Potential 2028 candidates and soon-to-be campaign managers
From
Josh Hendler
Subject
How to think about AI use on your campaign
Among the many gnarly decisions you’ll make when booting up a presidential campaign, the role AI should play will be a prominent moving target. Between breakneck capability growth, shifting public opinion, and an increasingly clear risk profile, predicting even a few months out is a fool’s errand. Then again, so is running for president.
I’ve worked on campaigns through three big technology shifts. In 2004, the web was turning online fundraising and organizing from a novelty into a core function. In 2008, I worked on Obama’s first run as social media changed how campaigns found and mobilized supporters. And I was technology chief at the DNC during President Obama’s first term, when Twitter started setting the news cycle in real time. Each time, the technology showed up well before campaigns knew what to do with it, with a real first mover advantage. This might be the biggest technological shift yet.
This memo is intended as a directional guide as of October 2026, not an exact roadmap. It is focused primarily on AI and the practice of campaigning. Policy and public opinion are out of scope (for public opinion research, see the Center for Shared AI Prosperity). It’s written for candidates, campaign managers, finance directors, field staff, and everyone in between.
For those running for President
We can’t unring the bell of transformational AI. Even with zero new model releases, what is already possible but not yet adopted, the technical overhang, will change campaigns forever.
If you only take away a single piece of advice from this memo: hire a campaign manager curious about AI who actually uses it.
AI will have profound implications for how events play out over the coming 24 months. Don’t give someone the wheel who lacks an even rudimentary understanding of what this technology is capable of.
What is already out of your hands
Maybe the right question isn’t so much, “what will you do with AI?” but rather “what will AI do with your campaign?” Reaching more into speculative territory, it seems perfectly plausible that in late 2027 and early 2028, we’re living in the following world:
Information agents
For a large segment of younger Americans, AI intermediates civic and political information consumption (e.g. a personal assistant agent like Meta’s Muse) gatekeeps information from emails, group postings, text messages, major news publications.
MemeticsSwarms of AI agents, either rogue or intentionally deployed, play a prominent role in the information ecosystem, influencing both other AI agents and social media content.
So whatBuild early warning systems for narratives related to your campaign and key issues.
SaaS-pocalypse
Information workers migrate to using frontier AI as their primary surface area for doing work, with complex user interfaces going the way of the dodo.
So whatBefore you pay for a new software tool, test whether the AI your staff already uses can do the same job.
New cyber threats
Presidential campaigns will be an even more appealing cyber target now that individuals with a model will have similar capabilities to what only coordinated state-actors could do not very long ago.
So whatHire a CISO who understands today’s threat environment and the potential for AI.
How AI can be most helpful to a campaign
Prioritizing cost and operational efficiency as a mental model for how AI can support a campaign is completely reasonable. Presidential primaries are, after all, a marathon and not a sprint, and caloric consumption matters. But my recommendation is to put it aside for the moment. The 5-10% cost efficiencies matter, but overly focusing on them misses the forest for the trees.
Focus instead on how AI can enable and accelerate effective learning loops: the continuous process of hypothesizing, testing and learning. This matters for three reasons:
As today’s information environment has evolved, many of the usual ways of reaching voters have faltered. It’s just harder to get in front of voters, made worse by continued declines in institutional trust.
Tactical innovations matter way more in presidential primaries. At least on the Democratic side, there will likely be a very crowded field, with harsh realities around fundraising and attention. Tactics alone are insufficient, but the innovation lane is very much open.
This is actually the place where campaigns can find real exponential gains in efficiency, get the flywheel running and get out of the way.
The Loop
Learning loops are a critical part of how today’s frontier models are trained, and AI can be a powerful accelerant. Think of the campaign itself as the “outer loop”, with many loops running inside of it:
The Campaign Loop Everything the campaign learns feeds the database of record and knowledge base, so each day’s listening starts with more than the last. Select a step to jump to its section.
Here’s an example of how this can all play out. Much of it runs on its own, with people stepping in to decide. Step through it:
Illustration. The messages, names and numbers below are invented to show how the loop could work. This is not a demo of a real product.
The campaign keeps a constant read on what people are seeing and talking about. AI does the reading and sorting that staff used to do by hand: polls, news coverage, social media, and the campaign’s own numbers.
What is the political press focusing on today?
How are LLMs answering questions about the race?
What are the weirdest new trends on TikTok and Instagram?
What are voters saying inside bellwether communities in early states?
What are our opponents saying, and what have they changed on their website?
#comms-dailythe agent posts here on its own
Campaign agent 6:02 amUnprompted
Morning brief. Three themes moved overnight:
Child care costs. Now the most-mentioned topic in door notes from Iowa and New Hampshire, up from fourth place last week.
