The Business of AI: The Hard Truths Builders Learn Too Late

The business of AI is easy to mistake for the software business. The product ships like software. It demos like software. It prices like software. But the economics underneath are different. Founders who ignore that learn it the hard way, usually right after the prototype works and the real company has to start.

This piece covers the business of AI end to end: what it takes to build, sell, and run an AI company profitably. It draws on public analysis from investors who track these businesses closely. It also draws on patterns from technical founders who made the jump from shipping code to running revenue. Treat everything here as a starting point to test, not a rulebook.

The margin problem in the business of AI

Classic SaaS has a famous property: build once, sell many times. Serving one more customer costs almost nothing. That's where fat margins come from.

AI breaks that assumption. Every query, generation, and inference carries a real compute cost.

Martin Casado and Matt Bornstein at a16z documented this pattern back in 2020. They found AI company gross margins often in the 50-60% range, below the 60-80%+ benchmark for comparable SaaS (a16z).

Bessemer Venture Partners made the same point more recently, in a February 2026 pricing playbook. They put AI margins at 50-60% versus 80-90% for SaaS (Bessemer). The SaaS benchmark differs slightly between the two pieces. Either way, the direction is the same: the business of AI carries real cost of goods sold.

Where does the money go? Three places, mostly:

• Cloud and inference. a16z found AI companies often spend 25% or more of revenue on cloud resources: training, inference, and storage of rich media (a16z). In the generative AI era, they estimated app companies spend roughly 20-40% of revenue on inference and per-customer fine-tuning (a16z).
• Humans in the loop. Data cleaning and labeling can take up to 10-15% of revenue, per a16z's 2020 analysis. Real-time human oversight adds more.
• Edge-case maintenance. AI products meet inputs nobody predicted. More on that below.

The lesson builders report: model your unit costs before you design your pricing. If the math fails at ten customers, it fails at a thousand. Bessemer makes the point directly (Bessemer).

One caveat. These figures come from different points in time. The a16z margins piece predates the generative AI wave, and costs per token have fallen since. But inference remains a real, metered cost. Build margin assumptions conservatively.

Where the money lands in the business of AI

Sequoia's David Cahn framed the revenue gap starkly in 2024. Infrastructure spending implied far more AI revenue than the market was producing. He called it the "$600B question" (Sequoia).

His math is worth understanding even if you disagree with it. GPUs are about half the total cost of ownership of an AI data center. He then assumes a 50% gross margin for the startups renting that compute. Run the numbers, and the revenue hole is enormous (Sequoia).

Two implications for founders:

• Compute is a commodity, metered per hour. Prices tend toward marginal cost. Don't build a business whose edge depends on compute staying expensive.
• Falling costs help builders. Cahn calls declining GPU prices "good for long-term innovation and good for startups." Your cost per interaction gets cheaper over time, if the product survives that long.

On value capture, a16z's 2023 analysis of the generative AI stack found infrastructure vendors capturing most of the dollars. App companies grew revenue fast but struggled with retention, differentiation, and margins (a16z).

Why distribution beats model choice

This is the uncomfortable part for technical founders. a16z could not find strong structural advantages anywhere in the generative AI stack. Apps use similar models. Models train on similar data with similar architectures. Cloud providers run the same GPUs (a16z).

If everyone can access the same models, the model is not your edge. So what's left? a16z lists the standard defenses: scale, supply chain, ecosystem, algorithmic, distribution, data pipelines. None tend to be durable long term, they note. But distribution is the one a founder can start building on day one.

Practical translation: pick the best model available today. Wrap it in a workflow customers value. Then spend your scarce time getting that workflow into buyers' hands. Model providers will keep improving. Your pipeline of qualified conversations compounds.

This is also where tooling starts to matter. Post-prototype, the founder's job shifts from writing code to running pipeline, outreach, content, and support. Doing that across five to eight disconnected tools adds overhead and scatters data. Some founders consolidate instead.

Zrevick, for example, combines CRM, lead data, outreach, and publishing in one AI-native workspace. An Identity Graph keeps one shared picture of each prospect across the team. Its MCP and API access matters for technical founders: go-to-market work can run through the same coding agents used for the product. That's one option worth a look.

The principle behind it is broader. In the business of AI, GTM deserves the same engineering care as the product itself.

Four failure points after the prototype

The prototype stage hides the hard parts of the business of AI. Builders consistently hit these traps.

1. Edge cases. a16z found 40-50% of intended functionality for AI products can live in the long tail of user intent (a16z). Users type anything into an open-ended interface. Each new customer generates data nobody has seen. One founder a16z interviewed called this the "time cost" of AI products. Every deployment starts with data collection and fine-tuning before real value shows up. Prototypes and betas cover common paths, not the tail.

2. Fake product-market fit. Five or ten friendly customers feels like PMF. Then the ML backlog balloons, deployments stretch, and new sales stall. a16z flags exactly this pattern (a16z).

3. Pricing that ignores COGS. Bessemer's warning is blunt: don't default to cost-plus. Test value-based prices and find the point where buyers hesitate (Bessemer). They also warn about soft-ROI copilots. Products that advise without closing the loop face a renewal cliff as pilots hit real budget cycles.

4. Unit economics that were never measured. Bessemer advises tracking the full cost stack from day one: inference, human oversight, founder time spent selling, support drag. Companies that skip this can scale into negative margins without noticing (Bessemer).

Lessons from builders who crossed to revenue

Pulling the threads together, founders who make the jump tend to do a few things.

1. They price before they scale. They know the cost per core interaction and design pricing that covers it.
2. They share models where possible. a16z saw a large COGS gap between per-customer models and one shared model across customers (a16z).
3. They narrow the task. a16z's advice: choose problem domains carefully, and often narrowly. High-scale, low-complexity tasks like moderation, data entry, and transcription minimize edge-case pain.
4. They build GTM like a product. Distribution, not the model, is the part they control and compound.
5. They keep humans in the plan. a16z suggests embracing services where the market demands it. Treat model maintenance and human failover as first-order problems.
6. They consolidate the business stack. One source of truth for customers beats five tools that disagree.

A starting checklist for founders

• Write down the cost per core interaction, including inference and human review time.
• Set conservative margin assumptions. Start near 50-60% gross, not 80%.
• Pick a charge metric customers understand, whether that's task, outcome, or hybrid, and test prices until buyers hesitate.
• Define your edge-case budget per new customer before signing them.
• Build one repeatable outbound motion before hiring anyone to run sales.
• Review unit economics monthly. Count founder selling time as a cost.

The takeaway

The business of AI rewards founders who respect its economics. Margins run thinner than SaaS. Costs are real and variable. Defenses are shallower where you hoped they were deep. But distribution, workflow depth, and unit-economics discipline are still yours to build, and compute keeps getting cheaper. The builders who win are not the ones with the best model. They're the ones still standing when the pilot invoices come due.

If the go-to-market side of this sounds familiar, start with Zrevick. It puts CRM, lead data, outreach, and publishing in a single AI-native workspace, and its MCP and API access means your coding agents can run go-to-market alongside the product. Wire up one repeatable motion, watch the unit economics, and let the margins tell you when to scale.