So Your Business Is An AI Wrapper. Is It Doomed?

A familiar story

Picture an established business that has rebuilt its operations around AI agents. The work is impressive, customers like it and the board is proud of it. Then the AI bills start arriving, and they are well above budget. Spending controls tighten, usage gets rationed and someone asks the obvious question: what happens when tokens get more expensive?

It's a story playing out in a lot of businesses right now, and the surveys below suggest it is closer to the norm than the exception.

Many founders, boards and investors are now asking some version of that question, usually quietly: can a business that depends on AI for almost everything stay viable over the long term? It's a fair question, and the honest answer depends on several things most businesses aren't yet measuring. So let's explore that question and find the answers together.

Three ways to use AI

Most businesses using AI fall into one of three broad types.

Decorative. AI makes the existing business faster, cheaper or better, but what the customer buys doesn't change: drafting proposals, triaging support tickets, helping developers write code. AI is an overhead. If prices doubled tomorrow, the IT budget would get worse but the business model would survive. This is often exactly the right choice, especially where customers are buying expertise, relationships or accountability. It becomes the wrong choice when competitors use AI to change what customers can buy, not just what it costs to deliver.

Foundational. AI makes possible an offer that couldn't exist before, or changes how a service is delivered. The value sits in what's built around the model: the workflow, proprietary data, integrations, trust and distribution. AI is now part of the cost of goods.

Wrapper. The model is most of the product: prompts, a user interface and some tuning on top of a frontier model, aimed at a specific job. It's fast to build, can do things that weren't possible before and is quick to copy.

What this looks like in practice

These classifications are our own view, based on public information. Businesses also move between types over time.

• Decorative: Klarna, and the law firms using Harvey. Klarna's AI assistant, built with OpenAI, handled two-thirds of its customer service chats in its first month, doing the equivalent work of 700 full-time agents (Klarna). That is a big efficiency gain, but customers still come to Klarna for payments, not for the chatbot. Similarly, more than 3,500 legal organisations use Harvey (Harvey), but their clients are still buying legal advice and the firm's judgement. If AI prices doubled, margins would suffer but neither offer would change.
• Foundational: Fin, formerly Intercom. Intercom launched its Fin AI agent in 2023, then rebuilt its product, infrastructure and team around AI, and eventually renamed the company after the agent (Fin). Fin charges US$0.99 per resolved outcome (Intercom), so the AI cost sits inside the price of every result. The company passed US$400 million in annual recurring revenue, and Salesforce agreed to buy it for about US$3.6 billion (Silicon Republic).
• Wrapper: Jasper. Jasper built an AI copywriting tool on OpenAI's models and raised US$125 million at a US$1.5 billion valuation in October 2022 (TechCrunch). Six weeks later, OpenAI released ChatGPT, free (OpenAI). Within a year Jasper had cut jobs and refocused on enterprise marketing teams (Austin Business Journal). Its supplier had become its competitor.

The most interesting businesses are the ones changing type. Cursor began as a coding tool built on other companies' models and, like many such tools, ran at negative gross margins. After launching its own Composer model, it reached slight gross margin profitability (TechCrunch), and SpaceX completed a US$60 billion acquisition in August 2026 (TechCrunch). Fin, too, has moved to its own purpose-built model (Fin).

The business in the opening story isn't a wrapper. Like Fin, it's an established company that has gone AI-native, and it still hits the cost problem. You don't need to be a wrapper to have a wrapper's cost problem. You only need AI in your cost of goods.

Is AI getting cheaper or more expensive?

Ask around and you'll hear confident answers both ways. So let's weigh the evidence on each side.

The case that AI is getting cheaper is strong. Epoch AI estimates the cost of reaching a given level of AI performance has fallen about 47% a quarter since 2023, roughly 13 times a year (Epoch AI).

The case that it's getting more expensive is also strong.

