Ask For The Odds

A tester did something clever with OpenAI's new Decisions API last week, and the result should change how you ask AI to predict anything.

Quick background. On October 7, OpenAI put its Decisions API into public beta on GPT-6 Luna. It returns typed answers instead of paragraphs. You can ask for the probability that a statement is true, a choice from a fixed list with confidence scores, or a score within a range. Input costs ten cents per million tokens and output is free.

The tester set up a weighted coin. Seventy percent heads, thirty percent tails. Then they asked the model about it two ways.

Asked as a choice, "will this flip land heads or tails," the model picked heads 98 percent of the time.

Asked as a probability, "what's the chance this flip lands heads," the model said 70 percent.

Nothing changed but the question. Ask it to pick and it bets the farm. Ask for the odds and it tells you the truth.

Now think about how you've been using AI for planning. "Will this deal close?" "Will we hit the deadline?" "Should we expect this client to renew?" Every one of those is a coin flip dressed up as a yes or no question. And if you've been asking it to pick, you've been getting the 98 percent version. Confident, tidy, and way too sure of itself.

This week's vault is about asking for the odds instead.

You don't need the technical explanation to use this, but the short version helps.

When you ask a model to choose, it does what any decent guesser does with a 70/30 coin. It picks the more likely side, every single time. That's the right move if you're betting on one flip. It's terrible if you're planning a quarter, because the 30 percent outcome shows up three times in ten and your plan never saw it coming.

Ask for a probability instead and the model has to describe the uncertainty rather than resolve it. That's what you actually need for planning. You don't need to know which way it'll go. You need to know how much room to leave for the other way.

The same thing happens with people, by the way. Ask your sales lead "will the Henderson deal close?" and you'll get "yeah, I think so." Ask "what are the odds, out of a hundred?" and you'll get "honestly, maybe sixty." That forty is the most useful number in the conversation.

Before the prompts, four rules. They apply to all seven.

Ask for a number between 0 and 100. Not "high, medium, low." Words hide disagreement. One person's "likely" is 60 and another's is 90.

Ask for the reasons on both sides before the number. If the model writes the number first, everything after it is justification.

Give it a base rate when you have one. "Historically, 35 percent of our proposals close" anchors the estimate in your reality instead of the model's general impression.

Write the prediction down somewhere you'll see it again. A forecast you never check is just an opinion with decimals.

Now the prompts.

Use this on any deal in your pipeline over a size that matters. Paste in the call notes, email thread or CRM history.

Run this across your top ten deals and add up the probabilities times the deal values. That's your weighted forecast, and it'll be a lot closer to reality than the column in your CRM where every deal sits at "commit" until the day it dies. If your pipeline lives in Go High Level, add a custom field called "AI odds" and paste the number in each week.

Never ask "will we finish by Friday?" Ask for a range.

Promise clients the P90. Plan your team's week around the P50. Celebrate the P10 if it happens. Most blown deadlines start when somebody promises the P10 date like it's a sure thing.

People and models both love to reason from the details of their own situation. This prompt forces them to look at how similar situations usually turn out first.

If you only steal one prompt for anything new, steal this one. A product launch, a new hire, a price increase, a new market. It won't always know the real base rate, so ask it to tell you how confident it is in that first number too. The point is to start from "most of these fail" and then argue your way up, instead of starting from your own excitement and arguing your way down.

You've probably done a pre mortem, imagining the project failed and asking why. This version attaches odds, which makes it actually usable.

The sort is the whole trick. A plain pre mortem gives you a scary list. This one gives you a ranked to do list.

FROM THE AI NEWSROOM

The exact systems behind everything in this issue. The audit sheets, the routing logic, the templates and the review cadences, built out step by step so you can copy them straight into your own stack. One time, forty seven dollars.

Get the Blueprint for $47

.....

Every forecast should come with the tripwires that would change it. Run this right after any of the prompts above.

Now you know what to watch. If the deal odds hinge on whether the CFO joins the next call, you know exactly what to push for. If the project timeline hinges on the vendor delivering by the 20th, that's the date to call them on the 15th.

For any business with repeat clients, this one earns its keep quick. Paste in the client's history: usage, support tickets, recent emails, payment history, meeting notes.

If you record client calls, Fathom transcripts make this far more accurate, because the model can hear the hesitation in what people actually said rather than in your summary of it.

This is the one that makes the other six get better over time. Once a month, take every forecast you logged and what actually happened.

If the things you called 80 percent only happen 55 percent of the time, you're overconfident, and now you know by how much. Most people are. Nothing personal about it. Treat it like a bathroom scale that reads three pounds heavy: once you know the error, you correct for it.

All of this falls apart without a log. Keep it simple: one sheet, five columns. Date, the question, the probability, the date you'll know, and what happened.

If you want it hands off, build it in Make.com. A form where you paste the question and the odds, a row appended to the sheet, and a reminder sent to you on the "date you'll know" asking what happened. Fill in the outcome, and the monthly calibration check has everything it needs.

Ten minutes to set up. After three months you'll know more about your own judgment than most business owners ever learn.

Garbage in. The model only knows what you paste. A deal history with no notes from the last call will get a confident number built on nothing. If something important isn't written down, the forecast can't see it.

False precision. A 63 percent estimate isn't meaningfully different from 60. Use the numbers to sort and plan, not to argue over decimals.

Treating it as the answer. The model's number is a second opinion. Write down your own number first, then run the prompt. Where you disagree by more than twenty points, that's worth a conversation, with yourself or your team.

Forgetting the base rate. Without your real numbers, the model falls back on its general impression of how things go. That's fine for a quick read and useless for a forecast you'll bet a quarter on.

FROM THE AI NEWSROOM

For operators who would rather build it with someone than build it alone. The full stack, installed alongside you, with the decisions made in the room instead of guessed at. Ninety seven dollars.

Join the Accelerator for $97

.....

Pick your five biggest open deals or projects.

Write down your own odds for each one first, a number from 0 to 100. Don't skip this.

Run the deal odds prompt or the deadline range prompt on all five. Then run "what would change your mind" on the two that matter most.

Start the log with those five rows.

Put a reminder on your calendar for thirty days from now to run the calibration check. You'll have your first read on whether you, the model, or both of you need to recalibrate. Stop asking it to pick. Ask it for the odds.

Jordan

The AI Newsroom | Practical AI for people with a business to run.