AI Academy · Book
Executives & Directors · Module 10 · Chapter 006

AI-Native Products and Business Models

An AI-native product sells the work done, not the tool for doing it. That changes three things at once: what you build, what unit you charge for, and what stops a rival with the same model from copying you. The second and third are where AI products most often go wrong.

≈ 14 min read

After this chapter you can

  • Distinguish AI-enhanced, AI-centric and AI-native products with the take-the-AI-out test.
  • Explain why AI products carry a cost of goods that varies by customer, and why outcome pricing still pays for failed attempts.
  • Choose among seat, usage, outcome and hybrid pricing from cost variability and outcome measurability, and recognize the input-pricing trap.
  • Explain why model access is not a moat and rank the layers that are harder to copy - context, workflow position, distribution and trust.
  • Identify the EU and US rules that make disclosure, liability and AI claims part of the product specification.

In December 2024 OpenAI launched ChatGPT Pro, its most expensive consumer plan, at 200 dollars a month with unlimited use of its strongest models. A month later the company’s chief executive, Sam Altman, posted that OpenAI was losing money on it: “people use it much more than we expected.” He had set the price himself, expecting it to make money1.

Read that twice. The product was so useful that its keenest customers made it unprofitable. In conventional software that sentence makes no sense, because one more hour of use costs the seller almost nothing. In an AI product every answer, every action and every minute of reasoning has a cost, and the customers who love the product most are the ones who run that cost up.

That is the business-model problem in miniature. The previous chapter, Autonomous Workflows and AI-Native Organizations, turned the redesign question inward, to how a company runs its own work. This chapter turns it outward, to what a company sells: what the product does, how it is priced, and what defends it.

The drill and the hole

Marketing teachers have a favorite line, popularized by the Harvard professor Theodore Levitt: people do not want a quarter-inch drill, they want a quarter-inch hole2. For fifty years the line was mostly a reminder about messaging. Software still sold drills: features, screens and licenses, with the customer’s own staff doing the drilling.

AI lets a product deliver more of the hole. A document tool can offer “get this contract prepared for signature” instead of an editor with templates. An analytics product can answer a business question and propose an action instead of handing over a dashboard to interpret. The product does more of the underlying work, and the customer supervises. That is a capability claim. Whether a given product delivers the hole reliably, on a customer’s real work and at a price that covers its cost, is something only its results can show, and public evidence on such results is still thin.

Three product types from AI-enhanced, where one feature improves, to AI-native, designed around work the product completes.AI-enhancedSame product; one featuregets smarterAI-centricCustomers spend much oftheir time working withthe AIAI-nativeDesigned around work theproduct now completesTest: take the AI out. Does the value proposition survive?
Figure 10.6.1 The spectrum is about what the customer buys, not how much AI is inside.

Products sit on a spectrum. An AI-enhanced product is the existing product with a smarter feature: a summary in search, a drafting button in a CRM. An AI-centric product puts AI at the heart of the experience, as research and coding assistants do. An AI-native product is designed from the start around work that the product itself now completes. A simple test separates them. Take the AI out. If the product still does roughly what customers pay for, it was enhanced. If nothing worth buying is left, it was native.

Neither end is better in itself. Enhancement is often the right move for a mature product with loyal customers. The mistake is to ship an enhancement and price, market or defend it as if it were native.

A drill-selling product offers features and is paid per user; a hole-selling product completes the work and is paid for the work done.SELL THE DRILLFeatures, screens and seats;the customer does the workPaid per userSELL THE HOLEThe product completes thework; thecustomer supervisesPaid for work donevs
Figure 10.6.2 AI moves products from the tool toward the result, and the business model has to move with them.

When the product does the work

Three design changes follow when a product starts selling the hole.

The interface moves from navigation to intent. Instead of menus and forms, the customer states a goal and the product works out the steps. That does not mean everything becomes a chat box. Much useful AI in products is invisible: a recommendation, a pre-filled form, an exception flagged before anyone asked. Conventional screens remain the right answer wherever the customer needs precision and visibility.

