AI for Case Management: Human-Owned, Auditable Workflows

A case is more than a conversation. It is a durable record of an issue that may span people, documents, deadlines, decisions, and channels. AI can reduce the clerical work around that record, but the case management system remains the source of truth and an authorized person remains responsible for the outcome.

This distinction matters in customer complaints, insurance and benefits administration, legal services, healthcare, employee relations, public services, and other settings where a plausible summary can affect someone's rights or access to help.

Give the case a human owner before adding AI

Every case needs a named owner, a current status, a next action, a deadline when one applies, and a record of what supports each decision. AI can propose or populate those elements. It should not make responsibility disappear behind an automated queue.

Design the lifecycle first:

1. Intake: What starts a case, which facts are required, and how is urgent risk escalated?
2. Triage: Which team should receive it, what priority rules apply, and who can change them?
3. Investigation: Which records, interviews, policies, or external evidence are needed?
4. Decision: Who has authority, which standards apply, and what review is required?
5. Communication: What must be explained, through which channel, and in which accessible format?
6. Follow-up and closure: What proves the action happened, and what can reopen the case?

This map exposes where a model can assist and where deterministic rules or human judgment belong.

Use AI for bounded casework

The most defensible uses save time without transferring decision authority.

Structure an intake without treating it as evidence

An assistant can ask approved questions, identify missing fields, transcribe a call, and create a draft intake record. Preserve what the person actually said or submitted. A model-generated paraphrase is a convenience, not the original evidence.

Avoid collecting a full life history "just in case." Ask only for information the defined workflow needs. If the issue changes, explain why another question or permission is needed. For urgent, sensitive, or out-of-scope disclosures, stop the routine flow and route to the designated person or service.

Prepare a brief with citations to the record

AI can summarize long notes, identify dates, group related events, and list unanswered questions. The reviewer should be able to open the underlying call, document, email, or database field from each important statement. When the source is missing or conflicts with another record, the brief should mark the conflict instead of choosing the smoother narrative.

The NIST Generative AI Profile identifies confabulation, privacy leakage, harmful bias, and prompt injection among the risks of generative AI. Those risks apply directly to case files. A malicious instruction hidden in an uploaded document must not override access rules, and a confident synthesis must not erase uncertainty.

Retrieve guidance and draft routine communications

An assistant can find the current procedure, surface the relevant section, and draft a status update or request for missing information. Restrict retrieval to approved, versioned sources. The case owner should review communications that interpret policy, set expectations, deny a request, or affect a person's options.

Coordinate work without deciding the outcome

AI can suggest a queue, create a task, calculate a deadline from approved rules, or remind an owner about missing work. Use deterministic validation for dates, jurisdiction, service levels, and authorization. High priority should be triggered by explicit, tested rules and confirmed by a person when misclassification could cause harm.

Separate the record, the model, and the decision

Putting a model in front of a case database does not make the database safe to use. Build distinct controls around three layers.

The case record holds the accepted facts, original evidence, status, owner, actions, and decisions. Fields should record provenance: submitted by the case subject, retrieved from a system, extracted by AI, verified by a person, or decided by an authorized role.

The AI workspace receives only the context needed for the current task. It may produce a draft summary, suggested category, extracted fields, or proposed response. Those outputs need confidence or review state where useful, and they should not silently overwrite verified fields.

The decision process applies policy, professional judgment, and authority. Eligibility, entitlement, liability, discipline, benefits, financial redress, clinical action, legal strategy, and safety interventions should not be delegated to a general language model. Use human approval gates and record who made the decision, on which evidence, under which policy version.

This architecture also makes corrections possible. If a case subject disputes a fact, the team can locate the source, correct the record, regenerate dependent drafts, and see which decisions used the old information.

Limit access and retention by task

Case files often mix identity data, free-form allegations, health or financial information, internal notes, and third-party records. A single broad permission for "the AI" is too coarse.

