AI-native service businesses: what they sell and how they work
How AI changes service delivery, what customers actually buy, and which costs matter before you scale.
- Written by
- Agentic Economy
- Published
An AI-native service business uses AI as a central part of delivering work customers pay for. The offer might be a researched company briefing, a campaign audit or an ongoing accounts-receivable service. The customer buys a defined contribution to their business; the supplier organises the software, information and people needed to deliver it.
For an established service firm, the useful question is which part of its expertise can become a repeatable offer. A task that takes less time to produce can support faster delivery, a different price or more customers. Whether that becomes a better business depends on quality, demand and the full cost of completing each job.
A service with responsibility for delivery
Y Combinator's April 2026 request for AI-native service companies describes companies delivering the work itself. It is an investment thesis about how services could be supplied, rather than evidence that every service category can already run autonomously.
Consider a small research firm. This illustrative business sells a briefing on a prospective customer: its activities, recent developments and likely relevance to the buyer's offer, with sources attached. AI could retrieve material and draft the briefing. A researcher could check ambiguous findings and release the finished work.
The firm's promise is the briefing's scope and quality. Its customer should not have to supervise each search or repair missing citations. The supplier needs a process for rejecting unsuitable requests, handling incomplete evidence and correcting a result that falls short.
That delivery responsibility is what makes the offer a service. The degree of automation can change as the firm learns which tasks are reliable.
A reported example: running finance operations
Rex's Y Combinator profile describes an AI-native service for enterprise order-to-cash operations: handling work from customer orders and invoices through to collecting and reconciling payment. Its offer includes work such as collections, portal submissions and inbox triage. The company says customers can use it as a managed service or operate the system with their own team, with supervised procedures progressing towards autonomy and exceptions retained for approval.
This is the company's account of its product and operations, checked on 7 September 2026. It establishes a concrete service offer; it does not independently verify its performance claims or show that the same approach will work in a smaller Australian firm.
The useful lesson is the specificity of the job. A finance team can compare an offer to carry out defined operational work with how that work is handled today. A general promise to “transform finance with AI” gives it much less to evaluate.
Three different ways a customer can buy
AI can change how a service is produced without changing who orders it. A person might still commission the work through a conversation, form or existing contract. An agent can also become the buyer's interface to that service.
| Offer | What the customer receives | What the supplier must establish |
|---|---|---|
| A project or retainer | Agreed work over a defined engagement | Scope, responsibilities, review and change requests |
| A repeatable service unit | One briefing, audit or processed document | Accepted inputs, output specification and charging event |
| A service available to agents | A callable way to order and receive work | All of the above, plus instructions software can use and clear purchase status |
An AI-native business does not have to sell through an application programming interface (API), the connection software uses to request work from another system. Likewise, a service bought by an agent does not have to use AI internally: a conventional database lookup can be useful to an agent. These are separate decisions about delivery and distribution.
Our explanation of how agents buy services follows the distribution side through a complete purchase.
Count the cost of accepted work
A cheap first draft can become an expensive service if every job needs substantial correction. Track costs against work the customer can actually use.
For the illustrative research firm, the direct cost of a briefing would include information purchases, model use, researcher review and delivery support. Rework belongs in that calculation too. Separately, the firm still needs to recover sales, administration and ongoing development costs.
Suppose a briefing sells for AUD 100. Assume AUD 10 in data and model costs, AUD 25 in review and AUD 15 in average delivery support and rework. That leaves AUD 50 before overheads. These invented figures demonstrate the calculation; they are not observed prices or an industry margin benchmark. If difficult assignments double review and rework costs, the offer needs a tighter scope, a different price or a better delivery process.
The concern has a longer history than today's agents. In a 2020 analysis of AI business economics, a16z identified computing costs, human support and edge cases as constraints on software-like economics. Its historical estimates are not current market benchmarks. The relevant question for your business is how much intervention your own service requires.
Price a result you can define
A price per document is understandable when documents have bounded size and complexity. A subscription may suit ongoing access or a recurring service. An outcome fee needs a result both parties can identify and a clear account of what the supplier controls.
For example, delivering a sourced company briefing is easier to attribute than winning the eventual sale. Charging for a sale introduces questions about the sales team's work, other marketing activity and the timing of the customer's decision. A higher-level business outcome can be valuable while still being a poor charging unit for one component service.
Start with an offer you can explain, deliver and evaluate consistently. Use the service preparation guide to specify its inputs, output, price and recovery path. If the delivery task itself is still uncertain, choose a first agent use case and test that before expanding the offer.