CRM, RevOps & Sales Ops: Which LLM for Which Job (Frontier, Light, Low-Cost)

Model map

Seven models cover every job in this guide. Prices are indicative API prices in USD per million tokens (input / output).

DeepSeek Flash also covers many mid-heavy jobs at Flash pricing, especially in batch, flexible or off-peak mode.

Sales Ops broken down into jobs

Job → model matrix

For each job: the CRM trigger, the recommended model, the low-cost alternative and the write policy.

Hands-on playbooks

Lead intake: normalize + soft enrich

Inbound form or CSV: first names in ALL CAPS, free-text countries, misspelled company names.

• Inputs: email, firstname, lastname, company, jobtitle, country, utm_source, form_id.
• Output: JSON with firstname_clean, lastname_clean, company_normalized, country_iso2, seniority (enum), enrich_confidence (0–1).
• Model: GPT-6 Luna, for volume, short extraction and low cost. Low-cost alternative: GLM-5.3-Flash, if your PII policy allows it.

Steps

1. CRM webhook → worker with the lead payload.
2. Prompt: normalize, map country → ISO2, map seniority to an enum (IC|Manager|Director|VP|C-level|Unknown).
3. Validate the output against a JSON Schema; reject if email is invalid.
4. Write soft fields only; don't touch lifecycle_stage or owner_id.
5. If enrich_confidence < 0.6, send the lead to a needs_review queue.

• Write policy: soft fields only, plus an ai_normalize_ts timestamp.
• Failure mode: wrong country → wrong territory. Validate ISO2 against an allowlist, otherwise write Unknown.

ICP / persona classification + territory routing

After soft enrichment, classify ICP fit and route the lead.

• Inputs: company_normalized, employee_count, industry, jobtitle, seniority, country_iso2, arr_band (if known).
• Output: icp_fit (A|B|C|Out), persona (e.g. RevOps|CRO|AE|Other), territory_code, routing_reason (one sentence).
• Model: Luna or GLM Flash. It's a closed classification, no long reasoning needed. Low-cost alternative: DeepSeek Flash in batch.

Steps

1. Load the ICP rules (YAML or Notion) into the prompt, 2K tokens max.
2. Classify icp_fit and persona with strict enums.
3. Territory: a table lookup after the model (country_iso2 + employee_count band). Don't use the LLM for pure geography.
4. Owner: round-robin or capacity rules in the CRM workflow, not in the LLM.
5. Log routing_reason for audit.

• Write policy: write icp_fit, persona, territory_code; the owner is set by native automation.
• Failure mode: a persona outside the enum. Reject it, set Out, alert RevOps on Slack.

Outbound first-line personalization

Sequence step 1: one personalized line, not an essay.

• Inputs: firstname, company, persona, trigger_event (funding, hire, tech), pain_hypothesis, offer_one_liner.
• Output: first_line (180 characters max), the personalization_tokens used, risk_flags (e.g. overclaim).
• Model: Luna, for short high-volume drafts. Low-cost alternative: GLM or DeepSeek Flash.

Steps

1. Pull the trigger from enrichment. Never invent a funding round.
2. Prompt for one line, with an explicit "no invented numbers" constraint.
3. Regex filter: no competitor URLs, no "AI-powered" spam.
4. Write to sequence_custom_field or a merge tag.
5. The AE sees a preview and sends manually, or it goes out untouched when the confidence score is high.

• Write policy: draft field only; never auto-send without an opt-in process.
• Failure mode: a fake trigger burns the domain. If trigger_event is empty, fall back to a generic first-line template.

Company enrichment batch (1k–10k rows)

A TAM list, or accounts missing firmographics.

• Inputs: a CSV batch with domain, company_name_raw, and optionally scraped site text.
• Output: per row, industry_naics_guess, employee_band, hq_country, tech_hints[], confidence.
• Model: DeepSeek Flash, for price, long context and off-peak batch. Luna if you're on an OpenAI-only stack. Low-cost alternative: Luna (more expensive at volume) or GLM Flash.

Steps

1. Chunk 50–100 domains per call with a structured extraction prompt.
2. Use off-peak or flexible mode if DeepSeek offers it.
3. Write to a staging table (enrich_staging), not directly to the CRM.
4. QA a 50-row sample, by a human or a frontier-model spot check.
5. Upsert to the CRM on domain match; skip rows below the confidence threshold.

• Write policy: staging → sample → CRM; never overwrite ARR or owner.
• Failure mode: a parked domain produces a garbage industry. Require confidence and a source URL field.

CRM dedup / merge proposals

Duplicate contacts or companies: email variants, domain variants.

• Inputs: candidate pairs record_a / record_b, with key fields and activity counts.
• Output: same_entity (bool), merge_winner_id, a field_resolution map, confidence, human_summary.
• Model: Sol, Grok 4.7 or MiMo, for multi-field reasoning and edge cases. Low-cost alternative: DeepSeek Flash, proposals only.