• Restrict tools and retrieval by case type, team, role, and current task.
• Redact or omit fields that are irrelevant to the prompt.
• Keep secrets and privileged notes out of general-purpose knowledge indexes.
• Define whether recordings, transcripts, prompts, model outputs, and tool logs are stored, where, and for how long.
• Prevent one case's context from entering another case or a shared answer.
• Log reads and writes to sensitive records and review unusual access.
• Test deletion, correction, legal-hold, and subject-access processes across vendors and derived data.

The NIST Privacy Framework provides a risk-management structure for identifying and managing privacy risk, while the NIST AI Risk Management Framework organizes AI risk work around Govern, Map, Measure, and Manage. These are frameworks, not substitutes for the laws and professional duties that apply to a particular case program. Healthcare, legal, employment, insurance, education, government, and cross-border workflows each require domain review.

Worked example: a disputed service charge

Consider a customer who phones about a recurring charge they do not recognize. The AI-assisted path can reduce repetition without deciding whether money is owed.

1. The phone flow identifies itself, confirms the caller's request, and performs the organization's approved identity checks.
2. A read-only tool retrieves the relevant account and transaction records. The model does not receive unrelated account history.
3. The assistant records the caller's explanation in their own words and confirms dates, amounts, and the requested outcome.
4. If an approved explanation fully resolves a factual question, the assistant can provide it with the source and next step. If the charge is disputed, identity is uncertain, policy requires judgment, or the caller asks for a person, it opens or transfers the case.
5. The case system stores the original transcript or recording according to policy, the retrieved transaction identifiers, a draft summary, and the fields the caller confirmed. It labels the summary as AI-generated until a case worker accepts or edits it.
6. A case worker reviews the evidence, requests anything missing, and decides the resolution within their authority. Any refund, denial, or exception follows the organization's approval rules.
7. The final communication states the decision, reason, next step, and review or appeal route when one applies. The case closes only after the intended account action succeeds and required communication is recorded.

If the account API fails at step two, the assistant should not guess from the caller's description. If the transfer fails at step four, the caller needs an honest fallback and a case reference, not an endless loop.

Test records and outcomes, not polished summaries

Create an evaluation set from representative, properly handled cases. Remove or protect personal data as required. Include incomplete forms, conflicting records, multiple issues in one contact, unusual language, emotional callers, unsupported allegations, changed policies, malicious attachments, and cases that should never be handled without immediate human attention.

Score the workflow at field and case level:

• required-field accuracy and unsupported-field rate;
• traceability of summary statements to source records;
• routing accuracy, urgency false negatives, and inappropriate escalation;
• time to first human action and time to resolution;
• human edit rate for summaries and drafted messages;
• missed deadlines, failed tools, duplicate cases, and unauthorized write attempts;
• transfer success and context preserved at handoff;
• reopen, correction, complaint, and appeal rates; and
• differences in error or service rates across relevant case and user groups.

Do not optimize only for fewer human touches. A low-touch flow can hide abandoned cases, missing evidence, delayed escalation, or decisions that are faster but harder to challenge. Review samples over time and after changes to models, instructions, knowledge, tools, or policy.

Where Dasha fits in phone-based case intake

At Dasha, we build a managed voice AI runtime for technical teams. A Dasha voice workflow can use webhook-defined tools to retrieve approved data or create a case through the organization's APIs. Those APIs should enforce identity, field validation, access, idempotency, and write permissions; the conversation prompt is not an authorization layer.

Dasha supports warm, cold, and HTTP-based call transfers, which can move an out-of-scope or sensitive conversation to the correct human path. Call Inspector exposes transcripts, recordings when enabled, model inputs and outputs, tool calls, timing, and latency for review. Post-call analysis can populate configured labels from a transcript, but consequential case fields should still be validated before they drive action.

Dasha does not supply the case policy, professional judgment, or system of record. It is a fit when a technical team wants to build a custom phone intake or status workflow and control the business logic behind it. It is not a replacement for case management software or case owners.

Start with one case type, one intake path, and read-only access wherever possible. Make evidence, ownership, escalation, and correction visible from the first pilot. Then expand only when the case record, human handoff, and final business action remain accurate under real review. Technical teams can assess the runtime through Dasha's Voice AI Backend and product documentation.