Key takeaways
- Separate observed fit evidence from engagement signals and missing fields; one unexplained score hides all three.
- Use deterministic rules for eligibility, territory, ownership and deduplication. Use model judgment only for bounded extraction from approved text.
- Qualification routes attention; it does not authorize claims, outreach, pricing, opportunity creation or a commitment to the buyer.
- Every routed record needs the criteria version, supporting evidence, unknowns, destination and a correction path a sales owner can audit.
A qualification record needs evidence, unknowns and a rules version
| Dimension | Acceptable evidence | Unknown behavior |
|---|---|---|
| Eligibility | Approved service area, account type or explicit exclusion | Hold or route to review |
| Fit | Observed company or use-case field with source | No points for absence |
| Need | Buyer-supplied problem in a form or conversation record | Do not invent intent |
| Engagement | Dated product, form or event signal under the approved policy | Label unavailable or stale |
| Ownership | Territory, account and queue rules | Default owned queue |
| Decision | Criteria version, matched rules and reason codes | Manual qualification route |
Run qualification from capture to owned next action
- 01Receive and identifyTriggerCreate one event ID and match the person, account and existing open record.
- 02Validate provenanceCodeCheck required fields, source, permission state, freshness and duplicate keys.
- 03Extract bounded evidenceAI judgmentMap approved free text to a fixed field or unknown, retaining the source span.
- 04Apply qualification rulesCodeEvaluate eligibility and fit through a versioned decision table.
- 05Assign deterministicallyCodeApply territory, account ownership, capacity and default-queue rules.
- 06Review exceptionsHuman approvalResolve conflicts, high-value ambiguous cases and correction requests.
- 07Choose the responseHuman approvalThe sales owner decides contact, timing, claims and commercial next step.
Do not let one score erase the reason for the route
| Bad shortcut | Better record | Why |
|---|---|---|
| Missing field equals zero fit | Unknown plus collection or review route | Absence is not negative evidence |
| Email click proves need | Engagement signal with source and date | Activity and fit answer different questions |
| Model says high intent | Extracted statement, source span and confidence/abstention | A label without evidence is not auditable |
| Score assigns a rep | Qualification result then deterministic ownership rules | Fit and routing are separate decisions |
Route ambiguity without losing the record
- Duplicate person or account: pause assignment, preserve both source events and resolve the canonical record.
- Conflicting firmographic fields: show each source and freshness; do not silently prefer the field that raises the score.
- No owner rule matches: assign an owned default queue rather than leaving the record unassigned.
- Criteria change: retain the prior decision and version; re-evaluate only under an explicit migration policy.
- Correction request: change the source field or rule outcome with actor, reason and timestamp, then rerun idempotently.
A high-engagement record remains unqualified
An inbound contact has several recent content events and asks about a service. The approved fit table requires a supported service region and a named operational use case. Region is present; the request describes only “AI help.”
The workflow records strong engagement, eligible region and unknown use case. It assigns the owned clarification queue with the criteria version and source text. It does not create an opportunity or send a generated claim. A rep asks the approved qualification question and owns the result.
Use the SOP and workflow blueprint as one versioned control
The SOP names the sales and revenue-operations responsibilities. The blueprint defines event identity, evidence fields, abstention, assignment, idempotency and tests. Replace the sample criteria with the organization's approved policy.
Lead qualification SOP
Editable procedure for validation, evidence-based qualification, deterministic assignment and correction.
Preview the fileHide preview
# Lead qualification SOP Version: 1.0 Owner: Revenue operations Effective date: [YYYY-MM-DD] Review cadence: [cadence and owner] ## Purpose and scope Validate, qualify and route one inbound lead record using approved evidence and versioned criteria. Qualification allocates internal attention; it does not authorize external outreach, claims, pricing, opportunity creation or a buyer commitment. ## Trigger, inputs and prerequisites - Trigger: approved inbound lead event with a unique event ID. - Inputs: person/account identifiers, source and permission state, declared request, fit fields, engagement events with dates, existing ownership and criteria version. - Prerequisites: approved eligibility/fit table, territory and assignment rules, default owned queue, correction path, data-retention policy and external-communication approval policy. ## Roles - Revenue operations: owns criteria, routing rules, versions and exceptions. - Marketing operations: owns source and permission evidence. - Sales owner: owns the commercial response and record corrections. - Data/privacy owner: approves fields, sources, retention and access where required. ## Procedure 1. Receive the event and deduplicate against person, account and open-record keys. 2. Validate source, permission state, freshness and required fields. 3. Separate eligibility, fit, stated need, engagement, unknowns and ownership fields. 4. If approved, extract a bounded field from free text with the exact source span; otherwise mark unknown. 5. Apply the versioned eligibility and fit decision table deterministically. 6. Apply territory, account ownership, queue and capacity rules separately from qualification. 7. Store reason codes, supporting evidence, unknowns, criteria version and destination. 8. Route conflicts and high-value ambiguity to the named review queue. 9. Let the assigned sales owner decide any contact or commercial next action. 10. Correct source fields or rules with actor/reason, then rerun without duplicating effects. ## Decisions, exceptions and approvals - Duplicate identity: hold assignment and resolve the canonical record. - Missing field: preserve `unknown`; never convert absence into negative fit. - Conflicting sources: show both source and freshness values for review. - No assignment match: use the owned default queue. - Criteria update: keep the original result/version; migrate only under an approved policy. - External communication, pricing, qualification override or opportunity creation: named sales approval. ## Output and completion Output one owned record containing qualification result, evidence, unknowns, reason codes, criteria version, assignment rule/version, destination, review state and correction path. Completion means the route is owned—not that a sale is likely. ## Measures Track duplicate rate, unknown-field rate, manual-review age, routing correction rate, unowned records, criteria-version distribution and false certainty found in review. Set targets from observed cases. ## Worked example (simulated) A record has recent content engagement and an eligible region but gives no supported operational use case. The workflow records engagement and region, leaves need unknown, and routes the owned clarification queue. It neither creates an opportunity nor sends a message. A rep asks the approved question and owns the update. ## Revision log | Version | Date | Change | Approver | | --- | --- | --- | --- | | 1.0 | [YYYY-MM-DD] | Adapted from public template | [name/role] |
Lead qualification workflow blueprint
Trigger, record contract, rules, bounded extraction, routing, retries and acceptance tests.
