Shadow mode: how agents earn autonomy
Shadow mode runs an agent beside your team, touching no real record, until a named person decides it has earned a gate. Checklist and exit criteria.
The short answer
To run an AI agent in shadow mode, have it do the real workflow alongside the person who does it today.
It reads the same inputs and proposes outputs, but nothing it produces reaches a record or a recipient. Mercury Digital records every difference, and a named gate holder decides from that record when the agent earns a gate.
Shadow mode
A stage of agent deployment where the agent runs the real workflow on real inputs next to the human doing it, with no ability to write, send or pay. Its outputs are recorded and compared with the human outcome. Autonomy opens only on a human decision made against that record.
Shadow mode vs other ways to test an agent
| Approach | Real inputs? | Can it affect records? | What it proves | What it doesn't prove |
|---|---|---|---|---|
| Demo / sandbox | No (sample data) | No | The agent can do the task on chosen examples | How it handles your actual inputs and exceptions |
| Shadow mode | Yes | No | How its proposals compare with your team's real outcomes | Inputs it never saw; what happens when it writes |
| Gated production | Yes | Only after a named person approves each item | Its proposals hold up under real approvals | Unattended behavior (by design) |
| Unattended for a step | Yes | Yes, for that step only | It performs within agreed limits | Steps it wasn't opened for |
How to run shadow mode: checklist
Pick one workflow and its consequential step
Capture the human baseline
Give the agent read-only access
Run in parallel on real items
Record every comparison
Write down the limits
Decide at the gate
What the shadow record can't tell you
Unseen inputs
It covers only the inputs that actually came through.
The human baseline may be wrong
Some mismatches are the person's error. That's a finding, not an agent failure.
Nothing about writes
Shadow mode never writes, so the first gated writes still need close attention.
Where this fits
In Mercury Digital deployments, every workflow starts in shadow mode and ships gated by default. Under Mercury's standard policy, the agent acts on its own only after it has run in shadow for at least 50 items of each type and at least 2 weeks, with 95%+ agreement with the human reviewer and zero critical errors. Autonomy is then opened gate by gate, on your decision. It's the Deployment phase of Audit → Design → Deployment → Governance. See How It Works.
Exit criteria (Mercury's standard policy, confirmed by the gate holder before shadow starts)
Shadow mode ends on a recorded decision, not a calendar date. Mercury's standard policy sets the minimums below; the gate holder confirms them, or sets stricter ones, before shadow starts, so the decision is made against them:
- Coverage: every input type and exception type the step routinely sees has come through at least 50 times, over at least 2 weeks of shadow running.
- Agreement: the agent's proposals match the human reviewer's decision on at least 95% of items, with every mismatch explained.
- Zero critical errors across the shadow run.
- Zero unexplained misses on the consequential field (amount, account, recipient).
- Rules updated: every mismatch caused by an incomplete written rule has had its rule fixed and re-tested.
- Limits stated: the record says in writing what it doesn't cover.
- Rollback ready: how to close the gate again, and who can do it, is documented.
- Decision recorded: the gate holder's name, date and decision (open / extend / stop).
What a shadow record contains
Every item that comes through shadow mode gets one entry in the record. That's what the gate holder decides from.
| Field | What it holds |
|---|---|
| Item | The input the step handled (an invoice, an entry, a request), with its type |
| Agent proposal | What the agent would have done, with the evidence it used |
| Human decision | What the human reviewer actually did with the same item |
| Agree / disagree | Whether the proposal matched the human decision |
| Reason | For a disagreement, the cause: incomplete rule, agent error or human baseline error |
Coverage, the agreement rate and any critical errors are read straight from these entries.
Related reading
- Related use cases: Approval gates · Readiness checklist · When not to automate · Automating month-end close · Invoice intake and exception triage
- Offer pages: How It Works · Agent governance · Book your audit
