Rep-assisted sales AI observes available evidence, proposes a bounded action, lets a named rep review it, executes only within granted authority, and reconciles the result. “Human in the loop” is too vague unless the workflow says which human, at which state, with what evidence, and with authority over which destination.
This operating model draws on NIST’s voluntary generative-AI risk framework, Microsoft’s current assist-versus-execute guidance, and Gangly repository facts accessed August 9, 2026. It does not establish legal compliance or performance lift for a particular product.
The short answer
Design rep assistance as five explicit states: observe, propose, review, execute, and reconcile. The rep should see the evidence and exact proposed change before consequential outbound or CRM action. Approval must be scoped to that action, and every execution must return a durable receipt or visible error.
NIST’s Generative AI Profile is intended to help organizations incorporate trustworthiness into AI design, use, and evaluation. Microsoft’s adoption guidance says assistive systems support human decisions, while executing systems need explicit ownership, authority, lifecycle management, and risk response. A sales workflow can contain both modes, but each transition needs a contract.
Define rep assistance by state
Each state should have one input, owner, output, and failure path.
| State | System job | Required evidence |
|---|---|---|
| Observe | Read a permitted signal, CRM record, email thread, transcript, or calendar event | Source, time, identity, scope, permission |
| Propose | Create a priority, message, question, note, task, or field suggestion | Inputs used, uncertainty, exact output |
| Review | Let the rep accept, edit, reject, defer, or suppress | Diff, destination, consequences, policy |
| Execute | Perform the approved send or write under bounded credentials | Approver, idempotency key, destination state |
| Reconcile | Confirm success, retry, reject, correct, or roll back | Source and destination IDs, timestamp, result |
Skipping reconciliation is a common design failure. A success toast is not proof that the message reached the intended draft folder, the task attached to the correct opportunity, or a retry did not create a duplicate.
Assign authority at field level
A general “CRM access” permission is too broad for a controlled workflow. Define authority by object, field, action, precondition, and actor.
- A call summary may be proposed but not attached until account and opportunity identity are confirmed.
- A follow-up task may be created after approval, while ownership and due date remain editable.
- A deal stage may be suggested but require the rep or manager to confirm the stage-specific exit evidence.
- A close date may be proposed from an explicit buyer statement but never inferred from silence.
- An email may be written to a drafts folder, while send remains a separate human action.
- A suppressed contact, customer, competitor, legal hold, or restricted account must hard-stop the action.
Authority also expires. Revoke it when credentials change, a source becomes stale, a field definition changes, or error rates exceed the agreed gate.
Design a useful review packet
Fast approval depends on evidence density, not hiding detail. Show the smallest packet that lets a rep catch a material error:
- person, account, opportunity, owner, and destination;
- source links and observed times;
- the exact event or conversation evidence;
- the proposed output and a diff from current state;
- uncertainty, missing context, and expired content;
- the available actions: accept, edit, reject, defer, suppress, or escalate.
Do not make accept visually effortless while burying source or identity. Measure edits and rejections as learning signals, not rep resistance. Repeated formatting edits differ from repeated wrong-account corrections and should produce different fixes.
Gangly repository facts describe rep review before outreach sends and before post-call work syncs to CRM. The relevant product question is whether the configured Sales Workflow System exposes enough evidence and authority control for your risk level.
Reconcile every execution
Every proposed action must end in one explainable state. Use a source event ID, proposal ID, approval ID, execution ID, and destination ID. Record rejected-by-policy, rejected-by-rep, expired, executed, failed, retrying, corrected, rolled back, or unresolved.
Idempotency prevents a network retry from sending twice or creating duplicate tasks. Optimistic concurrency prevents an older AI proposal from overwriting a newer human change. Reconciliation reports should expose missing receipts, duplicate destinations, stale pending work, and actions created after access revocation.
The AI sales ethics guide explains accountability and buyer-impact boundaries; the workflow automation guide covers state and failure design. Rep assistance connects them at the actual decision.
Pilot the workflow safely
Start with one bounded action and a labeled holdout. For example: create a post-call summary and proposed next-step task for review, without enabling automatic CRM writes.
Include wrong and ambiguous contacts, two accounts on one call, missing transcript, stale CRM state, a newer human edit, suppressed records, no next step, conflicting dates, repeated events, revoked credentials, rate limits, and destination outage. Reviewers label the correct identity, facts, omissions, and allowed action before seeing system output.
Measure material-summary accuracy, identity accuracy, accepted without edit, accepted after edit, rejected, suppressed, review time, duplicate proposals, execution success, correction, rollback, and unresolved errors. Expand one authority boundary at a time only when the failure is observable and recovery works.
Decision record: workflow version, sources, fields, approvers, pilot population, raw counts, hard failures, operator minutes, allowed authority, rollback owner, and retest date.
Rep-assisted AI succeeds when the rep can make a better-grounded decision with less reconstruction—not when the interface produces more suggestions.