How the AI review works, and why it never decides

The monthly decision meeting had one problem we kept running into. A lot of the time went to figuring out what a pitch actually meant. What’s the result? How would we know it worked? Which goal is it for? Those are fair questions, but they aren’t decisions, and the meeting is an expensive place to answer them.

So every pitch now gets an AI review before it reaches leaders. The job is narrow on purpose. It makes sure a pitch is ready to talk about, so the meeting can spend its time deciding.

The four checks

The review looks at four things. Each one is checked on its own, so a strong reason doesn’t make up for a missing result.

1. It ties to one of your big goals. That can be direct, or through a connection someone running your business would recognize. A tool that keeps customers from canceling helps a revenue goal even if the pitch never says "revenue".
2. It says what will be done by the end of the month. "Work on onboarding" doesn’t pass. "New onboarding checklist live for every new hire" does.
3. You can tell at month end whether it worked. For a project or a fix, a clear definition of done is enough. If the pitch promises to move a number, it needs a target, even a rough one.
4. There’s a real reason for it. A problem someone saw, a customer request, a recurring cost, a number, an incident. Something that actually happened.

Different pitches get different bars

The review first figures out what kind of pitch it’s reading: something to build or deliver, a fix for a known problem, a bet on moving a business number, or a short investigation that ends in an answer.

Each kind gets a bar that fits it. A new contract template doesn’t need a revenue forecast. A pricing test does need a target. If every pitch had to come with a spreadsheet, people would stop pitching.

What happens to a weak pitch

There are three results.

• Ready for the meeting. All four checks pass. Any ideas to make it better come along as suggestions, and suggestions never hold it up.
• Needs a rewrite. It ties to a goal, but something else is missing. The person gets at most two questions to answer and a template to fill in.
• Doesn’t tie to a goal. If none of your goals covers the work, it doesn’t go to the meeting.

The review never makes up numbers, evidence or facts the person didn’t give it. It can clean up the wording. It can’t make a pitch look stronger than it is.

We made it easier to pass

Our first version was too strict. When we ran it on pitches leaders had already approved, it sent back about three out of four. That was a lot of extra work for people with good ideas.

The current version starts from the idea that most pitches written in good faith, by someone who knows the business, should pass. On the same test it now passes most of the pitches leaders approved, and it still catches the ones with no result or no reason behind them.

Why it never decides

We could have let the AI rank pitches and approve the top few. We didn’t, on purpose.

Deciding depends on things the AI can’t see: what your team can take on this month, who’s already stretched, what a customer told you yesterday, what you’re willing to bet on. Those calls belong to the people who are responsible for the result. If a priority fails, nobody should be able to say "the AI approved it".

So the AI weeds out weak ideas and helps strong ones get sharper. Leaders approve, bank or reject every pitch in the decision meeting.

What happens to your data

To review a pitch, its text goes to our AI provider to produce the review. We don’t let the provider train its models on your pitches. Your pitches, goals and updates are your company’s data.

Calvin SeamonsHead developer of Unbreakable OS. He builds intelligent systems for a living and leads engineering across the product’s AI, web and mobile apps.