Six months ago, an AI startup could often lead with capability and adoption while leaving the cost curve for later. That order made sense when the product boundary was still moving. It becomes a liability when usage grows, customers depend on the feature, and the company still cannot say what it costs to deliver the service it sells.

For 2027, agent unit economics should be part of company-building, not a diligence appendix assembled under deadline. The investor question is not simply whether the model is expensive. It is whether the company knows the cost of the work it promises, how that cost changes with customer behavior, and which levers can improve it without degrading the product.

Choose a unit that survives the pitch deck

Cost per token is rarely the unit a customer buys. A startup may need cost per customer, per completed workflow, per feature, or per transaction. The right unit depends on the product and pricing model. It should be consistent enough to compare cohorts and specific enough to show where the economics break.

That unit also needs a boundary. Does it include retries, retrieval, tools, hosting, orchestration, and human review? Are test traffic and production traffic separated? Is the cost tied to the provider invoice or inferred from partial telemetry? A confident unit-cost figure with an unexplained boundary is not a dependable underwriting input.

Investor test: Ask the founder to move from one real customer outcome to the workload that delivered it, then to the reconciled cost and the gross-margin treatment.

Keep learning spend separate from the product promise

Research and experimentation are supposed to produce uncertainty. Production is different. When a workload supports a feature customers rely on, its cost belongs in the company's view of cost to serve. Combining research spend with production COGS can make both look less informative: experimentation appears wasteful, while the product margin remains hidden.

Six months ago, a single AI-spend line may have been a useful placeholder. By the next financing conversation, investors should expect the company to show the transition from exploration to production and state who owns each pool. Do not penalize a young company for experiments. Do challenge a company that cannot explain which experiments became recurring product cost.

Economics should shape product choices

Useful measurement changes a decision. It can reveal that a workflow needs a different model, a different amount of context, a revised price, or a product boundary that better matches customer value. It can also show that a more expensive model improves completion enough to justify its cost. Optimization is not synonymous with choosing the cheapest model.

Ask what the company would change if cost per completed unit rose, and what evidence would justify keeping the current design. If the answer is only "we will optimize later," the plan lacks an operating mechanism.

Make the next round an update, not a reconstruction

Investors can help portfolio companies establish a small evidence standard early: named owners, explicit cost boundaries, invoice reconciliation, an agreed unit, and a review cadence connected to product and pricing decisions. This makes comparisons more useful without pretending different companies have identical architectures.

Six months from now, ask what changed in the unit economics and why. A useful answer will show its assumptions, explain variance, and connect the evidence to an action. That is stronger than a polished forecast because it gives the next decision something real to build on.