Here’s a question worth asking any AI vendor: when your system extracts a number from a document, how do you know it’s right?

Most won’t have a good answer. They’ll hand you an output and expect you to trust it. No explanation of how confident the model actually was, no way to know which fields deserve a second look before you act on them. You either take the output on faith or you review every single field by hand, which defeats the point of automating the work in the first place. Neither option scales, and in a business where a single misread number in a rent escalation clause can represent millions of dollars, blind trust isn’t an option.

We built Outcome’s confidence score to close that gap. It’s the mechanism that tells you, data point by data point, how sure the model actually is, and why. Not a promise of perfection. An honest read on where the model is confident and where it isn’t.

How the Confidence Score Works at Outcome

Take lease abstraction as an example. When a lease runs through Outcome, every extracted field gets two things: a confidence score and a written explanation for that score.

The explanation matters as much as the number. If a tenant name appears differently in two places in the document, the reasoning column tells you that directly: something like “tenant name appeared differently in multiple places in the document.” You’re not guessing why a field scored low. You’re explicitly told.

The scores themselves are color-coded, so your team can scan a lease at a glance:

• Green (90–100%): High confidence
• Yellow (75–89%): Medium confidence
• Red (below 75%): Low confidence

When a field drops below the threshold, the workflow automatically flags it for manual review. The model tells your team exactly where to look, instead of leaving them to review every field in every document. That’s human-in-the-loop done right: the model handles extraction, and your team spends its time on the parts that actually warrant a second set of eyes.

The Technical Side: How We Keep the Score Honest

A confidence score is only useful if it’s grounded in something real. Ours is tied to coordinate-backed citations in the source document itself, not general pattern-matching from the model’s training. When we say a field scored 94%, that score traces back to a specific location on a specific page.

We also run a strict verifier on top of the extraction. A field only gets marked as verified if the extracted value matches the source value exactly. That verifier is what prevents false positives, meaning a score that looks confident but shouldn’t be.

Leases rarely come as a single document. You have the original lease plus a stack of amendments, and each one can change the picture. When we process a grouped document set, the pipeline reviews the extracted data across all of them together and generates one consolidated score. We don’t average the individual document scores together and call it a day. The model reasons across the full document set the way an analyst would.

If a single data point fails verification, only that data point gets re-extracted. The rest of the document stays exactly as it was. You’re not waiting on a full reprocessing cycle because one clause needed a second pass.

And because the model runs on fixed parameters, the same document produces the same score every time you run it. If your confidence scores shift every time you rerun a document, you can’t trust the score any more than you can trust the extraction.

Why This Matters More Than a Perfect Score

The goal is not a perfect score on every field. That’s not how extraction works, and any vendor telling you otherwise isn’t being straight with you. The goal is an honest score. One that tells your team the truth about how confident the model is, so you know when to move forward and when to check the work yourself.

That honesty is what makes the output something you can act on. Scores will improve as the model gets better, and they already have. But the transparency behind the score is the value your team gets today, not something you’re waiting on.

Plenty of AI systems will hand you a finished output and ask you to take it at face value. At Outcome, we’d rather show you what’s happening underneath the surface. Show the score. Show the reasoning behind it. Let your team see exactly why a data point was flagged, and let them make the call from there. That’s how trust is built between a real estate team and the AI it’s using.