A shadow period is the stretch of time before go-live when an AI agent works real cases alongside the people who handle them, but nothing it produces is used. People do the work as before, the agent does the same cases in parallel, and the two are compared afterwards. No customer feels the effect.
Why run a shadow period?
Because certification on past cases, however strong, is not the same as live work. A shadow period shows how the agent handles today's cases, today's phrasing and today's documents. It reveals gaps that past cases did not contain: a new product, a seasonal pattern, a change in how suppliers format invoices.
It is also the first real training ground for reviewers. During shadow, they can look at what the agent would have done, compare it with what they did, and learn to read its verdicts before any of them matter. Many teams run reviewers' supervised practice on the shadow output, so people and agent become ready together.
How does it work in practice?
The agent is connected to the same work as the team, with every capability capped at depth 1 or 2, so it reads and proposes but cannot act. For each case, the comparison records what the person did, what the agent proposed, and whether they matched. Mismatches are sorted: agent wrong, person wrong, or both acceptable. The agent owner banks what the agent got wrong; the team lead coaches on what people got wrong. The shadow run report template gives a format for the weekly summary, and how to run an AI agent in shadow mode covers the setup.
An example
Take a hypothetical insurer's travel claims team. For four weeks, its agent assesses every incoming claim while handlers carry on as before. In week one the agent and handlers agree on most claims; the mismatches show the agent misreading a new airline's delay certificate. The agent owner adds the format to its sources. In week three, a mismatch reveals that two handlers had been paying a benefit the policy excludes. That is a people finding, not an agent one, and it goes to calibration. By week four, mismatches are rare and explained, and the claims manager takes the results to the go/no-go decision.
How is it different from shadow AI?
The words collide, but the ideas are opposite. A shadow period is a deliberate, governed test run by the business. Shadow AI is the use of AI tools by staff without the business's knowledge or approval, outside any governance. One is a control; the other is a gap in controls. Stopping the second is covered in how do you stop shadow AI agents.
A shadow period also differs from a grace period, which is time allowed after a deadline before consequences apply, such as a few extra days to pay a bill. And it differs from draft-only mode, where the agent's drafts are used after a person approves them.
How long should it last?
Long enough to see the full range of the work, including a peak. For daily work with steady volumes, three to six weeks is common. Low-volume or seasonal work needs longer. End it on evidence, not on a date. The guide to AI agent skills by role shows who should read the comparison.
Questions people ask
Does the agent's output reach customers during a shadow period?
No. That is the defining feature. If its output is being used, even after review, the agent is in draft-only mode, not in shadow.
Who decides when the shadow period ends?
The process owner, at a go/no-go decision, using the comparison results and reviewer readiness as evidence.