The product carries context. It remembers the customer’s preferences, past work and the state of an unfinished task, much as the agents in Multimodal and Agentic AI use connectors to reach business systems. Context makes the product more useful and harder to leave. It is also customer data, held for a purpose, and it must be governed as such.

The product’s behavior changes between releases. A model update or a new prompt can change answers without a single new feature. Customers who rely on the product to do work will notice before you do. Treat behavior as part of the product: test it before release, announce material changes, and let business customers pin or roll back where the work is high-stakes. When the product acts rather than advises, the controls from Agentic AI and Autonomous Actions apply in full: least privilege, approval for consequential steps, reversibility.

The cost of goods comes back

The hook was not a one-off. As The Economics of AI in Module 08 showed, AI companies have run lower gross margins than comparable software firms because every unit served carries compute and often human review3. Unit prices for a fixed level of capability keep falling, as Where AI Is Going Next described, but products tend to spend those savings on harder tasks, longer reasoning and more steps4.

What is new for a product leader is not that cost exists but that it varies by customer, and the customer controls it. A seat sold to a light user and a seat sold to a power user cost the same in a classic software business. In an AI product the second can cost many times the first. A flat price with unlimited use therefore attracts exactly the users it cannot afford. That is what happened to ChatGPT Pro.

Outcome pricing has its own version of the trap. If you charge only when the work succeeds, you still pay for every attempt that fails.

An illustrative outcome price of 100 falls to a gross margin of 40 after model costs, human review, unbilled failed attempts, and support.100 per 100 of pricePrice percompleted case−15 per 100 of priceModel and tool calls−20 per 100 of priceHuman review−15 per 100 of priceFailed attemptsnot billed−10 per 100 of priceSupport and platform40 per 100 of priceGross marginILLUSTRATIVE NUMBERS
Figure 10.6.3 Illustrative figures: under outcome pricing, the attempts that fail are still paid for, by the seller.

The numbers above are illustrative, invented to show the shape rather than taken from any firm. In a product like this, the cost of human review and of failed attempts can matter more than the model price, and both depend on how reliable the product is on the customer’s real work. Reliability is not only an engineering metric. It is the margin.

Choosing the unit you charge for

Every price has a unit: a user, a request, a task, a result. There are four broad choices, and most real AI products combine them.

Four pricing units - seat, usage, outcome and hybrid - with what the customer pays for, when each fits and its main risk.UnitCustomer pays forFits whenMain riskSeat orsubscriptionAccess, per user or accountUsage and cost are predictableHeavy users erode marginUsage or creditsRequests, actions or creditsCost varies with useBills feel unpredictableOutcomeA defined resultThe result is countableand attributableDisputes; failures still costHybridBase fee plus usage or outcomesCost and value are mixedComplex to explain
Figure 10.6.4 No unit is best in general. The right one depends on how variable your cost is and how measurable your customer’s result is.

Two questions narrow the choice. First, how much does your cost to serve vary between customers? Second, can the customer’s result be counted, and credited to your product rather than to everything else that happened?

A two-by-two of cost variability and outcome measurability, pointing to subscription, subscription with targets, usage pricing or outcome pricing.CountableHard to countOutcomePredictableCost to serve · Varies by customerSubscription plus targetsFlat fee with service levelsOutcome pricingPay per defined resultSeat or subscriptionClassic software termsUsage or creditsRevenue tracks cost
Figure 10.6.5 Start from your cost behavior and the customer’s measurability, then design the hybrid around them.

Outcome pricing is the most discussed option and the hardest to run. Intercom charges 0.99 dollars per outcome for its customer support agent. The definition is where the business model lives. An outcome counts when the customer confirms the issue is resolved, when the customer does not ask for more help after the agent responds, or when the agent completes a configured workflow, including a handoff to a person5. Each of those is a reasonable proxy, and each is a negotiation. Before you sell outcomes, write down what counts, who measures it and what happens when the customer disagrees.

