Election offices face seasonal demand, exacting legal requirements, and little tolerance for incorrect information. AI can help with bounded administrative work and public information, yet the same system can create exclusion, security exposure, and loss of trust when it influences voter eligibility or election results. The useful question is where AI belongs in the electoral cycle, under which controls, and with whose authority.
What AI election management means
AI election management is the use of machine learning, optimization, language, speech, or computer vision systems by an election management body (EMB) to support election administration.
That scope covers several technology classes:
- Generative AI drafts, summarizes, translates, or answers questions in text or speech.
- Classification and matching models flag possible duplicate records, categorize documents, or detect anomalies.
- Forecasting and optimization models estimate demand, schedule staff, or suggest resource allocations.
- Computer vision and speech models transcribe meetings, read documents, or interpret images and audio.
These systems have very different failure modes. A scheduling suggestion can be reviewed before it affects anyone. A wrong answer about a registration deadline can prevent someone from voting. A false match in a voter roll can threaten an eligible voter’s rights.
The practical starting point is narrow, reversible work. A 2026 report from the National Conference of State Legislatures found that election administrators are using AI mainly for low-risk, verifiable tasks, including drafting communications, translating materials, and assisting with ballot proofing. Some offices are exploring retrieval-based assistants on websites and phone systems, though adoption is still early.
Election administration and campaigning require separate rules
Election administration is a public function. It includes voter registration, poll-worker operations, polling-place logistics, official voter information, tabulation procedures, audits, and records. Its goals include accuracy, access, security, neutrality, and public trust.
Political campaigns seek support, raise money, advertise, and mobilize voters. Their use of AI for persuasion, targeting, fundraising, or political content falls under a different set of election, communications, and privacy rules.
An official EMB system should never drift into persuasion. It should not recommend candidates, infer political preferences, tailor neutral voting information according to predicted ideology, or share campaign content. Keep the data, vendors, staff permissions, phone numbers, domains, and review processes for administration separate from campaign systems.
AI election use cases by risk tier
The right boundary depends on both the probability of failure and the harm that a failure could cause. International IDEA’s electoral management report maps potential applications across the pre-election, election, and post-election periods, along with their technical and human-rights risks. The tiers below describe operational risk. They are not statutory categories.
| Risk tier | Example workflows | Safe operating boundary |
|---|---|---|
| Lower | First drafts of internal emails, formatting documents, meeting transcription, training outlines | No sensitive data in general-purpose tools. A staff member approves every output before use. |
| Moderate | Poll-worker scheduling, turnout scenarios for staffing, document classification, translation drafts | AI produces a recommendation or draft. Election staff apply statutory rules, resolve exceptions, and retain the original source. |
| High | Public website or phone answers, polling-place lookup, deadline reminders, incident triage | Answers come from approved, current sources. Critical facts use deterministic data or approved templates. Unsupported questions transfer to staff. |
| Very high | Voter-roll matching, eligibility review, signature or biometric matching, ballot interpretation, fraud detection | AI cannot make the final decision. Use requires explicit legal authority, documented performance across affected groups, an appeal path, and an auditable human decision. |
| Unacceptable for generative AI | Removing a voter, determining eligibility, adjudicating a ballot, certifying results, alleging fraud, or predicting how a person will vote | Keep generative AI out of the decision path. Follow the jurisdiction’s authorized election procedures and certified systems. |
This classification is stricter than a typical enterprise AI matrix because election harm is asymmetric. A useful suggestion saves staff time. A wrong decision can deny a right or cast doubt on a valid result.
Good first uses are reversible
The U.S. Election Assistance Commission has documented real office use cases for social media drafts, poll-worker scheduling, turnout forecasting, meeting transcription, and role-specific training materials. Each one leaves staff with time to inspect and correct the work.
Reversibility is a useful selection test. A strong first workflow has:
- a clear source document or data set;
- an output that a qualified employee can review quickly;
- no autonomous effect on voter status, ballots, or results;
- limited or redacted personal data;
- a manual process that remains available; and
- an observable measure such as time saved, error rate, or completed transfers.
Public voter information needs stronger controls
An AI assistant can make official information easier to reach by web or phone. It can answer supported questions, offer language options, and transfer complex cases to an election worker. Voice can help people who prefer calling or have difficulty navigating a website. It should complement accessible official channels rather than replace them.
This is still a high-risk workflow. The Election Assistance Commission warns that voting dates, hours, and locations demand a higher degree of accuracy than many general AI tools provide. A fluent answer is not evidence of a correct answer.
For that reason, a public assistant should distinguish three kinds of response:
- Deterministic facts: Dates, hours, addresses, and deadlines come directly from a versioned database or approved template. The language model does not invent or transform the underlying value.
