Nine minutes in, you’re halfway through a slide about your model architecture and the partner across the table has started checking her phone. You haven’t lost because the tech is weak. You’ve lost because of the order you said things in.
Pitching AI is its own craft. The technology is unfamiliar enough that investors default to skepticism, and familiar enough that everyone thinks they already understand it. Here’s the sequence that tends to work, with the slides, numbers and answers that actually move people.
Lead With the Decision Your Buyer Is Already Trying to Make
Nobody wakes up wanting an AI platform. They wake up with a decision that’s expensive to get wrong: which claims to pay, which contracts to flag, which alerts to escalate at 3am.
Google’s $32 billion bet on Wiz is a useful model here. The security company didn’t sell investors on graph algorithms. It sold them on finding the paths an attacker would realistically take through a cloud environment, which is precisely the question a CISO is already losing sleep over. Framing your pitch around what Google’s cloud security acquisition actually does is a reminder that the winning pitch names the decision, not the technique.
Open with one sentence: “Your team makes 400 of these decisions a week, and each one takes a specialist 12 minutes.” That’s your first slide.
Two Problem Slides, Zero Technology Words
If the word “model” appears before slide three, you’ve jumped the gun. Get specific about the pain. Not “insurance companies struggle with manual review” but: a mid-size insurer processes 62,000 claims a month, a reviewer clears roughly 90 a day, and the error rate on complex claims sits near 4%.
A problem slide earns its place when it has all four of these:
- A named role who owns the pain (Director of Claims Operations, not “enterprises”)
- A number attached to it in hours, dollars or churn
- Evidence you’ve seen it firsthand, such as 15 customer interviews or a paid discovery sprint
- The ugly workaround they use today, described honestly, spreadsheets and all
That last bullet matters more than founders expect. If you can describe the current workaround better than the buyer can, they assume you understand everything downstream of it.
Show the Delta, Not the Model
Your third slide is one number pair. Twelve minutes becomes 90 seconds. Eleven-day turnaround becomes same-day. Cost per reviewed contract drops from $6.80 to $0.40.
Notice what’s missing: accuracy claims in isolation. “94% accurate” means nothing without a baseline. Say “our baseline reviewer catches 82% of the flagged clauses; the system catches 94%, and we verified that on 1,200 held-out contracts from your industry.”
Know your unit economics cold, too. If you’re running on a managed platform, be ready to say what an inference costs you today and what it costs at ten times the volume. Teams that demonstrate this fluency early tend to do well, and if you’re weighing platforms, it helps to understand what Google’s ML platform actually does and what it charges before someone asks you to compare.
Run the Pilot on Their Data Before You Ask for Money
The single biggest conversion lever in an AI pitch is showing a result from the buyer’s own environment, however small. A 30-day pilot with clear edges beats any benchmark.
A structure that works: one workflow, one team, one agreed success metric, and a baseline measured in week one before anything changes. So if the metric is review time, you measure the current median in week one, deploy in week two, and report on weeks three and four. Bring the pilot offer on a single slide with the timeline drawn out. Vague pilots die in procurement.
Answer the Moat Question Before It’s Asked
Every experienced investor is running the same calculation: what stops a better-funded team from copying this next quarter?
Founders get creative here, and that’s fine. One eSports startup raised $20M on the strength of a reframing trick that turned an AI-skeptical room around — proof that the story, not the stack, decides the outcome. Your honest moat is usually some combination of proprietary workflow data, switching costs from being embedded in a regulated process, and distribution you already have. Say which one it is. “We’ll stay ahead because we’re smarter” is not a moat, and everyone in the room knows it.
Close With an Ask Tied to a Milestone
End on three lines. Here’s the amount, here’s the 18 months it buys, and here are the three milestones we’ll hit: one production customer live by month five, $400K ARR by month twelve, gross margin above 70% by month eighteen.
Specificity signals you’ve done this before. It also gives a hesitant investor a reason to say yes to the small version of the round.
The Three Questions That End AI Pitches Early
Prepare a 30-second answer for each. Rehearse them until they sound unrehearsed.
“What happens when the model is wrong?”
Describe the failure path, not the accuracy figure. Who reviews it, how fast it’s caught, what the customer sees. A human-in-the-loop step with a defined escalation route beats any confidence score.
“Why can’t OpenAI just do this?”
Answer with the parts that aren’t the model: the integrations, the compliance work, the labelled edge cases, the fact that nobody wants to rebuild this themselves. Founders get defensive here. Don’t. The hard reality of selling AI to investors is that generic capability gets commoditised fast, and the durable value sits in what surrounds it.
“What does this cost per user at ten times the volume?”
Have the number. If margins compress badly at scale, say so and explain the plan: caching, smaller fine-tuned models, or renegotiated commitments.
Same Deck, Different Room
A pitch to a seed investor and a pitch to a procurement committee share roughly 40% of their content. Investors want the market and the moat. Buyers want integration effort, security review and what happens if they cancel. Build the deck as modules so you can drop slides in and out depending on who’s sitting down.
Audience fit is unforgiving. Google has spent the past year pitching an AI agent ecosystem to consumers who may not want it, which is a good reminder that a brilliant pitch delivered to the wrong room still reads as a failure.
Rehearse It Out Loud, Out of Order
Most founders practise the deck start to finish, which is the one order they’ll never get. Real meetings jump. Someone interrupts during slide two and asks about pricing. Someone else wants the security answer before you’ve explained the problem.
So rehearse in fragments. Have a friend open with question three, then five, then one. Time each answer to 30 seconds and stop talking at the end of it. Record yourself once on your phone and watch it on mute; if your hands and eyes are doing something different from your mouth, the room will notice.
Then do the unglamorous part. Send a two-page follow-up within four hours that restates the problem in their words, the pilot terms, and the number you promised to find. The pitch ends when that email lands, not when you leave the building.