• Frontier prices are rising. Google's own pricing page shows Gemini 4 Argon's "introductory" price doubling on 1 December 2026, from US$2 to US$4 per million input tokens and from US$10 to US$20 for output. Two Gemini Flash models double on 1 January 2027 (Google). One price tracker found list prices rose on 10 of 14 model tiers in the year to October 2026 (Synthor).
• Flat-rate deals are being unwound. GitHub moved every Copilot plan to usage-based billing, explaining that "a quick chat question and a multi-hour autonomous coding session can cost the user the same amount" and that the old model was "no longer sustainable" (GitHub).
• The industry's economics don't add up yet. AI providers are expected to spend nearly US$1 trillion on AI infrastructure in 2026. Gartner estimates that data-centre owners need US$8.5 trillion in cumulative revenue by 2030 to justify the spending, against no more than US$3.6 trillion of attributable AI revenue (Gartner). OpenAI's own plan, as reported, shows US$278 billion of negative free cash flow from 2026 to 2030 (Reuters).

At some point that gap has to close, and raising prices at the frontier is one obvious way to close it.

There are two reasons not to assume prices will rise across the board.

• Selling AI may already be profitable. Anthropic's prospectus implies an 85% gross margin in the first half of 2026, up from 12% a year earlier (CNBC). On those numbers, the losses come mostly from training the next generation of models and racing for capacity, not from selling tokens below cost. Slowing that race is another route to profit.
• Overbuilding usually pushes prices down. The closest parallel is the telecom fibre glut. In the late 1990s, telecom companies borrowed heavily to lay fibre-optic cable, betting that internet traffic would keep exploding. By April 2001, less than 3% of US long-distance backbone capacity was in use (Federal Reserve Bank of Dallas). Bandwidth prices fell sharply, with Morgan Stanley estimating a fall of 60% in 2001 on top of a similar fall the year before (Deseret News). Global Crossing, once worth more than US$50 billion, went bankrupt owing US$12.4 billion and sold majority control for US$250 million (Los Angeles Times). Buyers picked up whole networks for cents on the dollar, and over the following decade that cheap capacity carried streaming, cloud computing and the modern web (AI Bubble Question). The investors lost; the customers won. The difference this time is that GPUs lose most of their value within three to five years, so there won't be a decade-long glut to inherit.

Weighing it all up, the most useful way to think about it is that there isn't one price of AI. There are three:

1. Yesterday's capability keeps getting cheaper, and is heading towards a commodity.
2. The frontier resets with each new release and is now rising.
3. Heavy use on flat-rate plans is where the subsidy is ending, through caps, credits and usage billing rather than headline price rises.

A business that needs the newest model for everything, and charges its customers a flat fee, is exposed to the second and third. On the face of it, a business built on the first price has a strong tailwind. But that depends on yesterday's model still being there.

Will yesterday's model still be supported?

Cheaper older models only help if you can keep using them, and increasingly you can't, at least not for long.

• Anthropic retires Claude Sonnet 4.5 on 30 November 2026, about 14 months after its September 2025 release, with 60 days' notice, its stated minimum. Claude Sonnet 4 and Opus 4, released in May 2025, were retired in June 2026 (Anthropic).
• OpenAI removes GPT-5 and o3 from its API on 11 December 2026, about 16 months after GPT-5 launched. GPT-4o and GPT-4.1 follow on 14 January 2027, along with every fine-tuned model built on them. OpenAI commits to at least six months' notice for generally available models (OpenAI).

And the replacement is rarely like for like. Researchers at Stanford and Berkeley found that the accuracy of the "same" GPT-4 service on one task fell from 84% to 51% between March and June 2023, as the model was updated behind the scenes (arXiv).

For a wrapper, where the model is most of the product, that is shifting sand. Every forced move to a new model means re-testing prompts, re-checking outputs, fixing what has changed and sometimes explaining to customers why the answers look different. And it happens every year or so, whether you're ready or not.

So the falling cost of yesterday's capability helps less than it first appears. In practice, the capability gets cheaper because a new model arrives at a lower price point, not because the model you built on gets cheaper and stays put. You still pay a migration cost every time. There are two ways to shrink it: build an automated test suite of real customer cases so that a model change becomes routine, or run open-weight models you host yourself, which no one can retire. That second option brings its own costs, which we'll come to in the next Insight.

