What makes an AI system agentic.
A chat assistant waits for a question and answers it. An agent is given a goal and works toward it across several steps. It decides what to look at, uses tools to retrieve and write things, checks whether a step worked and chooses the next one. In government affairs the goals are ordinary: keep this issue covered, get this meeting ready, make sure the comment deadline is met.
Three capabilities separate an agentic system from a better search box. It runs on its own trigger, so it can notice a new notice or a schedule change without being asked. It reasons with context, meaning the organization’s positions, past meetings, commitments and relationships, not only the public text. And it can act in other tools: create a task, update a record, draft a document, prepare a message. Each of those actions is a place where someone has to decide how far the agent is trusted.
Every agent runs some version of the same loop. Move the slider to see which steps it handles and where it pauses for a person at each level of autonomy.
An animated loop of five steps: observe, reason with context, prepare, propose or act, and review. A slider sets the level of autonomy from one to five. At level one a person asks and the agent only answers. At level two it drafts on request. At level three it monitors and prepares on its own and waits for a person to choose the next step. At level four it runs internal steps on its own and pauses for a person's approval before anything external, which is where government affairs should stop. At level five it acts externally without review.
Five levels of autonomy in government affairs.
Autonomy is not one setting. It rises in steps, and each step changes what the agent does without being asked and what is left for a person.
Answers questions
You ask, and it answers from public sources and your own records. It does nothing until asked and keeps no work in motion. The person decides everything that follows.
EXAMPLE“What did we say about this rule in our 2024 comments, and what has changed since?”
Drafts on request
You ask for a work product and it prepares one: a meeting brief, a one-pager, a comment outline. It saves drafting time, but a person still has to notice the need, ask, edit and send.
EXAMPLE“Draft a one-pager on the hangar modernization request for Sen. Okafor’s staff.”
Monitors and prepares proactively
It watches sources and calendars on its own and prepares work before anyone asks, then proposes next steps for a person to choose and carry out.
EXAMPLEA markup is noticed overnight. By morning a brief on the provisions you track is waiting, with a proposed list of offices to call.
Runs multi-step workflows with approval gatesWHERE GA STOPS
It carries a workflow through its internal steps on its own, such as updating records, creating tasks, assembling drafts and scheduling internal reviews, and stops before anything goes outside the team.
EXAMPLEA meeting ends. Notes are logged to the office and issue, commitments become tasks, and the follow-up email waits in the attendee’s approval queue.
Acts without reviewPAST THE LINE
It sends, submits or publishes on its own judgment, and people learn what happened from the log. For external work in government affairs, this is past the line.
EXAMPLEAn agent emails a committee staffer the moment a bill moves, in the organization’s name, without anyone reading it first.
Where agentic government affairs should stop: approval before anything external.
The line is not about how capable the model is. It comes from what government affairs work is made of.
- Every external message speaks for someone. A note to an office, a comment in a docket, a coalition sign-on: each carries the organization’s or the client’s name, and the person accountable for it should read it first.
- Relationships are the asset. One mistimed or tone-deaf email to a staffer can cost more than every hour an agent saves. The person who holds the relationship decides when to use it.
- Some context is never written down. An office may be in a sensitive week. A position may have shifted on a call nobody logged. A person catches what the record does not show.
- Accountability stays with people. Disclosure filings and anything with legal weight remain the responsibility of the registrant or the organization. An agent can assemble the material; people review and file it.
Inside the line, agents should move quickly. Updating records, creating tasks, assembling drafts and flagging conflicts are internal and reversible. Requiring approval for each one recreates the copy and paste work agents were supposed to remove. That is why level four is the practical target: autonomy for internal steps, a person at every exit. Meri, the AI Chief of Staff inside the Capiro back office, works this way. It monitors, reasons with client memory, prepares the work and moves the internal steps, and nothing leaves the firm without a person approving it.
Designing supervision for an agent.
Approval gates are one part of supervision. The rest is deciding in advance what the agent watches, what it may touch and how people check its work.
Scope what each agent watches
Name the issues, clients or programs, the sources and the offices. An agent watching everything produces noise and reasons from the wrong context.
Write down internal and external actions
List what it may do on its own (log, tag, create tasks, draft), what needs approval (send, submit, publish, schedule with an outside party) and what it never does, such as filings or anything financial.
Name the approver for each exit
The person who was in the meeting approves the follow-up. The GA lead approves a plan that hands others deadlines. The approver sees the sources next to the draft.
Make it stop when unsure
Missing sources, conflicting positions or an ambiguous instruction should produce a question for a person, not a confident guess.
Review the runs, not only the outputs
Look at a sample of runs each week: what triggered them, what was read, what was proposed and what people changed. Repeated edits show what to fix.
Keep a pause for each workflow
Stopping one workflow, one client or one issue should not stop the rest. Pausing should be easy enough that people actually use it.
What a person should be able to see afterward.
Supervision depends on reconstructing what happened. For every run, the log should answer what started it, what the agent read, what it concluded and from which source, what it did on its own, what it proposed, and who approved or changed it. A finding without a source is an opinion, and in this field an unsourced claim in a brief is a liability.
A sample run log for an agent handling a new proposed rule. Four sources are listed: the Federal Register notice, an issue memo, a prior comment letter and a meeting log. Seven log entries follow: the agent observed the rule, reasoned about it, prepared a plan, created internal tasks, proposed outreach and waited for approval, a person approved the plan and held the outreach, and the agent logged the decision. Selecting a source highlights every entry that used it, and selecting an entry highlights its sources.
The trail also makes the work transferable. When someone asks in March why the organization held its outreach on a rule in November, the answer is in the log: the notice, the reasoning, the proposal and the person who decided.
Before giving any agent more autonomy, it is worth asking:
- For any output, can it show the source behind each statement?
- Can you list every action it takes without approval, and change that list yourself?
- What does it do when two sources disagree?
- Where do you see what it did overnight, and who reviews it?
- Can it be paused for one client or one issue without stopping everything else?
Frequently asked questions
What is agentic AI in government affairs?
AI that works toward a goal across several steps instead of answering one prompt. In government affairs that means an agent that notices a development, reasons about it with the organization’s or client’s context, prepares the work, proposes or takes the next step and records what it did, under rules people set.
What is the difference between an AI agent and an AI assistant?
An assistant responds when someone asks. An agent can start on its own trigger, such as a new Federal Register notice or a meeting ending, use other tools to retrieve and write things, and carry several steps in a row. That extra reach is why autonomy and supervision need to be designed, not assumed.
Should an AI agent ever contact a congressional office on its own?
Not in government affairs. A message to an office speaks for the organization or the client and draws on a relationship someone else holds. An agent can draft it, attach the sources and queue it, but the person accountable for the relationship should read and approve it before it is sent.
How do you audit what an AI agent did?
Through a run log that records, for each step, what triggered it, what the agent read, what it concluded and from which source, what it did on its own, what it proposed and who approved or changed it. Reviewing a sample of runs each week shows where the agent needs tighter scope or better instructions.
Can an agent update our records and create tasks without approval?
Yes, if the actions are internal, reversible and logged. Updating a stakeholder record, tagging a meeting, creating a task or drafting a document can run on their own. Requiring approval for each would recreate the busywork the agent was meant to remove.
Autonomy inside the team. A person at every exit.
Capiro is the AI back office for lobbyists and government affairs teams. Meri, the AI Chief of Staff inside the back office, monitors, prepares and moves the internal work with sources on every finding, and waits for your approval before anything leaves. AI should not replace the relationship. It should carry the work around it.
The category, explained: What is an AI back office for lobbyists?
