Hand off the work you can check faster than you can do.
The useful test for AI in lobbying is not whether it can do a task. It is whether you can verify the result in a fraction of the time it would take you to produce it. Reading a three-hour hearing to find the two minutes that matter to one client is slow to do and quick to check, provided the summary points you to the timestamp. Deciding whether to push a client’s ask before markup is the opposite: there is nothing to check, because the judgment is the work.
That test points to four jobs most lobbyists do every week. The workbench below runs each one on fictional material, so you can see the shape of good output and the checks a careful practitioner runs before using it.
Reading, sorting, first drafts, turning notes into tasks. Sources let you check in minutes.
Timing an ask, reading an office, deciding what to trade. AI can lay out facts; the call is yours.
Pick a job. Watch a long input become something you can check.
Fictional clients and documents. The highlighted passages are the ones the AI relied on; follow them into the result.
- [00:02:10] CHAIR: The subcommittee will come to order. Today we examine the budget request for…
- [01:02:40] REP. ALVAREZ: I want to ask about range capacity. We hear test windows are booked out.
- [01:03:15] WITNESS: We are looking at options, including partnerships with commercial test sites.
- [01:40:02] MEMBER: On a separate matter, the shipbuilding schedule…
- [02:11:05] RANKING MEMBER: I would like a briefing on workforce at the ranges.
- [03:09:48] CHAIR: Without objection, the record will remain open for ten days.
Nothing to check yet. The AI is reading.
An interactive example with four jobs: a hearing transcript becomes a five-line client note, a long bill becomes a list of affected clients with section and page, raw meeting notes become a follow-up draft and dated tasks, and a client’s ask becomes a first draft of an appropriations request. For each, the long input collapses into a short output, and then the checks a lobbyist runs before using it appear.
What to give it, what good looks like, and where it goes wrong.
The quality of AI output for lobbyists depends far more on what you hand over than on clever wording. Give it the client’s context and a clear format, and ask for sources on every line.
A hearing transcript becomes a five-line client note.
GIVE ITThe transcript or the official video link, the client’s current asks, and the members and issues you care about for that client. Ask for no more than five lines, each with a timestamp.
GOOD OUTPUTShort, specific and honest about what did not happen. “Your program was not mentioned” is a useful line. Every claim points to a minute you can find in seconds.
THE TRAPCharacterization. A summary may describe a question as support or a polite answer as a commitment. Tone is exactly what a client will quote back to you.
A long bill becomes “which clients, and where.”
GIVE ITThe bill text and a short profile of each client: programs, facilities, contract types, the terms they would use. Ask for client, section, page and one line of why.
GOOD OUTPUTA list you can verify by opening pages, ranked by how directly each provision touches the client, including the clients it searched for and did not find.
THE TRAPFalse comfort. A clean “not affected” is only as good as the search terms behind it. Bills use statutory language; clients use program names.
Raw meeting notes become a follow-up and dated tasks.
GIVE ITYour notes as they are, typos included, plus who you met and for which client. Ask for a short thank-you draft and a task for every commitment, with a date.
GOOD OUTPUTA draft that sounds like you on a good day, tasks that name who does what by when, and anything uncertain (a name, a date) flagged rather than filled in.
THE TRAPPromises you did not make. If the draft commits the client to sharing data, confirm the client agreed before the note goes out.
A client’s ask becomes a first draft of an appropriations request.
GIVE ITThe client’s email, last cycle’s request if there was one, and the office’s request form. Ask for a draft in that form with blanks wherever a fact is missing.
GOOD OUTPUTThe program, a two-sentence justification and optional report language, drafted from the client’s own history, with the account and amount left open for you to confirm.
THE TRAPConfident numbers. Account lines and program elements change between budget cycles. Every figure comes from the client in writing or from the budget books.
Find the three lines you should never send.
Reviewing AI output is a skill, and it gets faster with a short list of what to look for. Most errors in lobbying work fall into three kinds: a commitment nobody made, a number nobody checked at the source, and a date carried over from last year. Here is a draft client note with one of each.
Three of these six lines would embarrass you in front of the client. Tap the ones you would not send.
Look for promises nobody made, numbers you have not seen at the source, and dates without a current citation.
Five lines AI should not cross in a lobbying practice.
Handing off the reading and the first drafts works because a person still owns everything that leaves the firm. Keep these boundaries fixed and the rest of the time savings are low-risk.
Anything going to an office or a client goes out under your name, after you have read the exact words.
A note to a chief of staff you have known for years should come from you. AI can remind you what you discussed last time; it should not write the warmth.
Whether to push now, wait for markup or trade one ask for another is strategy. AI can lay out the facts.
If a statement cannot point to the page, minute or email it came from, treat it as unverified.
Client strategy, numbers and documents only go into tools your firm has approved for client work.
How to start using AI for lobbying work this week.
You do not need a firm-wide rollout to learn what works. You need one recurring job, two weeks and an honest comparison.
- STEP 1Pick one job.Choose the one of the four that lands on your desk every week, not the most impressive one.
- STEP 2Run it twice in parallel.Do your own version too. Compare what the AI caught, missed and invented.
- STEP 3Write your checks down.Start from the lists above. Keep them next to the output until they are habit.
- STEP 4Use an approved tool.Client material belongs in a system built for it, with sources on every finding.
- STEP 5Keep what you trust.After two weeks, keep the jobs that check out, drop the rest and add the next one.
In Capiro, the AI back office for lobbyists, these four jobs are part of the daily work of Meri, the AI Chief of Staff inside the back office. Meri reads hearings and bills against each client’s priorities, turns meeting notes into follow-ups and tasks, and drafts requests from the client’s own history. Every finding carries its source, and nothing leaves the firm until a person approves it.
Frequently asked questions
Can AI summarize a congressional hearing accurately?
Usually well enough to tell you where to look, as long as every line of the summary cites a timestamp. Treat it as an index into the hearing rather than a substitute for it. Before you repeat anything to a client, especially a reading of a member’s tone or intent, watch that minute yourself.
How do I check an AI summary of a bill against my clients?
Ask for a section and page number behind every claim, open each one and read the text. Then check the negatives: when it says a client is not affected, look at the search terms it used, because bills rarely describe an issue in the words your client uses.
Can AI write an appropriations request?
It can write a solid first draft from the client’s ask and last cycle’s request: the program, a justification and optional report language. It should leave blanks for the account, program element and amount until you confirm them, and the final version has to follow each office’s own form and deadline.
Is it safe to put client information into an AI tool?
Only into a tool your firm has approved for client work. Look for data kept private to your firm and not used to train public models, access controls by client, and answers tied to their sources. Pasting client details into a general consumer chatbot is the risk to avoid.
How long should it take to review AI output?
Less time than doing the work yourself, by a wide margin. If checking takes as long as producing, either the task is a poor fit or the output is not giving you sources and citations to check against. That is a signal to change the tool or the task, not to skip the review.
Start with one job. Check it. Keep what you trust.
Capiro is the AI back office for lobbyists. Meri, its AI Chief of Staff, reads, drafts and follows through from each client’s own context, with sources on every finding and nothing sent without your approval. More capacity for every client.
The category, explained: What is an AI back office for lobbyists?
