Terrence Abara-Flynn's first Congressional briefing lasted eleven minutes before the ranking member of the House Oversight and Accountability Committee stopped him cold. Terrence, then a newly appointed Deputy CIO (Chief Information Officer) at the Social Security Administration, had spent three weeks preparing slides on model architecture, training data pipelines, and F1 scores, a measure of how accurately a machine-learning model classifies cases. He got through slide four before Representative Delgado leaned into the microphone and said, "Sir, I don't need to know how the engine works. I need to know if my constituents are going to get their disability checks." Terrence flew back to Baltimore on a Thursday afternoon understanding something no technical certification had taught him: Congress is not a technology conference. It is a constituent-service forum with subpoena power.

Think of It as a Constituent Pipeline, Not a Data Pipeline

The controlling analogy for Congressional communication is a constituent pipeline. Every bill, every budget vote, every oversight hearing flows through one question: what does this mean for the people who elected me? Your AI program is a constituent-service delivery mechanism, whatever the architecture diagram says. Frame every claim, every risk disclosure, and every budget number through that pipeline before it leaves your mouth. The discipline is not simplification. It is deciding, sentence by sentence, which end of the system you are describing: the model, or the person waiting on the model's output.

When you say "our ML model achieves 94% accuracy," a staffer hears noise. When you say "our model helps process 1.2 million disability applications per year with fewer errors, meaning fewer beneficiaries wait more than 90 days for a decision," the staffer hears a win for their boss's district. The second sentence contains the same fact as the first. It simply attaches the fact to a person. Run every sentence through the constituent pipeline before you speak it aloud in a hearing room, and run your slides through it as well.

What Congress Is Actually Deciding

Congress shapes policy. Congress holds the power of appropriations. Congress oversees agency operations. Those three levers explain why Congressional understanding of your AI program is not a public relations concern but an operational one. Congressional misunderstanding of AI produces bad policy in three predictable shapes: overly restrictive rules that slow beneficial deployment, inadequate safeguards that permit harmful ones, and requirements written without regard for operational constraints that agencies then cannot actually satisfy. Your job in a briefing is to supply the real-world complexity that lets Members choose between those outcomes knowingly.

Assume nothing uniform about the room. Some Members understand the technology in detail. Most do not. Some arrive with strong, already-formed views on AI policy. Others are engaging with the subject for the first time because a committee assignment put it in front of them. Some care most about innovation and national competitiveness; others care most about risk and citizen rights. All of them are operating under severe time constraints and real political pressure from constituents, and all of them are weighing your program against competing proposals from other agencies and outside advocates.

Know Which Committee You Are Facing

Not all committees want the same thing. Walking into the wrong conversation with the wrong frame is the second most common mistake after leading with technical details. The table below summarises what each committee typically cares about, roughly how much technical depth and hearing time you can expect, and the framing that tends to land. Treat it as a starting hypothesis to check against the committee's recent hearing record, not as a substitute for that record.

Appropriators care about money. They want total program cost, cost per unit of output, the requested budget line, and return over a three-to-five year horizon. The SSA's Intelligent Document Processing pilot cost $4.2 million in fiscal year 2024 and eliminated 180,000 manual data-entry hours. That is a story Appropriations can use in a markup. A ROI (return on investment) figure, dollars saved per dollar spent, lands well here, and so does a credible alternative cost: what the same workload would cost handled manually. Have both ready before you sit down, with the assumptions written on the page.

Oversight committees are adversarial by design. Their job is to find failure, and they will ask about every error your system has made, every FOIA (Freedom of Information Act) request filed about the program, every IG (Inspector General) finding, and every public complaint. Prepare a brief chronology of incidents and corrective actions. Transparency about past problems, not defensiveness, is the posture that survives an Oversight hearing. "We identified the problem in March, paused the system for 14 days, retrained the model, and re-deployed with enhanced human review" is a credible answer. "We're confident in our systems" is not an answer at all.

Armed Services will ask about adversarial risk if your agency touches national security or defense acquisition: the possibility that a foreign actor could manipulate your AI system. Expect questions on supply-chain security in your model vendors, data residency, and system resilience in a crisis. Have your CISA (Cybersecurity and Infrastructure Security Agency) coordination documents on hand and know which ATO (Authority to Operate), the formal security authorization for the system, covers it. Know your vendor concentration too, because "we do not rely on any single supplier for a critical component" is the answer this committee is listening for.