Even very large software companies are still searching. Salesforce, which did as much as anyone to make per-user subscription the industry standard, has changed the unit for its agent product twice in under two years, from 2 dollars per conversation to credits per action to per-user editions with bundled credits67. That is not indecision. It is a seller learning which unit its customers can budget for and which one tracks its own cost. Expect to do the same, and design contracts and billing systems that let you change the unit without renegotiating every customer.

The input-pricing trap

There is a quieter problem for any business that charges for inputs. If you price per seat and your product lets one person do the work of three, your best customers will buy fewer seats. If you bill by the hour and AI halves the hours, your revenue halves with them. Pricing by the input punishes you for the product’s success.

The exits run in both directions. You can price the output instead and accept the cost and measurement risks above. Or you can keep a predictable fee and change what it buys: more volume, a service level, a guaranteed turnaround. What you cannot do for long is keep selling the drill by the hour while the customer can see the hole being made in minutes.

What defends an AI product

If every competitor can rent a model of similar quality, the model is not the defense. The Evolution of AI Models showed how quickly leading models converge. A product that is a third-party model behind a thin interface is exposed twice: rivals can copy it in weeks, and the model provider can change price or policy, or ship the same feature itself.

Chegg, whose product was the answer itself, is the cautionary example: its shares fell 48 percent in a day in 2023 after it said ChatGPT was hurting growth, and it cut 45 percent of its workforce in 202589. When what you sell is an answer a general model gives away, model access defends no one.

What lasts is what a rival cannot rent. Above the rented model sit four layers, from the most copyable to the least.

Defensibility rises from the rented model through proprietary context, workflow position and distribution to trust and accountability.Trust andaccountabilityA name that signs for the resultSlowest to buildDistributionCustomer relationships, channels, brandEarnedWorkflowpositionEmbedded where the work and records liveStickyProprietarycontextData and history rivals cannot buyMust be usableThe modelRented, converging, improving for everyoneTable stakesHARDERTOCOPY
Figure 10.6.6 The model is the floor of the stack, not the moat.

Proprietary context helps only when it is usable, relevant and linked to outcomes, as Proprietary Data, AI Moats and Differentiation set out in its seven questions. Workflow position means the product sits where the work happens and where records are kept, so leaving means rewiring the customer’s process. Distribution still decides many markets: a better product with no route to customers loses to a good one with a sales force. Trust is the deepest layer. When a product does consequential work, the customer is buying someone who stands behind the result.

Dependency on a model provider is the mirror image of this stack. Model and Third-Party Risk covers how to keep a tested way out; for a product, that also means a pricing model that survives a change in the provider’s price list.

Trust is part of the product

Selling the hole means answering for the hole. Customers need to see what the AI did, what it recommends, what a person can change and what it will do on its own. The law is moving the same way.

The wider landscape of regulation and national strategy is the subject of the next chapter. For a product leader the point is narrower: a claim, a disclosure and a liability position are now part of the product specification.

Story: the repair reports that billed by the hour

This is an illustrative composite, not a single firm. Picture an aerospace supplier of a few hundred people that makes and overhauls landing-gear components. A steady part of its revenue comes from repair assessments: when a damaged part comes back from a maintenance shop, the shop needs a written decision on whether and how it can be repaired. Each one is bespoke. An engineer gathers the part’s drawings, inspection results and service history, runs a stress analysis where the damage needs it, and writes a long report that a senior engineer signs. The firm bills by the hour, and a typical assessment takes about three weeks.

In 2025 the team starts using AI to assemble the records and draft the standard sections. Drafting time drops sharply. So does fee income from the work, because the firm bills hours. The managing director names the contradiction at a board meeting: the better the tools work, the less the firm earns.

The board considers three choices. Keep hourly billing and say nothing about the efficiency. Offer the same reports faster at a discount. Or turn the service into a product.

Before, the client bought engineer hours at an hourly rate; after, a signed assessment at a fixed price by complexity band, with AI drafting and engineers signing.ElementBeforeAfterWhat theclient buysEngineer hoursA signed assessmentPrice unitHourly rateFixed price by complexity bandAI doesNothing formalGathers records; drafts; screens damageEngineers doEverythingAnalysis judgments; sign every reportUnusual damageSame processRouted to bespoke work, billed hourlyDefenseReputationRecords, signature, authority relationships
Figure 10.6.7 Illustrative composite: the firm stopped selling hours and started selling a signed result, with people on every report.