Steps

1. Compute blocking keys (email, domain, LinkedIn) in the ETL.
2. Send only ambiguous pairs to the frontier model.
3. Write a proposal object to a dedupe_queue.
4. RevOps reviews in a UI: approve, reject or change the winner.
5. Merge through the CRM API after a human click.

• Write policy: proposal only, zero auto-merge.
• Failure mode: merging the wrong contacts loses data. Use soft-delete, an audit log and a documented rollback playbook.

Deal research memo

Opportunity at SQL or demo stage, for AE / SE prep.

• Inputs: opportunity fields (amount, stage, close_date, competitors_mentioned), the account website, the last 5 emails or notes, a call transcript excerpt.
• Output: a Markdown memo with context, stakeholders, risks, discovery questions and the competitive angle. Draft only.
• Model: Opus 5.5 (long, careful), Grok 4.7 (web and X) or Sol (tools). Low-cost alternative: MiMo-V2.6-Pro, if your policy allows Chinese providers.

Steps

1. Assemble a packet that fits the context window; trim the transcripts.
2. Use web search tools with Grok or Sol for public news.
3. Output to Opportunity.Note or a linked Notion page.
4. Tag it ai_draft=true.
5. The AE edits it; nothing syncs to amount or forecast_category.

• Write policy: draft note only; no writes to financial fields.
• Failure mode: an invented competitive win. Require citation URLs, or "unknown".

Weekly pipeline / forecast narrative (CRO)

Monday: a pipeline snapshot with commentary.

• Inputs: an export of open opportunities by stage, week-over-week delta, commit / best case, slipped deals, top 10 by amount.
• Output: a 400–700-word narrative, risk bullets, questions for managers.
• Model: Opus 5.5 or Grok 4.7, for careful synthesis without hallucinated numbers. Low-cost alternative: Sol.

Steps

1. Numbers are pre-computed in SQL or BI; the LLM doesn't recompute totals.
2. Prompt: comment only on the deltas provided.
3. Draft to Slack #revops and email to the CRO.
4. Human approval before it goes to any wider channel.
5. Archive the version and a hash of the inputs.

• Write policy: draft channel, human approval.
• Failure mode: the LLM "corrects" a total. Forbid calculation and inject the numbers verbatim.

RevOps agent: stage hygiene + missing next step

An opportunity with no next step, or stuck in a stage for more than N days.

• Inputs: stage, stage_entered_at, next_step, next_step_date, amount, owner_id, activity over the last 14 days.
• Output: hygiene_flag, suggested_next_step, suggested_task_due, severity.
• Model: Sol or MiMo, as a scoped agent with CRM tools. Low-cost alternative: DeepSeek Flash, flag and suggest only, no writes.

Steps

1. Query the stale set every night.
2. The model proposes task text and a due date.
3. Create a Task linked to the opportunity and ping the owner on Slack.
4. Never advance stage automatically.
5. Cap writes at N per owner per day.

• Write policy: scoped writes, a task and a flag field only; no stage skip, no amount edit.
• Failure mode: task spam. Rate-limit and send a daily digest instead of one task per opportunity.

Matrix template

Before go-live, fill in one row per automation with these columns: job, CRM trigger, model, low-cost alternative, input fields, output schema, write policy, failure mode, owner.

A row without a write policy doesn't go to production.

Go-live checklist

• Legal / security: explicit approval if GLM, DeepSeek or MiMo receive PII that leaves the EU, or mask it.
• Logging: prompt version, model id, record_id, latency, cost, output hash.
• Eval set: 50 leads (or 50 opportunities) with gold labels; precision / recall on critical enums.
• Cost alert: a daily budget, plus an alert when spend spikes above 2× baseline.
• Write scopes: least-privilege API keys, no delete unless needed.
• Rollback: a documented way to undo a merge or a task spam.
• Pricing: recheck vendor pricing between September 2026 and your production date.
• Human approval: on merges, forecast publishing, and anything from a deal memo that touches financial CRM fields.

Conclusion

Seven rules cover most routing decisions:

1. Volume + short schema (extract, classify, first line) → Luna, GLM Flash or DeepSeek Flash.
2. Batches of 1k+ rows, off-peak → DeepSeek Flash, or Luna if you're locked into one vendor.
3. Merges, CRM agents, tools → Sol, Grok 4.7 or MiMo, policy permitting.
4. Long memos, forecast narratives → Opus 5.5 or Grok 4.7, as a draft with human approval.
5. PII + EU → no Chinese provider without legal approval; otherwise anonymize.
6. Any destructive CRM write (merge, stage, amount) → human in the loop.
7. Numbers → computed outside the LLM. The LLM only comments on them.