Judiciary members focus on civil rights, privacy, and due process. They will ask whether your system produces disparate outcomes for protected classes, how an individual can appeal an AI-assisted decision, and what legal authority permits automated decision-making in your program at all. Have your equity impact analysis ready. Know your statutory authority by citation, not just by name, because a Judiciary staffer will look it up while you are still talking. This is also the committee most likely to ask what happens to a person who does not know AI was involved in their case.

Reading the Member, Not Just the Committee

Within any committee, individual Members fall into four recognisable postures. Tech champions know the field, understand the technology and advocate for innovation; give them technical depth and discuss opportunity, while recognising that they may push for less oversight than the public will accept. Tech skeptics do not understand the technology, are wary of AI and are focused on risk; explain plainly, address risks directly and show the safeguards, recognising that they may restrict beneficial work.

Pragmatists understand the technology reasonably well and want to know real-world impact; focus on operations, show what works and acknowledge what does not. There is no significant risk in dealing with a pragmatist provided you are being honest, which is the whole point of the category. Disengaged Members have not yet paid much attention to AI; brief them on why it matters and what the implications are for their district, because the alternative is that they default to whatever the loudest stakeholder in their inbox is saying. Four postures, one rule: adapt the depth, never the facts.

The Briefing Pyramid

Good briefing structure works whether you have five minutes or fifty. Lead with the bottom line up front, the single point you want the room to retain: that the agency is deploying AI in compliance with federal requirements, or that a system will cut processing time while protecting individual rights, or that you need Congressional approval to expand a capability. Do not bury it. Then give brief context that answers why Congress should care, and be specific about impact: who is affected, how many people, and what is at stake if the program stops or expands.

Follow with three to five key facts that support the point. Use specific numbers rather than adjectives: how many systems you have deployed, how many impact assessments you have completed, how many critical risks those assessments found. Use concrete examples of a system doing something a Member can picture. Use a chart where a chart shows impact better than a sentence. Then address the concerns you know are coming before anyone asks, which signals that you have thought about the downside rather than been surprised by it.

Close with the call to action. What do you actually need from Congress? A funding level, an authority, a statutory change, or simply their support in a markup. Be specific, because a vague ask is an ask that no staffer can act on. The pyramid is not a script. It is an ordering principle that survives interruption: if the hearing derails after ninety seconds, the room has still heard your bottom line, your context, and at least one fact that supports both.

Fitting the Brief to the Clock

Three common formats come with standard time allocations. For a five-minute briefing: thirty seconds on the bottom line, thirty seconds on context, two to three minutes on key facts, thirty seconds on the ask, and a minute reserved for questions. For a fifteen-minute briefing: a minute on the bottom line, a minute on context, six to eight minutes on key facts, two to three minutes addressing concerns, one to two minutes on the ask, and one to two minutes for questions. For a forty-five-minute hearing appearance: a five-minute prepared statement, leaving forty minutes of questions.

Check the arithmetic before you rehearse. The hearing allocation sums exactly. The two shorter allocations are stated as ranges, and at the top of every range they overrun the slot: the five-minute components run from four and a half to five and a half minutes, and the fifteen-minute components run from twelve to seventeen. Treat those ranges as a budget to be cut, not a schedule to be followed. Decide in advance which key fact you drop when the gavel moves faster than you do, and rehearse the shortened version rather than discovering it live.

The One-Page Policy Brief

Every briefing should produce a one-pager that a chief of staff can read in 90 seconds. Staff read these standing in hallways between votes, which is the real design constraint. Structure yours in four blocks, in this order.

1. The problem in constituent terms. "Each year, 340,000 Social Security applicants wait more than six months for a disability decision because of manual document review backlogs."
2. What the agency is doing. "The SSA's automated document sorting tool reviews incoming medical records and flags complete versus incomplete applications, letting claims examiners focus on substantive review rather than paperwork."
3. The safeguards. "No application is denied by the system. Every flagged case is reviewed by a person before any action is taken. The system has been audited for accuracy across demographic subgroups."
4. The ask or the bottom line. "The FY 2026 budget request includes $6.8 million to expand the tool to all ten SSA processing centers, reducing average wait times from 180 days to 120 days."

Do not bury the constituent outcome in paragraph three. It goes first, always. And write the safeguards block only as strongly as your documentation can support: every clause in block three is a claim an Oversight staffer can test against your records, and a safeguard you cannot evidence is worse than a safeguard you never claimed. If a person reviews every flagged case, be able to show the review log. The 60-day reduction in block four is likewise a projection, and it should be labelled as one.