They chose the product, and designed its economics before its interface. Maintenance shops upload the inspection data and get a screening result within a day and a signed assessment within a week. The price is fixed by complexity band, because stress analysis is the variable cost and it depends on the damage. One round of queries from the shop’s airworthiness reviewers is included, which is the outcome promise the firm can actually measure. Damage that fails the screening rules goes to bespoke work at hourly rates; that is the exception path from the previous chapter. Every report is still signed by an authorized engineer.

The defense was never the model, which any competitor could rent. It was three decades of drawings and repair histories, checked first for the right to reuse them; the engineer’s signature and professional insurance; and the firm’s relationships with the aviation authorities. Counsel was asked whether the self-service screening tool counts as software under the revised Product Liability Directive. The lesson the directors drew was simple: the AI changed the cost of the work, and that forced them to decide what they had been selling all along.

What this means for leaders

For any product with AI inside, ask what the customer is really buying, and whether your price and your defense match that answer. The failures this chapter has described come from a mismatch: an enhancement priced as a revolution, an outcome sold without a definition, or a thin layer over someone else’s model presented as a moat.

Check yourself

  1. Adding a generative AI feature makes a product AI-native.
  2. A flat subscription with unlimited use can lose money on the customers who like the product most.
  3. Outcome pricing removes the seller’s cost risk.
  4. Per-seat pricing can shrink revenue as an AI product succeeds.
  5. Access to a strong model is a durable competitive advantage.
  6. In the EU, software including AI systems is covered by no-fault product liability for products placed on the market from December 2026.

Reflection: what are you really selling?

What comes next

Products and business models do not compete in a vacuum. What you may build, where you may sell it and how secure your supply of chips and models is all depend on rules and rivalries well beyond your market. The next chapter, AI, Regulation, Geopolitics and Global Competition, widens the lens to those forces.

Laws referenced

Not legal advice. Laws change; verify before relying on this, and consult counsel for decisions.

EU AI Act · EU

Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744

Risk-based rules. Prohibited practices include social scoring, untargeted scraping of facial images, and emotion recognition in workplaces and schools (with narrow exceptions). High-risk systems (Annex III: biometrics, safety components of critical infrastructure such as energy, water and traffic, employment and worker management, credit, education, essential services, law enforcement, migration, justice) need risk management, data governance, documentation, logging, human oversight, human oversight that keeps people able to understand the system, notice automation bias (over-reliance on its output), override it or stop it (Art. 14(4)), appropriate accuracy, robustness and cybersecurity (Art. 15), automatic logging of events (Art. 12), a provider quality-management system (Art. 17) and conformity assessment. An Annex III system is not high-risk if it poses no significant risk of harm, for example a narrow procedural or preparatory task that does not replace human assessment; systems that profile people are always high-risk, and a provider relying on this exception must document it and register (Art. 6(3)). Deployers of high-risk AI must use it as instructed, assign competent human oversight, monitor its operation, keep logs for at least six months and report serious incidents (Art. 26); employers must inform workers' representatives (Art. 26(7)). Public bodies, private providers of public services, and deployers of credit-scoring or life and health insurance pricing systems must carry out a fundamental-rights impact assessment before first use (Art. 27). Providers must run post-market monitoring (Art. 72). A deployer that puts its name on a high-risk system, substantially modifies it, or changes its purpose so that it becomes high-risk takes on the provider's obligations (Art. 25(1)). A substantial modification (Art. 3(23)) of a high-risk system needs a new conformity assessment, unless the change was pre-determined and documented at the first assessment, as with planned continuous learning (Art. 43(4)). Providers of general-purpose AI models (from 2 Aug 2025) must keep technical documentation, have a policy to comply with EU copyright law including text-and-data-mining opt-outs, and publish a sufficiently detailed summary of training content (Art. 53). Research, testing and development before a system is placed on the market or put into service is outside the Act, except testing in real-world conditions (Art. 2(8)). Since the 2026 Omnibus, the Art. 4 AI-literacy duty is an obligation of effort (take measures to support literacy), not of result. Fines reach EUR 35 million or 7% of global turnover for prohibited practices.