Most blowouts are driven by excess token consumption, not price increases

AI providers bill by the token, a unit of text roughly three-quarters of a word long. Every question, every document an agent reads and every answer it writes is counted in tokens. And the story at the start of this piece is far from unusual.

• Spending is running well ahead of plan. Menlo Ventures found the average enterprise planned to grow AI spending 1.9 times in 2026 and actually spent 4.7 times its 2025 level, about 2.5 times its budget (Menlo Ventures).
• Overruns are the norm. In IDC's July 2026 survey, 67% of enterprises overran their AI agent budget by more than 10% in the previous year, and 24% ran significantly or extremely over (IDC). McKinsey's smaller survey found 93% of respondents had exceeded their AI budgets (McKinsey).
• Many projects won't survive the bill. Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls (Gartner).

Almost none of this is caused by price rises, which have only just started. Most of it comes from consuming far more tokens than planned, through usage and design:

• agents that loop, retry and call tools without limits
• several agents passing large amounts of context to each other
• voice, images and video, which cost far more than text
• frontier models used for routine work that a smaller model, or ordinary software, could handle
• customers paying a flat fee while usage grows without limit

Buyers are also leaning harder on the most expensive models. Between Q3 2025 and Q3 2026, frontier models' share of enterprise AI spending rose from 35% to 63%. Open-source models handle 19% of usage but take only 4% of spending (Menlo Ventures).

So the real risk isn't only that prices rise. It's that usage grows unchecked first, and then rising frontier prices land on top.

"It's just like relying on AWS" is only half true

The comparison with the cloud is the strongest counter-argument, and it deserves a fair hearing. Almost every SaaS business depends on infrastructure it doesn't control, and that has turned out to be a very successful business model. The cloud market is competitive, and switching providers is painful but possible.

Three things make AI different.

The rent is a much bigger share of revenue. The median private B2B SaaS company spends about 5% of annual recurring revenue on hosting (Gart Solutions, citing SaaS Capital). ICONIQ found AI-native companies run gross margins roughly 15 points lower than their peers: 56% against 70% for companies under US$100 million in annual recurring revenue, and 63% against 80% above it (ICONIQ). A 20% rise in hosting costs is an irritation. A 20% rise in the cost of running AI models can wipe out a year's margin improvement.

The landlord opens a shop next door. On 17 September 2026, OpenAI launched Astra for Law, its most powerful model combined with a legal research index, offered directly to law firms. Legal AI companies Harvey and Legora, which build on OpenAI's models, will also be able to build on it (Reuters). A day earlier, Anthropic launched Claude Docs and Claude Slides (Reuters). Cloud providers have competed with their customers too, but AI labs are doing it faster and closer to the product that customers actually use.

The lease terms keep changing. Models are retired, behaviour changes between versions, rate limits tighten and "introductory" pricing ends.

So the landlord metaphor needs updating. Rent on the older shopfronts is falling. Rent on the best shopfront in the street is rising. And the landlord is also opening shops in the same street.

Starting, running and investing in each type

A wrapper can still build a moat. It just won't be the technology, because anyone can call the same model. The moat has to be something that is extremely hard to replicate:

• Distribution. A loyal customer base, a community or a channel that competitors can't easily reach. In our view, Cursor's edge was never exclusive access to a model, since rivals could use the same ones. It was the loyalty of the developers who adopted it.
• Strategic relationships. Exclusive content, data or partners. Harvey, which builds on frontier models, formed a strategic alliance with LexisNexis to bring LexisNexis case law, statutes and citation checking into its platform (LexisNexis). When OpenAI launched its own legal platform, that content was something it couldn't simply copy.
• Trust and accountability. Regulatory approvals, certifications, or a willingness to stand behind the outcome in a way a general-purpose model provider won't.
• Embedded workflow and switching costs. Integrations, customer configuration and data that build up over time and are painful to rebuild elsewhere.

