In short: investors now price AI application companies on workflow ownership, margin and retention, so raise on that evidence and put the model in the appendix.
Key points
- Model companies are priced on capability and compute; application companies on retention, margin, distribution and replaceability.
- Expect questions on gross margin after inference, provider feature risk, model switching, data rights and free usage.
- A company that has swapped models once is more fundable than one that says it could.
- Show cloud and model credits as a cost reduction with an expiry date, never as revenue or margin.
- Lead the deck with the customer and workflow; architecture and eval results belong in the appendix.
This is operating guidance, not investment advice.
The market has split
Two years ago "AI" in the deck earned a premium on its own. That has ended for application companies. Investors now sort the market into model companies and the companies that build on them. The two get priced on different logic.
Model companies are priced on capability, compute access and research talent. Most founders reading this are not building one. Application companies are priced on what any software company is priced on: retention, margin, distribution and how hard the product is to replace.
The difference is that an AI application carries two extra costs. Inference, which hits gross margin. And platform risk, because the model provider can ship your feature. Part 11 covered how to answer the "wrapper" objection. This part is about how that answer shows up in a raise.
The five questions you will get
Expect these in the first meeting, not the last.
Gross margin after inference. Investors want the number net of model and compute costs, and they want the trend. A flat or rising cost per task is a problem. A falling one is a story.
What happens when the provider ships your feature. The honest answer is that some of your product will be absorbed. The question is what remains. Data, distribution or workflow depth, from Part 11, are the three defensible answers.
Model dependency and switching cost. Can you move to another model in a week, a month, or never? Have you done it? A company that has swapped models once is more fundable than one that says it could.
Data rights. Do your contracts let you use customer data to improve the product? Many early contracts do not, because nobody asked. Fix that before the raise, not during it.
Paid or free usage. Free-tier activity is not evidence of demand. It is acquisition cost. Investors will strip it out, so strip it out first.
The evidence that counts
Part 3 argued for raising against evidence rather than narrative. For an AI company the evidence has four parts.
Production usage with retention. Not pilots, not demos. Teams using the product weekly and still there at 90 days.
Gross margin trend. Three or more months of margin net of inference, moving the right way.
Cost per task falling. This is the operating proof that your margin story is real. Show the work: caching, routing to cheaper models, prompt reduction.
A security pack that has passed review. Part 12 covered the pack. A pack that has been through a real buyer's security process is worth more than any claim about enterprise readiness.
The capital stack
An AI raise usually has more than one source of capital, and they do not behave the same way.
| Source | What it is for | What it is not |
|---|---|---|
| Equity | Runway, hiring, go-to-market | Cheap, or quick |
| Cloud and model credits | Reducing inference cost for a period | Revenue, or margin |
| Startup programmes | Credits, introductions, sometimes distribution | Validation of the business |
| Strategic investors | Distribution or data access | Neutral capital |
Credits are runway, not revenue. Show them in the plan as a cost reduction with an end date. Do not let them flatter the margin line in the pitch. Investors will find the expiry date and ask what margin looks like after it.
Strategic investors from cloud or model providers want something: distribution of their platform, or data from your use case. That can be a good trade. Know what you are giving before you take the cheque, and keep model switching possible in the terms.
Narrative discipline
The deck should read like a software company that happens to use AI, not an AI company looking for a use case.
Lead with the customer and the workflow. Which job, for which team, replacing what. Then the evidence from the section above. Then the market and the plan.
The model, the architecture and the eval results go in the appendix. They matter in diligence. They do not carry the first meeting, and leading with them signals that you have not found the customer yet.
Run the round the way Part 4 describes: a pipeline, a timeline, a named owner. AI investors are busy and have seen a lot of similar decks. A tight process is itself evidence.
The checklist
- [ ] Gross margin net of inference, three months of trend, in the deck
- [ ] Cost per task, with the three things you did to reduce it
- [ ] Model switching tested once, with the time and cost recorded
- [ ] Data rights confirmed in every live contract
- [ ] Free-tier usage separated from paid usage in every metric
- [ ] Credits shown as a cost reduction with an expiry date, not as revenue
Common questions
How do investors value AI application startups?
On what any software company is priced on: retention, margin, distribution and how hard the product is to replace. AI applications also carry inference cost and platform risk.
What questions will investors ask an AI startup?
Gross margin after inference, what happens when the model provider ships your feature, model dependency and switching cost, data rights in customer contracts, and how much usage is paid versus free.
Should cloud credits count as revenue in an AI startup pitch?
No. Credits are runway, not revenue. Show them as a cost reduction with an end date, because investors will find the expiry date and ask what margin looks like after it.
What evidence matters most when raising for an AI company?
Production usage with teams still active at 90 days, three or more months of improving gross margin net of inference, falling cost per task, and a security pack that has passed a real buyer's review.