- Bounded explanations: The assistant retrieves approved passages from current election materials and explains them within a defined scope.
- Human-only matters: Eligibility determinations, unusual legal circumstances, complaints, threats, accessibility accommodations, and record corrections transfer to authorized staff.
Where Dasha fits
Dasha helps technical teams build and run production voice AI agents through a managed runtime, REST APIs, and a web application, with telephony, integrations, testing, monitoring, and large-scale call execution.
For an EMB, the defensible Dasha use case is a bounded voter information line. The agent can receive a call, classify a supported question, retrieve an approved answer or call an official data service, read the result, and transfer exceptions to staff. The application layer should enforce the policy boundary. A prompt that says “answer only from official sources” is too weak on its own.
Dasha should not determine eligibility, authenticate a voter from biometrics, interpret a ballot, or make an allegation of fraud. Those functions carry different legal, evidentiary, and certification requirements.
A safe architecture for voter information
A production system needs controls around the model. A better prompt is insufficient. Use five layers.
1. An authoritative source registry
Give every approved source an owner, effective date, jurisdiction, language, and expiration rule. Separate stable explanations from changing facts such as deadlines, polling places, and emergency notices. When two sources conflict, the system stops answering that question until the content owner resolves the conflict.
2. A policy and routing layer
Define the intents the system may handle. Route each intent to a specific tool, template, or retrieval collection. Block candidate advocacy, election predictions, legal advice, and personal-status decisions. Require a human transfer or official callback path for every blocked or unsupported intent.
3. Deterministic tools for critical facts
Use an API or structured data lookup for dates, hours, and locations. Preserve the returned value through speech generation. Attach a source ID and version to the interaction log. A language model may introduce the result, but it should not calculate a deadline or infer a location from partial information.
Personalized lookups need additional controls. Minimize the information collected, authenticate through an approved process, limit what the agent can retrieve, redact logs, and set a retention period. If the caller cannot complete the approved authentication flow, transfer the call without exposing record data.
4. Human operations and a safe fallback
Staff need a queue, context for the transfer, and authority to correct the source registry. The service also needs a safe degraded mode. If a data source is stale, unavailable, or producing conflicts, play an approved message and connect the caller to staff. Keep the public website, staffed phone line, and other established channels available.
5. Traceability and change control
Record the source version, tool call, model version, prompt version, response category, transfer outcome, and staff correction. Redact personal data before analytics. Treat model, prompt, knowledge-base, and telephony changes as production releases with named owners and rollback plans.
Run a pilot with election-grade release gates
A narrow pilot should prove a complete operating loop before expanding scope. Start outside an active voting period with a small set of high-volume, low-ambiguity questions.
1. Create the approved answer set
Select a limited jurisdiction and question set. Assign an election official to each source. Include direct answers, acceptable paraphrases, required disclaimers, expiration rules, and escalation conditions.
2. Build an adversarial evaluation set
Include ordinary questions and hard cases: outdated dates, similar place names, conflicting documents, code switching, background noise, interruptions, ambiguous residency questions, prompt injection, candidate questions, and requests for personal records.
3. Set release gates before deployment
The pilot does not launch with any incorrect date, time, location, deadline, or eligibility statement in the approved evaluation set. Unsupported questions must abstain or transfer. Language variants need qualified human review. Security exercises must show that a caller cannot expand the agent’s scope, reveal hidden instructions, or retrieve another person’s data.
4. Measure the whole service
Track critical-fact errors, unsupported-answer rate, successful transfers, abandoned calls, source freshness, tool failures, latency, staff corrections, and repeat contacts. Break results down by intent and language. A single average can hide a severe failure in a smaller group.
5. Expand one capability at a time
Add a new jurisdiction, language, or data tool as its own controlled release. Keep a change log and repeat the evaluation suite. Pause expansion when content ownership, staffing, security, or accessibility cannot support the larger scope.
Governance that survives election day
The AI policy should name who can approve content, production changes, new data access, and emergency shutdown. It should also define:
- permitted and prohibited uses;
- data classification, retention, and deletion;
- accessibility and language review;
- vendor access and incident reporting;
- records retention for prompts, outputs, and decisions;
- performance review across affected groups;
- public disclosure of where AI is used;
- complaint, correction, and appeal routes; and
- continuity procedures when the AI service is unavailable.
Procurement belongs inside this governance model. Require vendors to document data flows, subprocessors, model changes, security controls, log access, deletion, portability, and incident response. The EMB must retain control of approved content and the ability to disable the service. A contract cannot transfer public accountability to a vendor.
AI election management earns its place when it makes a bounded process easier to operate without weakening rights, neutrality, or evidence. If an inbound voter information line is your first controlled workflow, evaluate Dasha with your authoritative sources, transfer rules, and release gates from day one.