On May 7, 2026, OpenAI sent a note to every developer using its self-serve fine-tuning API. Organizations that had not already started a training job could no longer create one. Existing customers keep access until January 2027, at which point the door closes. The company framed the shutdown as a rationalization: newer base models are strong enough that fine-tuning is less often needed, and the ones who genuinely need custom work will move to enterprise contracts.
Four months later, on September 11, Salesforce named seven agents. Casey handles customer service. Paige handles IT and HR. Carter handles commerce. Hunter, Marshall, Piper and Fin cover outbound sales, supply chain, inbound pipeline and customer experience. The naming was the announcement. Software vendors have historically named products for what the product does. Salesforce named these products for who they replace. Two weeks earlier, on August 26, Salesforce and Anthropic had announced Claudeforce, a plugin bringing thirty-seven prebuilt sales skills into Claude with governed access to Salesforce data. Dreamforce 2026 opens tomorrow at the Moscone Convention Center and runs through September 17.
Somewhere in the middle of all this, Anthropic's revenue run rate crossed sixty-five billion. It was nine billion at the end of 2025. Fourteen billion in February. Nineteen billion in March. Thirty in April. Forty-seven in May. Sixty-five in July. Q2 delivered the company's first quarterly operating profit, roughly $559 million. One thousand enterprise customers now spend more than one million dollars a year each.
Read those three headlines together and the executive framework that every enterprise strategy deck opened with for four years just stopped making sense.
The Framework That Broke
The framework said: for any AI capability, choose between building your own model and buying a vendor product. Buy when the workflow is commodity and the vendors are mature. Build when the capability is core and the data is proprietary. McKinsey wrote it. Forrester wrote it. Bain wrote it. It is the first slide in every when-to-build-vs-buy-AI deck ever presented to a board.
The framework had a hidden assumption. It assumed the "build" column meant owning the model weights, and it assumed the "buy" column meant licensing a fixed piece of software. Both assumptions have quietly stopped being true.
The build column got hollowed out from the top. OpenAI's self-serve fine-tuning shutdown is the visible piece, but Anthropic never sold fine-tuning to the general market at all. Their position has always been that unrestricted fine-tuning degrades Constitutional AI properties and introduces safety regressions they cannot underwrite. Anthropic offers custom model work only under committed enterprise contracts through its professional services team. Bedrock still exposes some managed fine-tuning against Claude models, but the surface is deliberately narrow.
So what did "build" really mean? It meant renting a base model from OpenAI, adjusting it with your data, and running inference against your adjusted copy. You did not own the base weights. You did not own the training method. You could not port the fine-tune to another provider. Your custom model was a rented apartment with your furniture in it. The landlord just announced they are not renewing the lease.
The buy column got hollowed out from the bottom. When Salesforce introduces Casey, Paige and the rest, the pricing runs per-outcome: resolved tickets, closed opportunities, dispatched shipments. That contract sells labor equivalence, unit by delivered unit. When the vendor prices against your outcome, they capture the value of your workflow data. Every ticket Casey resolves becomes a training example for the next version of Casey that Salesforce sells to your competitor.
The framework asked you to pick a column. Both columns collapsed into something else.
What Actually Lives In The Middle
Between the model provider and the shrink-wrapped agent, there is a layer that most enterprise CIOs are only now starting to recognize as a distinct thing. Call it the composition layer. It contains everything that translates general model capability into your specific business.
The composition layer includes the data adapters that connect your systems of record to the model's context window. The retrieval pipelines that fetch the right rows from your data warehouse before the model reasons over them. The evaluation harness that grades model output against your definition of correct. The governance layer that logs every call, redacts what needs redacting, and refuses what needs refusing. The agent orchestration that decides which model handles which subtask, what tools it can call, and when to escalate to a human. The prompt library, the guardrails, the audit trail, the kill switches.
None of this arrives as a shrink-wrapped product. The composition is code, configuration, taxonomy, and policy that only makes sense against your specific business. Every part survives when the underlying model gets swapped, upgraded, or replaced.
The when-to-build-vs-buy-AI question dissolves into a different question: how deep does your composition layer go? A shallow composition layer means dependency on the vendor's understanding of your business. A deep composition layer means the model is a commodity input, swappable at will, and the value lives in what you built around it.