A rival’s town hall answer on drug prices is spreading on TikTok and Instagram, mostly among voters under 35.
The political press is focused on debate qualification rules, not issues.
Read 1,240 door notes · 18,300 posts · 64 articles
Comms director 7:15 am
Child care is the one. Draft narratives on it for today, and keep an eye on the drug prices clip.
AI drafts many ideas, and the strongest rise to the top. It writes dozens of possible messages and hundreds of versions of ads, videos, and posts, then narrows them down using past results and quick tests with real people.
Humans review and decide what to focus on today, using AI drafts as a starting point.
Staff can see why a draft was picked, such as “this one tested well with voters in Iowa.”
#comms-dailythe agent posts here on its own
Campaign agent 8:40 am
I drafted 24 narratives on child care and tested 6 with the panel and the predictive model. The top three:
Narrative
Predicted lift
“Child care is the new mortgage”
+4.1
“Parents shouldn’t have to choose”
+3.2
“The rival’s record on care”
+1.4
Red-team note. “New mortgage” invites a fact-check on cost figures. Use the state-level number, with a source.
Ready for your sign-off.
Comms director 9:05 am
Approved. Go with the state number and the source. Nothing goes public without a human sign-off.
The content is deployed through many channels, personalized to each audience when possible. Volunteers, aligned creators, and surrogates get tailored guidance for what to share.
Guidance is personalized to each person based on geography, their social networks, and the issues their friends and networks care about.
A volunteer in a small town gets a different message than a surrogate with a large online following.
Organizing staff get a first-draft message or volunteer recruitment pitch tailored for each group of supporters.
#digitalthe agent posts here on its own
Campaign agent 10:30 am
The approved narrative is ready to go out:
Clip drafted, 30 seconds, with two suggested edits.
Guidance ready for 412 super-volunteers in 9 counties, each matched to the networks they told us they are part of.
Three donor email versions by giving history.
Sample guidance
For Maria, Ankeny: Share the clip in your parents’ group chat. Lead with the state cost figure and your own story about finding care.
Should I send the volunteer guidance now?
Digital director 10:48 am
Send the volunteer guidance. I’ll take the first edit on the clip, then post it by noon.
What happens next feeds back into the plan. The campaign tracks which messages worked and which didn’t, in numbers and in what people said.
Which ads and posts got attention?
What did volunteers hear at the door or on the phone?
How did group chats react to different messages?
Each day’s lessons shape what the campaign tries the next morning.
#comms-dailythe agent posts here on its own
Campaign agent Next day, 6:01 amUnprompted
Results from yesterday’s child care push:
The clip was watched to the end by 61% of viewers, against a 38% average.
Volunteers report 1,900 group-chat shares and 214 replies, most of them about cost.
37 new volunteer sign-ups came from shared links.
narrative child care is the new mortgage
status tested · approved · deployed
result above average, keep testing
saved to database of record
Today’s morning brief already builds on this.
Analytics lead 7:20 am
Nice. Test it against “Parents shouldn’t have to choose” this week, and split the results by age.
1Foundation
Most of what follows depends on a few basics being in place from the first day. (A day-one checklist is in the box at the end of this section.)
AStand up one AI stack on day one, and give it to everyone.
Early on in the campaign, pick either ChatGPT or Claude and set up a business or teams account for every staffer. Make sure no one uses a personal account. This is the most important AI precaution a campaign can take.
Connect it to Google Workspace so that agents with permissions can read the campaign's docs, email and calendar (with customizability per user / level of sensitivity).
Offer the latest commercial tools (e.g. Claude Cowork), not half-baked political tools.
Slack, or other chat platforms, are good homes for campaign agents that interact with multiple teams. This not only puts AI where people already work, but it gives you optionality to swap in cheaper models to save costs over time. Campaign agents should be experts on the state of the effort, with access to the knowledge base and as many data systems as is practical. Another route to consider is making Slack the primary way that many campaign staff access AI in the first place.
BInvest in a robust, modern database of record that agents can query through MCP.
Getting value from AI requires one database of record that goes well beyond voters, with data pipes that are not lossy. This is where everything the campaign learns ends up.
This should be adjacent to a way of storing qualitative campaign information like a knowledge base.
CBuild with skills and agents first, custom apps last.
Ship the first workflows in the first week: campaign clips, tracking inbound press emails, recurring drafts (press emails, schedules, daily updates), and one shared home for documents and assets. Let users build their own agents with skills and plug-ins before anyone reaches for Claude Code or Codex to build a new application that needs ongoing maintenance (or even purchases a new “AI” SaaS app).
DMake AI use a hiring requirement, and hire AI-forward operations leaders.