How to reduce the risk

• Route each task to the cheapest model that does it well. McKinsey's advice is blunt: "Few tasks warrant the use of expensive frontier models" (McKinsey). ICONIQ reports that many builders now send most of their work to smaller or fine-tuned models and use about 3.1 model providers on average (ICONIQ).
• Cache what repeats. Reusing the same context can cut repeated input costs by up to about 90%, and companies that actively manage their AI usage typically save 20 to 30% (McKinsey).
• Don't use a language model where ordinary software will do. Forecasting, scoring and routing rules have been handled cheaply by conventional software and machine learning for years.
• Put limits on agents. Cap retries, tool calls and loops, and give every workflow a budget and an owner.
• Know your cost per outcome, especially for your heaviest 10% of users. If you can't name it, you don't know whether your AI product makes money.
• Price so that costs pass through. ICONIQ found companies moving customers to consumption pricing are seeing margins improve (ICONIQ). Avoid promising "unlimited AI". It's an expensive promise to walk back.
• Keep a second provider tested and ready. If switching would take months, you don't have a fallback.
• Build a test suite before you need it. A set of real customer cases, run automatically against any new model, turns a forced migration from a crisis into a routine release.
• Read the contract. Look for price protection, notice periods before models are retired and sensible volume commitments. The telecom bust is a reminder that capacity commitments signed at the peak of a boom can age badly.

Where the opportunity is

None of this is an argument against building on AI. The value is shifting to the application layer: 49% of companies building AI products say their main differentiation comes from the application layer, such as user experience, workflow, integrations and how they use data, compared with 14% who rely mainly on their own models (ICONIQ).

The businesses best placed to last own something the model can't replicate: proprietary data, a workflow customers rely on, deep integrations, trust and compliance, distribution, or accountability for an outcome. They're also designed so that the falling cost of capability works in their favour, with the testing in place to switch models without drama.

Questions for boards and investors

1. The deletion test. If your model provider disappeared tomorrow, what would you still have?
2. The next-release test. If the next frontier model did your core job natively, would customers still need you?
3. The rent ratio. What share of revenue goes on AI usage, meaning what you pay model providers every time your product uses AI, and is it rising or falling?
4. The doubling test. Re-run the margin model with frontier prices doubled, as Google is doing on 1 December. Does the business still work?
5. The heavy-user test. What do your heaviest customers cost to serve compared with what they pay?
6. The switching test. How many model providers have you actually run in production?
7. The migration test. When your main model is retired, how long will it take to move, and how will you know nothing has broken?
8. The ownership test. Who in the business owns AI spend, and how often do they report on it?

Seven myths, busted

1. "AI will keep getting cheaper." Half true. The cost of a given capability will keep falling, but the model you built on will be retired, and the frontier and heavy use aren't getting cheaper.
2. "AI costs are about to balloon." Also half true. For most businesses, the bill is ballooning because of usage, not price.
3. "Wrappers have no future." Many valuable AI businesses started as wrappers. What matters is what they build next.
4. "Decorative AI is a missed opportunity." Often it's the right answer.
5. "AI-native businesses are safer than wrappers." They carry the same cost exposure, plus the cost of building everything around the model.
6. "It's just like relying on AWS." Only if AI usage costs are as small a share of your revenue as hosting is for SaaS, and the cloud provider isn't selling your product.
7. "You need your own model to be defensible." Only 14% of AI builders say their main differentiation comes from proprietary models (ICONIQ).

So, is it doomed?

Not necessarily. A business wrapped around AI can be very viable over the long term, but only if it is built with the economics in mind. That means knowing its AI cost as a share of revenue, using the newest models only where they earn their keep, pricing so that heavy users pay their way, treating model changes as routine rather than a crisis, and building something around the model that a supplier can't simply copy. The businesses that do this have one of the strongest tailwinds in technology behind them. The ones that don't are running on a subsidy that is already being withdrawn.

Next: should you own the stack?

Some businesses are responding to cost blowouts by investing in their own AI hardware, models and infrastructure. Is that the answer, or simply a different way to take on the same risk? That's the subject of our next Insight.

Rev-IQ Advisory works with technology businesses on growth, pricing and commercial performance, including how AI changes the economics.

This article is general information only, not financial or investment advice. The classification of named companies as decorative, foundational or wrapper businesses is Rev-IQ's own view, based on public information. Model prices, retirement dates and vendor terms change often; check current terms before relying on them. Sources checked 13 November 2026.