  • 2024-08-01 — Entered into force
  • 2025-02-02 — Prohibited practices (Art. 5) and the AI-literacy duty (Art. 4) apply
  • 2026-07-27 — Omnibus softens Art. 4: providers and deployers must take measures to support AI literacy; no specific level must be guaranteed
  • 2025-08-02 — General-purpose AI model obligations apply; governance and penalties regime in place
  • 2026-08-02 — Transparency duties (Art. 50) apply: disclose AI interaction, label synthetic and deepfake content (marking for generative systems already on the market: 2 Dec 2026)
  • 2027-12-02 — High-risk obligations for Annex III systems (e.g. hiring, credit, education, essential services) - moved from 2 Aug 2026 by the 2026 Omnibus
  • 2028-08-02 — High-risk obligations for AI in products regulated under Annex I

Last verified 2026-10-06 · official text

EU Product Liability Directive (revised) · EU

Directive (EU) 2024/2853

No-fault liability now explicitly covers software, including AI systems and SaaS, and updates or the lack of security updates. Easier proof for claimants with complex products.

  • 2026-12-09 — Applies to products placed on the market from this date

Last verified 2026-10-06 · official text

US enforcement against false AI claims ("AI washing") · US - federal

Federal securities antifraud rules and Investment Advisers Act Marketing Rule (SEC); FTC Act Section 5 (unfair or deceptive practices)

There is no AI-specific federal statute, but existing law already applies to what companies say about AI. Regulators have penalized firms that overstated their use or capability of AI to investors (SEC) and to consumers (FTC). Claims about AI in marketing, investor materials and product descriptions need the same substantiation as any other claim.

  • 2024-03-18 — SEC's first AI-washing cases: Delphia and Global Predictions settle for USD 400,000 in total civil penalties
  • 2024-09-25 — FTC launches Operation AI Comply, with five actions over deceptive AI claims and uses

Last verified 2026-10-08

References

  1. TechCrunch. OpenAI is losing money on its pricey ChatGPT Pro plan, CEO Sam Altman says. TechCrunch. 2025.
  2. Clayton M. Christensen, Scott Cook and Taddy Hall. Marketing Malpractice: The Cause and the Cure. Harvard Business Review, 83(12), 74-83. 2005.
  3. Martin Casado and Matt Bornstein. The New Business of AI (and How It's Different From Traditional Software). Andreessen Horowitz. 2020.
  4. Ben Cottier, Ben Snodin, David Owen and Tom Adamczewski. LLM inference prices have fallen rapidly but unequally across tasks. Epoch AI. 2025.
  5. Intercom. Pricing. Intercom. 2026.
  6. Salesforce. Salesforce Introduces New Flexible Agentforce Pricing to Accelerate the Digital Labor Revolution. Salesforce (press release). 2025.
  7. Salesforce. New Salesforce Editions Bundle Everything Businesses Need for Agentic Transformation. Salesforce (news). 2026.
  8. CNBC. Chegg CEO calls 48% stock plunge over ChatGPT fears 'extraordinarily overblown'. CNBC. 2023.
  9. CNBC. Chegg slashes 45% of workforce, blames 'new realities of AI'. CNBC. 2025.
  10. European Parliament and Council of the European Union. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. 2024.
  11. European Commission. New EU product liability rules will apply to online platforms and software from December 2026. European Commission (Transition Pathways). 2025.
  12. US Securities and Exchange Commission. SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence (Press Release 2024-36). SEC. 2024.
  13. US Federal Trade Commission. FTC Announces Crackdown on Deceptive AI Claims and Schemes (Operation AI Comply). FTC. 2024.

Further reading

Sources last verified 2026-10-08.