Preview the fileHide preview
# Lead qualification workflow blueprint
## Trigger and termination
- Trigger: `{event_id, person_key, source, permission_state, received_at}`.
- Success: qualification and deterministic assignment recorded with an owned destination.
- Stop: invalid source/permission, unresolved identity conflict, revoked criteria version or owner-issued stop.
## State model
`received -> validating -> duplicate_review | extracting -> qualifying -> assignment -> review | assigned`
Terminal alternatives: `rejected_source`, `stopped`, `superseded`. Corrections create a new evaluation linked to the prior one.
## Data contract
Keep eligibility, fit, stated need, engagement and ownership as separate objects. Every fact has `{value, source_id, observed_at, freshness_status}`. Extracted text also has `{source_span, extractor_version, confidence_or_abstain}`. Result has `{criteria_version, matched_rules, unknowns, reason_codes}`.
## Ordered steps and actors
1. **Trigger:** receive the lead event.
2. **Code:** deduplicate person/account/open records and validate provenance.
3. **AI (bounded):** extract allowed fields from approved text or abstain.
4. **Code:** apply eligibility and fit rules.
5. **Code:** apply territory, account, queue and capacity rules.
6. **Human:** resolve conflicts, ambiguity and correction requests.
7. **Human:** choose external and commercial next action.
8. **Code:** record final owned route and evidence packet.
## Idempotency, retry and rollback
- Keys: `capture:{event_id}`, `evaluation:{canonical_record_id}:{criteria_version}:{input_hash}`, `assignment:{evaluation_id}:{assignment_version}`.
- Retry enrichment/reads with bounded backoff; preserve unavailable rather than defaulting.
- Before retrying an assignment write, read the destination system by effect ID.
- Rollback reassigns through a corrective event; do not delete the original route.
## Permissions and data handling
Allow only approved source fields. Minimize retention and access. Do not place sensitive or undeclared attributes into model prompts, criteria, logs or test fixtures. Keep external communication outside this workflow unless separately approved.
## Exception routes
- Duplicate/conflicting identity -> record stewardship.
- Missing or stale source -> source owner.
- Extraction abstention -> qualification review.
- No territory/owner match -> default owned queue.
- Criteria dispute -> revenue-operations owner.
- No exception owner -> safe stop.
## Acceptance tests
1. Duplicate capture creates one evaluation.
2. Missing fit field remains unknown and does not subtract points.
3. Conflicting sources show both values and block automatic route where policy says so.
4. Unsupported free text returns abstain with its source span.
5. Qualification and assignment rules remain distinct in the record.
6. No assignment match reaches the default owned queue.
7. Retry does not create a second assignment effect.
8. Criteria correction preserves and links the prior evaluation.
## Audit record
Retain source events, canonical record decision, input/rule/model versions, extracted spans, unknowns, matched rules, assignment result, reviewers, corrections, retries and external effect IDs.
Limitations and when not to use this
- This playbook does not claim an all-agents connector to a CRM, enrichment, advertising, email, calendar or sales-engagement system.
- Qualification is an internal attention route, not proof of purchase intent, buyer authority, creditworthiness or likely revenue.
- Outreach, consent, privacy, discrimination and sector rules depend on the actual data and jurisdiction; this is not legal advice.
- The example is simulated and represents no conversion, response-time, pipeline or revenue outcome.
Sources
- Set Up Assignment Rules — Salesforce Help Accessed 10 August 2026
- Protecting Personal Information: A Guide for Business — Federal Trade Commission Accessed 10 August 2026
- AI Risk Management Framework Core — NIST Accessed 10 August 2026
Download the qualification SOP
Replace the sample fit dimensions and destinations with approved criteria, owners and evidence sources.
Download the qualification SOPUli Prantz
Builds and operates all-agents
Uli Prantz builds all-agents, the process-automation platform this site documents. He writes about the operational side of automating recurring business work: where deterministic code beats model judgment, where it does not, and where a human still has to approve.