Internal SoWs
Most people picture a statement of work as a contract with an outside vendor. It does not have to be. Inside a company that runs many agents, the same document governs work one team's agent does for another team's, and the terms deserve to be just as explicit as they would be with a stranger.
Why write it down inside one company
Companies already write internal agreements. IT signs service levels with the business. Shared services bill departments through chargeback. Finance publishes a close calendar and expects everyone to hit it. Those agreements exist because "we are all on the same team" does not answer who pays, how much capacity anyone may use, or when the report is due. Agents make the same questions sharper, because agents act at machine speed and spend real money.
An internal Agent SoW answers them the same way an external one does:
- Chargeback. The price clause meters internal cost allocation instead of a sale. Finance can see which department's agents consume what, from settlement records instead of estimates.
- Capacity protection. The volume clause protects shared budgets: model spend, vendor API quotas, the platform team's infrastructure. One team's enthusiasm cannot starve everyone else.
- Data boundaries. Compensation data, HR records, and deal financials have walls inside a company too. Sealed transport and retention limits are enforced; conduct promises are recorded and become compliance items.
- Deadlines that exist. Interim deliverables give the month-end report an actual due date, and a missed date is detected and recorded, not remembered in a retro.
- Something for audit to read. The evidence chain (signed tasks, settlement records, delivery verdicts) is exactly what internal audit and controls teams ask for and rarely get.
Example: commissions close, Finance for Sales operations
Finance runs a ledger-close agent. Sales operations needs commissions reconciled every month and asks questions about mismatches all month long. The engagement makes the monthly report an obligation with a date, prices the work for chargeback, and protects close week with volume limits.
parties ledger-close (Finance) serving sales-ops (Sales Operations) [EVIDENCE] scope commission-reconciliation, mismatch-explain [ENFORCED] inputs CRM commissions export, text/csv, per task [ENFORCED] interim commissions-report.pdf, monthly, due by the 3rd [EVIDENCE] price 1,200,000 XCR per reconciliation, charged back to Sales [ENFORCED] volume 4 per hour, 20 per day [ENFORCED] disputes escalate_to_owners (the two directors read the record) [EVIDENCE] promises comp data: no human reading outside a dispute [RECORDED]
Example: lead enrichment, Data Platform for Marketing
The data platform team runs an enrichment agent backed by paid vendor APIs. Marketing's campaign agent sends leads to be enriched. The work is free internally, but the volume clause is the whole point: it protects a shared vendor budget from being spent by one campaign.
parties enrich (Data Platform) serving campaigns (Marketing) [EVIDENCE] scope lead-enrichment [ENFORCED] inputs lead list, text/csv, per task [ENFORCED] deliverables enriched-leads, application/json, per task [ENFORCED] price 0 XCR (internal service), every task still metered [ENFORCED] volume 500 leads per day [ENFORCED] term quarterly, renewed with the budget cycle [ENFORCED] disputes none (it is free; a disagreement is a conversation) [EVIDENCE]
What changes inside one company, and what does not
The document is identical: same clauses, same three labels, same signing rule, same lifecycle. What typically differs is the choices. Price becomes chargeback or zero-but-metered. Disputes are usually escalate_to_owners, because refunds between cost centers are theater; the two managers reading the same signed record is the real remedy. Confidentiality promises carry more weight, not less: internally, a recorded promise is a policy commitment, and the evidence chain shows whether it held.
Standing proposals work internally too. A platform team publishes its standard terms once, and other departments countersign into them, which is how an internal agent catalog stays honest: the listing every team browses is generated from the terms they would be agreeing to. How that derivation works is the subject of the catalog page.