Cash Flow Volatility: How Startups Can Stay Liquid

Every startup deck mentions AI now. That is precisely why mentioning it no longer does any work. Investors have moved on to a narrower question: what does AI do to this company's unit economics?

It is a fair question, and most founders answer it badly — not because the answer is bad, but because they answer at the level of capability ("we use AI to…") rather than at the level of the financial model. Here is where AI actually shows up in the numbers.

It moves cost from opex into COGS

Traditional software has near-zero marginal cost per user. AI-native software does not. Every inference call has a price, and that price scales with usage rather than with headcount.

The consequence is structural: spend that would once have sat in operating expenses moves into cost of goods sold. That reclassification changes gross margin, and gross margin drives your multiple. A company at 80% gross margin and a company at 45% are valued on different planets at identical revenue.

If you are building on third-party models, work out inference cost per active user per month and carry it explicitly. Do not bury it in "infrastructure."

Margin compression is the diligence question

The pattern is consistent: early gross margins look excellent because usage is low and provider credits are covering the gap. As usage scales, the credits run out — and margins compress exactly when the company is trying to raise on growth.

Model the unsubsidized case before you need to. If margin compresses with scale, say so and show the path back: caching, smaller models for routine calls, moving high-volume paths in-house. Investors respond far better to a founder who has already priced the problem than to one surprised by it.

Where it genuinely reduces burn

The reliable savings are unglamorous and mostly internal — first-line support triage, sales research and list-building, contract review, reconciliation and close preparation, and the long tail of document handling that quietly consumes finance and operations time.

These are real. They also tend to be smaller than founders project, because the work rarely disappears; it converts into review time. Budget the review.

Where it does not

AI does not compress the parts of a business that were already hard: enterprise sales cycles, regulatory approval, hiring senior people, building distribution. If your path to profitability depends on AI removing one of those, the plan has a gap in it rather than a strategy.

What diligence looks at now

Three things come up consistently:

Data rights. What are you permitted to train on, and can you evidence it? Ambiguity here is a discount, or a broken deal.
Model dependency. If one provider changes pricing or terms, what happens to your margin? Concentration risk applies to model providers exactly as it applies to customers.
Substance. Whether the AI is doing load-bearing work, or sitting on top of a conventional product as positioning. Buyers can tell, and the second case is priced accordingly.

What to put in the model

Four lines, carried separately rather than lumped together: inference cost per active user; the same figure without provider credits; review time attached to any AI-driven workflow you are claiming savings on; and a sensitivity showing gross margin if model pricing moved 30% against you.

Founders who bring that to a conversation change its character. The discussion stops being about whether the AI story is credible and starts being about the business — which is where you want it.