Copilots vs background agents
Both are real, and they do different jobs. A copilot is an assistant: it makes the person using it faster. A background agent is a worker: it does the work whether or not anyone asks. Most enterprises end up needing both. The mistake is to expect one to do the other's job.
| Copilot / assistant | Background agent | |
|---|---|---|
| Starts when | A person prompts it | The work arrives: an invoice lands, a bank feed posts, a deadline nears |
| Helps | One person, one task at a time | The whole queue: matched, posted and reconciled before the team logs in |
| Output | Drafts, answers, summaries a human then acts on | Completed transactions in the system of record, with exceptions escalated |
| Measured by | Usage, adoption and satisfaction surveys | Event logs and the four numbers: share of work in the system, cycle time efficiency, first time right, cost to serve |
| Governance | The user's judgement on each output | Scoped permissions, confidence thresholds, human review queue, full audit trail |
| Cost profile | Per-seat licences plus tokens on every interaction | Mostly deterministic code; the model is called only where judgement is needed |
| Shows up in | Someone's morning | The close, the exception queue, the P&L |
of enterprise GenAI builds deliver no measurable return to the P&L MIT NANDA, 2025
of finance leaders who deployed AI report high impact from it Gartner survey of 183 finance leaders
Run both. Expect different things.
Keep the copilots. They are worth having. But if this year's goal is a shorter close and fewer exception hours, that number moves when background agents do the work: inside the systems you already run, with scoped permissions and an audit trail, and measured in your own event logs. Agents go where the work is already in the system and done right first time. Measure that first.