Translating Technical Concepts Without Losing Accuracy

Translation is not dumbing down. It is choosing the right level of abstraction for the audience. Lawmakers are sophisticated about budgets, political risk, and constituent impact. They are not trained in statistics. Use analogies drawn from processes they already understand, and test the analogy on someone outside your field before you use it on the record. An analogy that survives a skeptical colleague will survive a hearing; one that only makes sense to you will collapse under the first follow-up question.

Instead of "the model uses a transformer architecture trained on 40 million labeled records," say "the system works like an experienced examiner. It has reviewed millions of past cases and learned to recognise whether a file is complete, the same way a veteran employee knows at a glance whether a folder is ready for review." Machine learning generally translates as a translator: show a model thousands of examples of English rendered into Spanish and it learns the patterns, then applies them to sentences it has never seen. Your system learns from historical data the same way.

Neural networks translate as layered processing: early layers learn simple patterns and later layers combine them into more complex recognition, roughly the way vision works. A large language model translates as prediction: you give it a prompt, it predicts the most likely next word, then the next, until it has produced a full response, and that is sophisticated prediction rather than comprehension. Bias translates as bias: just as people can carry bias, an AI system can amplify the bias in its training data, and if historical records show a group being denied benefits unfairly, the system learns that pattern.

Concrete beats abstract every time. Instead of "the system uses supervised learning with labeled data," say "we trained it on ten years of actual benefit determinations, marked which ones were correct and which had errors, and it learned from those examples." Instead of "the model achieves 94% accuracy on the validation set," say "when we tested it on cases it had not seen before, it made the right decision 94% of the time. That is better than our human processors, who are at 90%, and we still review the decisions."

The same rule kills jargon. Rather than "we implement XAI techniques using SHAP values to provide attribution explanations," say "we can show which factors mattered in a decision: income was over the limit, assets exceeded the threshold, work history showed disqualification." Rather than "the system exhibits hallucinations in generative tasks," say "it sometimes makes up information, and if you ask it for a report on a family it may invent details that are not true." Rather than "we conduct red team exercises against frontier models," say "we have security experts try to break the system, and we fix what they find."

Contextualise every accuracy number. A 94% accuracy rate sounds alarming in isolation. Comparing it to the baseline the system replaced gives lawmakers a meaningful frame, provided you say plainly what the comparison measures and on what data. And every acronym must be spelled out the first time you use it. Do not assume a Member on an Appropriations subcommittee knows what an LLM (large language model) is. Spell it, define it in one sentence, and move on without a pause that suggests you think the question was foolish.

Addressing Political Concerns Before They Become Questions

Members focused on civil liberties will raise the concern that the system discriminates against vulnerable populations. Acknowledge it first, because it is a real risk with AI and pretending otherwise costs you the room. Then show what you have tested and what you found, describe the ongoing monitoring cadence and the escalation path when monitoring flags something, and describe what remediation has actually looked like when you found a problem. Close by inviting oversight rather than tolerating it. An offer to report findings on a regular schedule is worth more than any assurance you can give in the moment.

Members focused on innovation will raise the concern that your requirements slow things down. Acknowledge that innovation matters and that you want beneficial systems deployed quickly. Then show that you are not over-restricting: a fast-track path for low-risk systems where assessment takes weeks rather than years, if you genuinely have one and can show the throughput data. Explain why safeguards enable deployment rather than block it, since public trust is the precondition for scale. Then point to what you are actively enabling: sandboxes for experimentation, funded research, shared infrastructure other agencies can use.

Fiscal hawks will raise cost. Show the cost-benefit with the inputs visible, not just a headline ratio. Show efficiency gains from shared infrastructure and what has happened to cost per agency. Show the alternative, which is what manual processing or no system at all would cost over the same period. Show the long-term shape: an initial investment that is high, operating costs that fall in later years, and the point at which the program becomes cheaper than what it replaced. If you cannot show that point with real numbers, say so rather than gesturing at it.

Members focused on national security will ask whether the supply chain is secure and whether critical components depend on a strategic competitor. Acknowledge that supply chain security is essential. Show your vendor vetting, including which components are sourced domestically or from allies. Show continuous vendor monitoring and the contingency plan if a vendor fails. Show resilience through avoiding single-vendor dependence for anything critical, with named alternate suppliers. This is the one concern where an incomplete answer will generate a follow-up letter rather than a follow-up question.