Using AI, and being willing to deploy it in service of winning, should be a hard requirement for hiring anyone on the campaign. Period. Not everyone will be at the same level, but there should be a floor. The biggest efficiencies sit across operational and administrative teams, so AI-forward leadership there matters most.
EMake the CTO the AI chief, and hire a CISO very early.
The Chief Technology Officer should also be the innovation chief and chief AI advocate. They should report to the Campaign Manager.
The Chief Information Security Officer should own cybersecurity, IT, identity and access, vendor risk, and AI security as one function. They should report to the CTO.
FDeploy agentic pods, and shift headcount from data to analytics.
Have a group of engineers partner with multidisciplinary teams across the campaign to build agents for their workloads (see Uber's agentic pods model). A lot of traditional "political data" work, like building a dashboard from the voter file, should be far quicker and of higher quality now. Analytics teams should be more productive and less held back by data hygiene problems.
GEvaluate what SaaS tools you actually need.
On the one hand, campaigns can waste valuable time re-building tools that already exist. But there are also clear opportunities to save money and gain efficiencies by not having hundreds of different SaaS tools. The perfect example here: there is virtually no reason that campaigns should be investing in a data visualization platform.
2Listen
ACapture every qualitative signal, and turn it into themes.
2028 campaigns must collect as much qualitative data as possible from every interaction with every person, from a voice note after a conversation at the door to a volunteer's feedback after a relational conversation. A machine should take all of it and surface what's being heard. This is not a replacement for polling, but it gives "on the ground" insight you wouldn't ordinarily get.
Pair it with frontier models as a source of community knowledge: baked-in knowledge plus deep web research can map the gathering places and local voices that matter in physical and online communities.
BRun one monitoring machine for social, oppo and GEO.
Social: with a social listening provider, get a daily feed of relevant posts, and use AI to transcribe them and see what gets traction.
Oppo: an AI should read every statement, transcribe every podcast and video, and audit every website edit from every other campaign and their surrogates, all day. It looks for unrealized opportunities and for tactics the campaign can adopt.
GEO: track and optimize how AI answers describe your candidate, campaign, opponents and surrogates. There is a set of clear best practices, but they will change many times. Today ChatGPT and Google summaries matter most, but agents like @grok play an outsized role in setting the narrative on X.
This should run on its own and follow events as they develop. Some of what it produces will be content, even if only panel participants see it. For now, require meaningful human review for anything public.
3Decide
ATest 100% of voter-facing content, inside the workflow.
Build testing into how teams create content, not as a separate step. The pipeline decides whether to run a live panel test or use predictive AI testing, and suggests changes based on past performance. Every staffer should also have real-time messaging guidance drawn from all current and historical polling. This is a solved problem.
BRed-team the campaign constantly.
Use models to keep asking, "what are we missing?" What exposure does the campaign have that no one has exploited yet? This is one of the better defenses against groupthink.
4Act
APower up volunteers, and stop pre-cutting turf.
Super-volunteers should get the same real-time message guidance as staff, plus tools to organize across their offline and online networks. They should be able to look up, instantly and with sources, what the candidate has said on any topic. There's some risk around data access and autonomy, but models can help supervise volunteer activity and data use.
Turf should be optimized in the moment, based on where each volunteer is and what needs doing. Swing Left's People app works mostly this way today. At a minimum, buy an auto-cutting tool. Better, pair it with distributed tools where volunteers start work on their own and field offices act as volunteer support centers.
Build deeper power maps of regions. Find the people who hold local cultural influence, build relationships with them, and help them become online spokespeople.
BBuild a rapid-response content engine.
Teams should react faster and harder to changing narratives, with AI finding stories, producing first drafts, and suggesting where to push them. Models should be constantly testing, clipping and ideating on new content. Even without a fully automated pipeline, automated clipping is a big opportunity for extra reach.
CPersonalize donor asks and find new prospects.
Even in 2026, we've seen considerable use of AI to personalize small- and high-dollar asks (based on both the candidate and the prospect) and to identify new donors.
Feed it back in
Everything above produces data: test results, door notes, clip performance, donor responses, oppo finds. All of it should flow back into the database of record, so each day’s listening starts smarter than the last. That's the self-learning campaign.
Where this all goes
This memo will likely seem like a dinosaur in six months. Campaigns have never had the luxury of a fixed technical or information ecosystem over the course of a cycle, and they especially won’t over the next two years. That takes us back to where we started: hiring a campaign manager, and a full team, who actually meaningfully uses and engages with this technology, and orienting these efforts towards more efficient testing and learning. Campaigns have the opportunity to make this a norm, and put their entire teams to work solving problems, and becoming a campaign that takes lessons from its own efforts: a self-learning campaign.