already running.
• Transparency: Citizens' Right to Know frames the same duties from the citizen's side.
• Drafting Agency AI Policies is where consultation output has to land if it is going to change anything.
• State and Local Government AI Policy covers the jurisdictional frameworks that determine what your consultation is legally required to do.

Closing

Siobhan's consultation took longer and cost more than the process the authority had run for a decade. What it produced was a public record that included the riders the scheduling tool would actually affect, asked them about choices the authority genuinely had to make, and left the agency able to say specifically what changed as a result. That last capability is the one that survives contact with a City Council hearing, a reporter, or an advocacy group that disagrees with the outcome. An agency that can point to the comment and the line it changed is in a different position from an agency that can only point to an attendance sheet.

The work is not exotic. Decide who is missing from your record before you choose a channel, build for them, then ask questions specific enough that the answers can move a design. Design for the people who always show up and you will get a record that tells you what you already believed. Design for the people who are missing and you will get one that tells you something you did not know, which is the only reason to run a consultation at all.

Key Takeaways

• Standard comment processes consistently miss the most affected populations. Transit-dependent riders, non-English speakers, night shift workers, and others most likely to be affected by AI decisions in government services are the least likely to participate in Tuesday evening hearings and 30-day website comment windows.
• Start from the gap, not the calendar. List who the system affects, list who appears in your existing public record, and design the consultation for the difference. The channel, the time, the place and the language all follow from that list.
• Bring sessions to communities, not communities to City Hall. Listening sessions held in community centers, transit hubs, libraries, and churches, at varied times, with interpretation services, reach significantly more residents than centralized hearings and produce substantially more comments.
• Specific, answerable questions produce substantive policy input. Broad questions about AI generate rhetorical responses. Specific questions about real design choices produce input that can actually inform policy, and writing them forces the agency to identify what is still open.
• AI can assist comment analysis but requires human review. AI categorization of large comment volumes speeds the analysis process but must be reviewed by human staff for misclassification, and human judgment is required to determine which comments raise issues that need policy response.
• Published comment summaries must accurately represent opposition. Selective summaries that do not reflect the range of concerns expressed in public input are an administrative law risk. The published record must be comprehensive, not curated to support the preferred outcome.
• Transparency requires explaining how public input affected the decision. Publishing a comment summary and then making the decision the agency intended all along destroys public trust. Agencies must be prepared to show, specifically, which concerns the public raised and how those concerns changed what the agency decided.
• Running the process is not the same as discharging the duty. A held session and a published response show that the agency asked. Whether a given process satisfies a legal participation requirement is a question for counsel, and legitimacy comes from what visibly changed.
• Deliberative panels produce representative input from populations standard processes miss. The $150,000 to $200,000 cost of a well-designed deliberative panel is justified when the AI system being considered will significantly affect large or vulnerable populations whose standard participation in government processes is limited.

Frequently Asked Questions

We do not have $180,000 for a deliberative panel. What is the minimum viable version? The expensive part of a panel is representative recruitment and professional facilitation, and neither is what makes the underlying idea work. What makes it work is reaching people who would not otherwise participate, giving them accessible material before you ask for input, and asking about real choices. Listening sessions in community settings deliver a large share of that at a fraction of the cost, which is why Siobhan ran nine of them alongside everything else. Choose the panel when the system will significantly affect large or vulnerable populations and the authority needs a defensible, representative record.

How do we handle an organized campaign that floods the comment channel? Read it and count it separately. A campaign is real public input and tells you that an organized constituency cares, which is useful, but treating volume as the decision rule lets the best-organized group set the policy. Report near-identical submissions as what they are, a campaign of a stated size, and analyze distinct concerns on their content. The single comment describing a specific harm nobody anticipated deserves more staff time than another near-identical copy of a form letter.

What if the consultation tells us something we cannot act on? Say so, specifically, in the response document. Residents accept constraints far better than they accept silence. A response that names the concern, explains the budgetary, legal or operational reason the agency cannot address it, and says what the agency will do instead preserves credibility. A response that omits the concern entirely reads as though nobody was listening, which is the outcome the whole process exists to avoid.

Does running a consultation satisfy a legal participation requirement? That is a question for your counsel and for whoever interprets the instrument that created the requirement, not one an engagement office can settle by pointing at a process. Design your consultation to be substantively good, document what you did and what it changed, and get the legal sufficiency question answered in writing before you rely on it. Treat comment windows and notice periods your agency sets as the agency's own commitments unless the underlying authority says otherwise.

How early should consultation start? Early enough that the answers can still change the design, which in practice means before a vendor is selected and before the requirements are locked. If the timeline does not allow real consultation, the honest options are to move the timeline or to announce the decision rather than dress it as an engagement.