Anticipating Hostile Questions

Prepare for three categories of hostile question, and assume all three will appear in the same hearing. The pattern for answering each is the same: acknowledge the question as fair, show the evidence, show the process that generates the evidence continuously, and invite scrutiny of both.

AI risk and errors. Members will ask what happens when the system is wrong. Have a concrete answer: "When the system flags an application incorrectly, a human examiner catches it before any action is taken. In FY 2025, our quality-assurance review found 2,100 flagging errors, and examiners corrected all of them. We are not aware of any applicant receiving an incorrect determination as a result of the system that year." Quantify the error rate, quantify the catch rate, and describe the human backstop precisely. Say that human review is the control, and say what the control's own failure mode is, because a Member who has read one IG report knows that reviewers rubber-stamp.

Privacy. Members will ask what data the system uses and who can see it. Know your Privacy Impact Assessment (PIA), the formal analysis of how a system collects and uses personal information, and whether it has been published. Know your data retention rules. Know whether your vendor has access to citizen data and under what contractual restrictions. "Our PIA was completed in November 2024 and is publicly available on SSA.gov" is a full answer. "We take privacy seriously" is not an answer at all. Be ready for the follow-up: what the PIA covers, and what it does not.

Job displacement. This question will come from Members whose districts hold large federal-workforce concentrations. Have workforce numbers ready. "This system did not eliminate any positions. It reassigned 47 claims examiners from document sorting to substantive case review, where demand exceeded capacity. Our union partnership agreement, signed in January 2025, covers all workflow changes." Connect the tool to redeployment rather than reduction wherever you can do so honestly. Where reduction has occurred, be direct and explain the attrition or transition plan rather than letting a staffer find the headcount in a budget table.

Three question patterns recur regardless of category. "How do we know this is not just another way to discriminate?" gets acknowledgement, the specific findings, the recurring test cadence, and an invitation to independent audit. "What happens when the AI makes a mistake?" gets honesty that AI systems do make mistakes, the comparison to the alternative, the oversight that catches most of them, and a real example of an error you caught and fixed. "How do we know you are actually following your own AI governance policy?" gets the reporting record, the audit findings, the published inventory, and a channel for anyone who thinks you are out of compliance to say so.

The hardest question is the one where the honest answer is that you do not know. Asked whether your bias testing is adequate, "we are confident our testing is adequate" is the worst available response. Better: you do not know for certain, this is a genuinely hard problem, you are using the testing approaches the field currently accepts, you are transparent about their limits, you are funding work with universities to improve them, and you will adopt better methods when they emerge. Humility is not a weak position in front of a committee. It is the position that survives contact with the next IG report.

Terrence put it this way afterwards. The question he feared most was never about the algorithm. It was always some version of: can you tell me what happens to a 58-year-old widow in Youngstown if your system makes a mistake? He learned to prepare that answer first, before he prepared anything else.

Preparing Formal Testimony

Preparation starts with the committee, not with your program. Find out what bills they are considering, what recent hearings they have held on AI, what positions individual Members have already staked out, and what they asked the last witnesses who sat in your chair. That research determines which of your facts are relevant. Then build talking points: your core message in two or three sentences, supporting points for each, examples that illustrate them, and written answers to the questions you expect. Keep any visual aids simple enough to be understood without narration.

Written testimony is submitted 48 hours before a hearing and enters the public record. Journalists, advocates, and opposing counsel will read it. Draft at a level accessible to an engaged non-specialist, comprehensive but scannable, with exhibits where they help. Have general counsel review every factual claim. Have your IG liaison check it against any open audit findings, because a discrepancy between your testimony and an IG report generates follow-up hearings, and the follow-up hearing is always worse than the original one.

Oral testimony should run five minutes or fewer. Members will interrupt; that is not rudeness, it is the clock. Speak clearly and deliberately, make eye contact, refer to your notes rather than reading them, and pause for questions. Do not speculate. If you do not know something, say so and offer to follow up in writing. Answer the question asked rather than the question you wish had been asked, keep answers to about thirty seconds unless pressed, correct your own misstatements immediately and on the record, and stay respectful of Members who are not being respectful of you.

Rehearse in front of people who will be unkind. Staff call these murder boards: practice sessions where colleagues ask the worst questions they can imagine, in the worst tone they can manage. Run at least two before any high-stakes hearing, and make sure one of them includes a participant with no technical background, because the questions that damage witnesses are almost never the technically sophisticated ones. Time yourself in every run. Refine based on what actually failed rather than what you felt uncertain about.

