Work Understanding: Copilot
Objective. Ask the copilot a process question tied to a decision, inspect the evidence it points to, and treat its answer as an AI suggestion to confirm rather than a fact.
The copilot answers natural-language questions about the discovered work and points to the evidence behind the answer. It answers 'can a non-analyst get a grounded answer quickly?' - as long as the evidence is inspected, not the wording trusted.
See it in the demo
What this screen answers
The customer question is: can someone who is not an analyst ask a question about the process and get a useful, evidence-backed answer? Sales and CSM people open the copilot to show that the insight in the app is reachable in plain language - 'where is the time going in this process', 'which step varies most' - without building a report.
The value is speed to a grounded answer. The risk, which you must name, is that fluent language reads as authority. A good copilot answer is one you can click into and check, not one you repeat because it sounded confident.
The task
- Open the copilot from the overview.
- Ask a question tied to a real decision and a defined scope - a specific process and period, not an open-ended prompt.
- Read the answer, then immediately open the evidence it cites: the underlying step time, variant or finding.
- Confirm the answer matches the evidence and the scope you intended, then use it. If the evidence is thin or off-scope, treat the answer as unconfirmed.
Do not invent prompt syntax or feature names. The teachable move is 'ask, then verify against evidence', not a specific phrasing.
How to read it
A copilot answer is an AI suggestion that needs human confirmation. The classic misread is taking the sentence as settled fact and quoting it to a customer. Always check three things: does the answer rest on evidence you can open, is the scope the one you asked about, and does it respect the same honesty rules as the screens (opportunity is not saving, association is not cause, overlapping findings do not sum).
If the copilot states a savings or causality claim, that inherits the same limits as the source screen - it does not become firmer for being phrased by an AI.
Honest limits
This screen does not turn a modelled estimate into a proven result, and it can be wrong - misreading the question, picking the wrong scope, or over-stating a confidence. It is a faster path to the evidence, not a replacement for judging the evidence.
Confirm with product what the copilot can access, how it is grounded, and what it is approved to answer for a customer. Never let a copilot sentence become a customer commitment without the evidence behind it checked and the scope confirmed.
Interpretation check
A copilot answer is an AI suggestion to confirm against its cited evidence and intended scope, never a fact to repeat because it sounded authoritative. It inherits the honesty limits of the screens it draws from.
In the field
An RCM lead asking the copilot 'why is this claim type slow' should land on the waiting evidence and open it - the honest answer is the evidence path, not the one-line summary repeated in a QBR.
Practical task
On the demo copilot, ask one scoped question tied to a decision, open the evidence behind the answer, and write a short note: did the evidence and scope support the answer, and what would you still verify before using it with a customer?
Self-check
The copilot states 'this step wastes 120 hours a month'. Can you quote that to a prospect?
Not as given. Open the evidence, confirm the scope, and recognise it is a modelled opportunity with the same limits as the source screen - a candidate to validate, not a booked saving.
Why is a fluent copilot answer a risk as well as a benefit?
Fluency reads as authority. The benefit is speed to the evidence; the risk is repeating a confident sentence without checking what it rests on or whether the scope matches.
What three checks turn a copilot answer into something usable?
The answer rests on evidence you can open, the scope matches what you asked, and any savings or causality claim respects the app's honesty rules.
Does phrasing a claim through the copilot make it more reliable than the underlying screen?
No. It inherits the exact limits of its source - estimates stay estimates, associations stay associations. The AI wording adds no proof.
Verified vs demo. Screen is live in the demo at /#/copilot; exact behaviour, grounding and figures are SME-validation-pending, and what the copilot can access and is approved to answer for a customer must be confirmed with product.