AI Drafts, Humans Send: How Response Drafting Should Work
Every inbound reply is a voter who chose to engage. The programs that convert those replies answer them fast, in a human voice, with facts the campaign stands behind. The programs that do not are staring at a queue of hundreds of unanswered conversations at 11pm.
Lincoln is the AI agent built into Political Comms. On broadcast projects, it drafts a reply to every inbound message and puts that draft in front of a staffer. What happens next is the point of this article, because the review step is where AI texting is either done right or done badly.
How a draft gets written
When a voter replies to a broadcast, Lincoln builds a draft from three sources, and only these three:
- The full conversation. The original outreach message and everything the voter and the campaign have said since.
- Every unanswered message. If the voter sent three texts since the campaign's last reply, the draft answers all three together. One reply, not a stack of fragments.
- Campaign context. Each project carries a goal, such as turnout, RSVPs, or volunteer signups, and a set of approved facts: poll hours, an event address, a donation link. This is the only factual material Lincoln is permitted to use.
That last constraint is the difference between drafting and guessing. Lincoln states facts that appear in the approved context or earlier in the thread. It does not produce a polling place it was never given or a link nobody vetted. When the context has no answer, neither does the draft.
Drafts are also held to the standard of a message a staffer would actually send. Before a draft ever reaches the review queue, code validates it: one SMS segment, plain language, and an outright rejection if it contains telltale AI phrasing, em dashes, smart quotes, or emoji. A draft that fails is dropped, not shown.
The review step is the feature
Every draft lands in a review queue with the voter's message quoted above it, so the staffer sees exactly what they are answering. Three actions are available:
| Action | What happens |
|---|---|
| Approve and send | The draft sends as-is through the normal reply path |
| Edit | The draft loads into the composer; the staffer adjusts it and sends it as their own message |
| Reject | The draft is dismissed and nothing sends |
There is no fourth option where the system decides on its own. A draft that nobody reviews sends to nobody. For high-volume days there is a project-wide review list of every pending draft, each shown with the inbound it answers, and an approve-all action for when the queue has been read and the drafts hold up. Even then, each draft sends exactly once, and each send traces back to a person who chose it.
Conversations also do not go stale under review. If the voter sends another message while a draft is pending, Lincoln revises that draft to cover the new message too. The staffer always reviews a current answer, never yesterday's.
What this looks like in practice
The River City Campaign runs a turnout broadcast to 40,000 voters on a Tuesday morning. By noon there are 1,800 replies: polling place questions, "who is this," a few donors asking where to give, and a share of hostility.
Without drafting, that queue is a day of staff time, and the slow answers convert worse than the fast ones. With Lincoln on, each of those 1,800 conversations already has a proposed answer built from the campaign's own approved facts. Staff work through the queue as editors: approve the clean ones, tighten the ones worth tightening, reject the ones that deserve a personal touch or no answer at all. Hostile threads get restraint, because backing off is part of how Lincoln drafts. The afternoon's work becomes judgment instead of typing.
The math favors the campaign that answers. Reply-handling capacity stops being the ceiling on how large a broadcast can responsibly be, because every reply arrives pre-drafted and grounded.
Why the human stays in the loop
Requiring review is not caution for its own sake. A text from the campaign is the campaign speaking, and someone has to own that. Review keeps accountability where it belongs, keeps compliance obligations attached to a person, and keeps the campaign's voice under the campaign's control. Drafting removes the typing. It does not remove the decision.
This is the workflow in full: Lincoln reads, Lincoln drafts, a staffer decides. It turns on per project with a single toggle and is off by default.
See it end to end on the Lincoln page, or talk to us about running it on your next broadcast.
Frequently Asked Questions
Can Lincoln send a reply without human approval?
No. Every draft waits in a review queue until a staffer approves it, edits it, or rejects it. There is no auto-send mode. A draft that is never reviewed never sends.
Can staff edit an AI-drafted reply before sending?
Yes, and editing is built into the workflow. Choosing Edit loads the draft into the composer, the staffer changes whatever they want, and the edited message sends as their own reply. Review and editing are required parts of the workflow, not optional extras.
Where does Lincoln get the facts in its drafts?
From two places only: the conversation itself and the campaign context the organization supplies, which includes the goal of the outreach and approved facts like poll hours, event details, or a donation link. Lincoln does not invent facts, links, or dates. If the approved context lacks an answer, the draft will not contain one.
Do AI-drafted replies sound like AI?
No. Every draft is validated in code before a staffer ever sees it: it must fit in a single SMS segment and it is rejected outright if it contains telltale AI phrasing, em dashes, smart quotes, or emoji. Drafts read like a staffer typed them, because anything that does not is thrown away.
What happens if a voter sends several messages before the campaign replies?
Lincoln revises the pending draft so one reply answers everything the voter has sent since the campaign's last message. There is one pending draft per conversation, kept current.