Anti-Patterns

• Assuming Congress understands the technical details. You explain gradient descent optimisation and the room glazes over, having learned nothing about why any of it matters. The message does not land and your credibility goes with it. Use analogies, use examples, cut jargon, and test the explanation on someone outside your field first.
• Answering concern with denial. A Member asks whether there are risks with the system and you say it is completely safe. Congress knows no system is completely safe, so the answer reads as concealment. Acknowledge the risks, explain how you manage them, and show that you have thought seriously about the downside.
• Claiming more certainty than the evidence supports. You testify that the system achieves 99% accuracy and real-world performance turns out to be 87%. The committee concludes it was misled, and every future briefing you give starts from a deficit. Use ranges and name the conditions: 85 to 90% in controlled testing, with real-world performance still being characterised.
• Not preparing for the hardest question. You brief on the benefits of AI in hiring, a Member asks whether it discriminates against protected classes, and you have no answer ready. Identify what each Member will care about and write the answers before the hearing, not during it.
• Selling human review as a guarantee. "A person reviews every flagged case, so no one is harmed" describes the design, not the outcome. Human review fails in known ways: reviewers under quota pressure approve the queue, and a reviewer shown only the system's recommendation tends to agree with it. Report a clean year as evidence that the backstop held, not as proof that it cannot fail, and report what you do to check that reviewers are actually reviewing.
• Offering the published inventory as proof of compliance. An inventory entry shows that a system was listed. It does not show that the listing is current, that the governance described actually ran, or that a system missing from the list does not exist. Offer the inventory as one artifact among several, alongside the audit findings and the reporting record that test whether the artifact is true.
• Reciting "no disparate impact" as though it closed the question. What a test can support is that no statistically significant difference was found, in the groups you measured, on the data you had, at the power your sample allowed. Say it that way. A finding of no difference in three demographic groups says nothing about a fourth you did not measure, and a committee that discovers the gap later will treat the original phrasing as spin.

Practice Prompts

1. Pick one committee that oversees your agency. Research what it is currently interested in, what bills it is considering, what its Members have said publicly about AI, and what it asked the last agency witness who appeared. Write down what that committee would want to hear from you and what it would not care about at all.
2. Design a fifteen-minute briefing on your agency's primary AI system for that committee, using the pyramid: bottom line, context, key facts, concerns addressed, call to action. Then cut it to five minutes and note which facts you dropped and why.
3. Explain one technical aspect of your system, how it works, how you test it for bias, or how you provide explanations, to a Member who knows nothing about AI. Use an analogy and a concrete example. Read it to a colleague outside your field and revise wherever they hesitate.
4. Write down the five hardest questions Congress could ask about your AI systems, including at least one where the honest answer is that you do not know. Draft answers. Have a colleague ask them out loud and refine what fails.
5. Draft a five-minute opening statement for a hearing on your agency's AI work. Run it as a murder board with at least one non-technical participant. Time every run and revise based on what broke, not on what felt uncomfortable.

Reflection

Take twenty minutes with these questions. Which committees have jurisdiction over your agency on AI matters, and when did you last read their hearing record? What do those committees care about most, and does your current briefing material reflect it? What tough questions will they ask about your systems, and which of those questions do you currently answer with an assurance rather than evidence? How prepared is your team to brief them, in the honest sense of having documents you could hand over today? And what one change would most improve your agency's Congressional engagement over the next year?

Glossary

• BLUF (bottom line up front). The key point or conclusion stated at the beginning of a briefing, so the audience gets the main message even if they do not stay for the rest.
• Mark-up. The committee process in which bill language is debated and amended before a vote.
• Floor vote. The process by which the full chamber votes on legislation.
• Appropriations. Congressional action allocating funding to specific programs or agencies.
• Oversight. The Congressional responsibility to monitor agency operations and compliance with law.
• Murder board. A rehearsal in which colleagues ask a witness the hardest and most hostile questions they can devise, to surface weak answers before the hearing does.
• PIA (Privacy Impact Assessment). The formal analysis of how a system collects, uses, shares and retains personal information.
• ATO (Authority to Operate). The formal security authorization permitting a system to run in a production environment.

