The compensation copilot: deploy business AI for pay questions without creating HR risk

If you want a “real” AI use case (not a demo), look at where people already ask for help. One signal this week: OpenAI highlights that workers send millions of daily messages asking about compensation and earnings. (source)

What can go wrong (and why “just add a bot” fails)

• Hallucinated numbers (the worst kind of error: confident + specific).
• Implicit promises (“you should be on $X” / “you’ll get a raise”).
• Privacy leaks (another employee’s pay, manager notes, performance info).
• Policy drift (the bot uses last year’s bands, or unofficial guidance).
• Compliance surprises across regions, awards, contracts, or internal rules.

The pattern: grounded answer + policy gate + draft-first

The safest version of this product is not “AI decides pay”. It’s “AI explains your organisation’s policy, shows the sources, and routes edge cases to humans”.

Inputs (what the copilot is allowed to use)

• Approved pay bands (role/level/location) with versioning and effective dates.
• FAQ/policy docs (bonus policy, review cycle, promotion guidelines).
• Public market data (optional) — but clearly labeled as external and non-binding.

Lane-based governance (three lanes that work)

• Green (general): explain policy, definitions, timelines, how bands work.
• Amber (personalised): employee asks “where do I sit?” → allow only if inputs are limited to their own data and pay bands.
• Red (sensitive): anything involving another person’s compensation, performance notes, negotiations, disciplinary items → refuse + route.

Escalation rules (make them mechanical)

Escalation is how you keep the bot helpful without letting it freestyle. Here are rules you can encode today:

• Numeric outputs require citations: any dollar figure must be traceable to a band row + effective date.
• Negotiation language: “counteroffer”, “threaten to leave”, “legal”, “discrimination” → route to HR/legal playbook.
• Uncertainty: model returns low confidence → route to a human response template.
• Policy mismatch: employee’s role/level not found in the band table → route to comp team.

Workflow blueprint (minimal viable implementation)

1. Classify the question (green/amber/red) + detect negotiation/compliance triggers.
2. Retrieve only approved sources (bands + policy + employee’s own record if allowed).
3. Generate a draft answer with:
   • Plain-English explanation
   • Quoted policy excerpts / citations
   • A “next steps” checklist
4. Run a verifier pass (can be small/fast): check for uncited numbers, promises, and privacy leaks.
5. Approval gate for amber/red outputs (or for any message being sent externally).
6. Log everything: inputs, retrieved docs, generated draft, verifier result, final send.

Metrics that tell you if it’s working

• Containment rate: % resolved in green lane without escalation.
• Escalation accuracy: % of escalations that humans agree were necessary.
• Policy drift incidents: uncited numbers or outdated bands (should trend to ~0).
• Time-to-answer: employee experience metric.

Freshness note: this post was prompted by items observed via OpenAI’s News RSS feed (e.g. “Equipping workers with insights about compensation”).

Where Workflow ADL fits

Workflow ADL is a “workflow-first” way to do AI: lanes, gates, tools, and audit logs. A compensation copilot is exactly the kind of system that should ship with guardrails by default.