Kerbtally Seed Extension
AI startup pitch deck example with the inference cost shown
An investor looking at an AI company in 2027 has one question the pitch has to answer without being asked, which is what a model call costs and who pays for the humans behind it. Kerbtally is invented, and its eleven slides put inference on its own bar and count the review team inside gross margin.
Every slide, in order
The whole deck as it renders, one slide after another. Read it the way the audience would.
The structure
What each slide is doing, so you can reuse the order even with different content.
- Slide 1cover
- Name the customer, not the technology, and put the margin on the cover.
- Slide 2intro
- Why this can be built now, argued from the cost of a model call rather than from hype.
- Slide 3pillars
- Four things the product does, with the human review step named rather than hidden.
- Slide 4comparison
- A month of bookkeeping done both ways, hour by hour, on the same one van business.
- Slide 5metrics
- The three headline numbers with the margin definition stated, since that is where AI decks cheat.
- Slide 6metrics
- Price down to margin per customer per month, with inference as its own bar.
- Slide 7timeline
- The model milestones, so the cost curve on the last slide has a history behind it.
- Slide 8pricing
- One price, one card, with the thing that is deliberately not charged for listed.
- Slide 9content
- Four things that are true about the data, because this is where the diligence questions land.
- Slide 10investment
- Four million in a table, sized to a break even month rather than to a valuation.
- Slide 11closing
- Close on the one figure that decides the round, with nothing decorative around it.
Slide by slide
Slide one names the customer rather than the technology. Slide two argues why now from the cost of reading one receipt. Slide three is four things the product does, with the escalation step named. Slide four compares a month of bookkeeping done both ways. Slide five is the three headline numbers with the margin definition written out. Slide six is the price to margin waterfall in dollars. Slide seven is five model milestones and what each one moved. Slide eight is the single price card. Slide nine answers the four data questions. Slide ten is the four million ask as a table with a break even month. Slide eleven closes on the eleven cent figure everything else follows from.
How to adapt this deck
A company selling to enterprises rather than sole traders swaps the single price card for a contract value slide and adds a security appendix, because the buyer is procurement rather than a person with a phone. A business whose inference runs on its own hardware replaces the waterfall with a capacity slide, since the cost question becomes utilisation instead of price per call. A deck raising a Series A rather than an extension moves the model milestone timeline to the back and puts cohort retention in its place, because by then the question has shifted from whether the margin works to whether the customers stay.
What this audience asks for
The Australian tax office publishes what a business activity statement has to contain and when it is due, and the Tax Practitioners Board publishes who is allowed to prepare one for a fee. Those two documents are why this deck has a registered agent reviewing every quarter and says so on the product slide, because a bookkeeping product that quietly does regulated work is a liability rather than a margin. An investor who has funded accounting software before checks that first, ahead of any question about the model.
What makes this deck work
Margin is defined before it is claimed
Slide five says gross margin here is after inference and after the human review team, and then says plainly that a lot of AI companies quote a number that excludes the people who catch what the model misses. Defining the number is what makes 72 percent worth reading.
The waterfall is in dollars, not percentages
Slide six runs 38 dollars down through 6.10 of inference and 4.50 of human review to 27.40 of margin. Percentages let an audience assume a different cost base. Four bars in dollars leave nothing to assume, which is the point of putting the chart in the deck at all.
The moat slide says what the moat is not
Slide nine states that the defensible asset is not the model but 2.9 million corrected receipts mapped to a trade specific chart of accounts, and that two providers are tested monthly so inference can switch in a day. Naming the thing that is replaceable is what makes the rest credible.
Questions people ask
What goes in an AI startup pitch deck?
Why the cost curve makes the product possible now, what it does including where a human is involved, the adoption numbers, unit economics with inference as its own line, the model history, pricing, the data position and the ask. This example uses eleven slides.
Should I show inference cost to investors?
Yes, as a separate line. Investors have learned to assume the worst when it is folded into cost of revenue. This deck shows 6.10 dollars of inference against a 38 dollar price and puts the number that drove it, eleven cents a receipt, on the closing slide.
How do I handle the human in the loop in a deck?
Name the percentage and cost it. This example says six percent of receipts go to a person, that the reviewers cost 4.50 dollars per customer per month, and that they are inside gross margin. Hiding a review team is the fastest way to lose a technical investor in diligence.
What is the moat in an AI application company?
Usually the corrected data and the workflow, not the model. This deck says 2.9 million receipts corrected by a human and mapped to a trade chart of accounts, which took escalation from 31 percent to 6. A deck claiming the model itself is the moat invites a very short conversation.
How many slides for an AI seed or seed extension?
Ten to twelve. This one is eleven. Architecture diagrams are the usual filler and they belong in a technical session. What earns space instead is the cost history, because it shows the margin is a trend rather than a snapshot.
Do I need a separate slide about data handling?
If your customers upload their own documents, yes. This deck uses one slide for four answers: documents stay in the customer tenant, inference runs in a named region, two providers are tested monthly, and the corrections belong to the company. All four come up in diligence anyway.
Build your own in about a minute
The button below opens the generator with this use case already described. Change the wording to match your own, generate, then edit anything you like.
Make one like thisOther presentation examples
Construction tech pitch deck measured in defects per site
A construction tech pitch deck has to prove the software changes what happens on site, because builders do not pay for apps, they pay to avoid rework. Sitewell, an invented Brisbane inspection product, shows rework items per site falling 34 percent over eight months while logged defects more than double.
Gaming studio pitch deck for a publisher meeting
A publisher is not buying a game idea, they are buying a team that finishes on time and an advance they can get back. Tidelamp Studio is invented, and its ten slides lead on wishlists, prove delivery with two shipped titles and end on a recoup schedule rather than a valuation.
Food and beverage pitch deck built on velocity by store
A grocery investor and a category manager are asking the same question in different words, which is how many units leave the shelf each week. Fizzroot is invented, and its eleven slides are built on that number, broken out by retailer and then taken apart into cents per can.
Want the steps in the builder? Read Make a pitch presentation with AI, then Add charts, maps and diagrams to slides. For everything this generator can do, see the presentation maker.
Written and checked by the OneCraft team. Last checked .