Related Lessons

• Drafting Agency AI Policies supplies the internal policy record that a committee will ask you to describe under oath.
• Navigating the Federal AI Landscape maps the federal requirements a Member is most likely to ask whether you meet.
• State and Local Government AI Policy matters when a Member's question is really about what their state is doing differently.
• Congressional and IG Reporting on AI covers the written reporting obligations that sit behind every oral briefing.
• Oversight Mechanisms: IG, GAO, Congress explains how the three oversight channels interact and where your testimony lands in each.
• Communicating AI Projects to Leadership is the internal analogue of this skill and the place most people practise it first.
• Public Consultation on AI Policy covers the constituent-facing engagement that Members will ask whether you have done.

Closing

Congressional briefing is part craft and part discipline. Understand the audience, communicate clearly, anticipate the concerns, and build credibility across repeated contact rather than in a single appearance. Get it right and Congress becomes a partner in developing workable AI policy. Get it wrong and Congress becomes a constraint that costs your agency options it needed. The agencies that do this well treat it as a real capability: they staff it with people who can communicate, they prepare carefully, they engage honestly, and they educate committees over years rather than lobbying them over weeks.

Terrence's second appearance before the same committee ran the full scheduled time and produced two follow-up requests in writing rather than an interruption on slide four. Nothing about the underlying system had changed. What changed was that every claim he made started at the end of the constituent pipeline and worked backwards, and every safeguard he described came with a document he could hand over. Congress will make AI policy whether or not your agency helps. Helping is the cheaper option.

Key Takeaways

• Lead with constituent outcomes, not technical specifications. Every claim should answer the question a lawmaker's constituent would ask before it answers the question a conference reviewer would ask.
• Know your committee before you walk in. Appropriations wants cost and return, Oversight wants incident history and corrective action, Armed Services wants security and supply chain, Judiciary wants rights and legal authority, Science wants research and standards.
• Read the Member as well as the committee. Champions want depth, skeptics want safeguards, pragmatists want operational reality, and the disengaged will default to whoever briefed them last. Adapt the depth, never the facts.
• Structure the one-pager in four blocks: the constituent problem, what you are doing, the safeguards, and the ask, in that order, with nothing in the safeguards block you cannot evidence on request.
• Translate every technical term into a process the Member already understands and spell out every acronym on first use, without exception and without a pause that implies the question was foolish.
• Prepare written answers to three hostile categories before any hearing: AI errors and the human backstop, privacy and the published PIA, and job displacement with workforce transition data.
• Say what a control does and how it fails. Human review, a published inventory and a bias test are evidence, not guarantees. A committee that discovers the limits after your testimony will treat your original phrasing as spin.
• Written testimony enters the public record. Confirm every factual claim with counsel and check it against open IG findings before submission.
• Run at least two murder boards, including one with a non-technical participant, before any high-stakes committee appearance, and time every run.
• The constituent pipeline is the discipline. Run every sentence, every slide, and every briefing document through one question: what does this mean for the people who elected the person sitting across from you?

Frequently Asked Questions

How much technical detail is too much? The test is not the quantity of detail but whether each detail changes what a Member would decide. Model architecture rarely does. Error rates, human review coverage, appeal rights and cost per case usually do. If you cannot say what decision a fact informs, it belongs in the written testimony rather than in the five minutes you have out loud.

What if I do not know the answer in the hearing? Say so, then offer to provide the answer in writing, and then actually provide it on the timeline you promised. Speculation in a hearing room is how agencies acquire commitments they never intended to make. A witness who says "I do not have that figure with me and will provide it" is treated far better than a witness who guesses and is corrected later by their own IG.

Should I disclose a problem the committee has not asked about? If the problem is material and you know it, assume the committee will learn it eventually, and that learning it from a third party will be worse than learning it from you. Disclosure is also not a one-time act: raising it in a briefing does not close it, and the committee will expect an account of what you did about it.

Who should actually deliver the briefing? Whoever can answer follow-up questions without turning around to a staffer. Technical depth in the room matters less than the authority to speak for the agency and the ability to say plainly what the agency does not yet know. Bring the technical expert as a second chair rather than as the witness.

How do I handle a Member who is clearly hostile? Answer the question that was asked, stay in the constituent frame, and do not match the tone. Hostility in a hearing is frequently performance for a district audience rather than a considered position, and the transcript outlives the exchange. The respectful, specific, boring answer is the one that reads well six months later.

How often should we brief, and about what? One briefing changes very little. Consistent contact does: a short update when a system changes materially, a note when an incident is resolved, an offer to walk staff through a new report before it publishes. Relationships built between hearings are what make the hearing itself survivable.