Setting an Outcome Baseline

Every AI program eventually meets the same question from a CFO: what changed because of this? Without a baseline, the honest answer is that nobody knows. The number may have moved, but it may have been moving anyway, or the definition may have shifted, or the team may be counting a different thing than finance is.

An Outcome Baseline settles that before any work starts. It takes a few weeks, it is mostly careful measurement, and it is the difference between a result you can report to a board and an anecdote.

1. Pick one number

Not a theme, a dashboard or a set of KPIs. One number that the business already cares about. A good outcome number passes four tests:

• Leadership already sees it. It appears in a board pack, an operating review or a plan.
• One person owns it. A named leader would be asked about it if it fell.
• It can move within a quarter. Not a five-year metric that hides every change.
• Your systems already hold it. It can be counted from records, not from a survey.

2. Define it exactly

Most disputes about results are really disputes about definitions. Write the number down so precisely that two analysts would count it the same way, and have finance agree to it.

Outcome definitionEligible referral-to-Rx rate

Counts

Prescriptions signed for referred patients

Out of

Eligible referrals received in the same cohort week

Excludes

Duplicate referrals; patients outside coverage

Window

Prescription within 60 days of referral

Source

EHR referrals and the order system

Owner

VP Patient Access

Agreed by

Finance, before work starts

3. Measure the history

Pull at least a quarter of history, and a full year if seasonality matters. Look at the level, the trend and the spread. The baseline is not one number on one day: it is where the number normally sits, and how much it wobbles on its own.

The shaded band matters as much as the median. A week above the median proves nothing. A sustained move above the band does.

4. Find where it stalls

Map the path the number travels: each stage, who owns it, which system records it, and how long items normally wait there. Then find the leak, the stage that loses the most. The leak is where the first Outcome App will work, and it is what makes the target believable.

On the Referral to Rx page you can see a worked example: five stages, five owners, and a benefits step that loses one in four patients.

5. Agree the target and the date

Size the target from the leak, not from ambition. If the leak loses 26% of patients and the app can recover a third of that, the target follows. Write down the target, the date, and the owner who signs it. A target that the owner did not agree to will not be defended when the number is reviewed.

6. Agree how you will count

Before anything goes live, agree how movement will be reported: weekly, against the baseline band, from the same source systems, with any change to the definition logged and dated. When the number moves, there is then nothing left to argue about except what to do next.

The Baseline checklist

0 of 8

• One number, already seen by leadership
• One named owner for that number
• A written definition finance has agreed
• At least a quarter of history, from source systems
• The normal range, not just the median
• The path mapped, and the leak found
• A target and a date the owner signed
• How movement will be counted, every week

Who does what

If the Baseline shows that an Outcome App is not the right lever for your number, we will say so in writing. That is part of the point.