Anthropic's growth pattern is the receipt. The company doubled from May to July, most of the growth from enterprise contracts. The reason enterprises are paying is that Anthropic's Model Context Protocol, released in November 2024 and steadily expanded through 2026, standardized how a model connects to enterprise data. MCP is a composition-layer standard. Anthropic's revenue is the composition tax, paid to the vendor whose interface is easiest to compose against.
Salesforce read this correctly. That is why the Claudeforce plugin is a composition play, not a model play. Salesforce provides the data, the workflow logic, and the governance. Anthropic provides tokens. Neither company is trying to sell you a fine-tuned model.
The Trusted Enterprise AI Harness As Confession
Read the Salesforce September 11 announcement carefully and there is a confession embedded in it. The Trusted Enterprise AI Harness has six pillars: Trusted Context, Trusted Agency, Trusted Action, Trusted Governance, Trusted Security, Trusted Models. Underneath sits an AI Control Plane that registers agents from any vendor, sets identity and access policy, manages their lifecycle, evaluates their performance, observes their behavior, and tracks their cost.
Salesforce is describing a way to run any vendor's agents on Salesforce data. The Harness is a governance layer over a multi-vendor agent world. Which means Salesforce, one of the loudest voices selling shrink-wrapped agents this year, has already conceded that no enterprise is going to be a single-vendor agent shop.
That concession is worth more than the seven named agents. The buy-column-only strategy is not what Salesforce itself expects its customers to run. The seven agents are decoys. The Harness is the actual product. And the Harness is composition tooling with a Salesforce logo on it.
Take Casey, Paige, Carter, Hunter, Marshall, Piper and Fin at face value and you have bought the wrapper. You have handed Salesforce your workflow data, your outcome definitions, and your competitive playbook. Take the Harness at face value and you are being sold the observation deck for a composition layer that Salesforce hopes to own. Different bet. Same vendor.
Why The Fine-Tune Door Closing Was Never About Fine-Tuning
The OpenAI shutdown announcement in May was framed as a rationalization: base models are so good that most fine-tuning is unnecessary. That framing is partially true and largely misleading.
The real reason to stop offering self-serve fine-tuning is that fine-tuning is a bad business for the lab. Every fine-tuning job is compute the lab burns on a specific customer's data, in exchange for a training fee that runs much smaller than the inference revenue that customer would generate over the same period on the base model. Fine-tunes also create support burden: a customer whose fine-tuned model regresses will call their sales rep, not their intern. Fine-tunes create switching cost against the lab, because a customer who has invested in a fine-tune becomes harder to migrate onto a newer base model.
Every one of those problems disappears if the lab kills fine-tuning and points customers at longer context windows, better retrieval, and more capable base models. The revenue moves from training fees to inference. The support burden shrinks. The upgrade path is trivial. And the switching cost from lab to lab, which fine-tuning used to hold in place, evaporates.
The OpenAI shutdown is a business decision. The model layer wants to sell pure tokens by the million, fresh from a frontier model, forever. The labs are commoditizing themselves upward. They want your money paid in inference, on the frontier model, month after month.
Which means the last vestige of "building your own AI" for anyone below hyperscaler size just left the menu. And it means the composition layer is now the only layer where an enterprise can build real durable difference.
The Consulting Ask That Broke My Week
Two weeks ago, a mid-market operations executive at a US insurance carrier asked me for a build-vs-buy assessment on claims triage. The internal team had spent eight months getting quotes from four vendors, plus a scoped statement of work for an internal build using a Databricks endpoint against Anthropic on Bedrock. The vendor quotes clustered between $180,000 and $410,000 a year in seat-and-usage license fees. The internal build carried a $1.2 million capex proposal and a $340,000 annual run rate.
I asked one question: what does your claims triage output look like six months from now, if the underlying model gets three times better at half the cost? The answer was that nothing in either proposal contemplated that trajectory. The vendor products would upgrade behind the paywall and the customer would keep paying the same price. The internal build had no evaluation harness that would even detect the improvement, let alone route to it.
The correct answer sat outside both columns on the spreadsheet. It was a composition project: a data adapter into their claims system, a retrieval pipeline over their policy documents, a governance layer that met New York and California regulatory requirements for adverse action, and an evaluation harness that would grade model output against senior adjuster judgment. Build the composition, rent the model, treat the model as fully replaceable inside eighteen months. The total came to $340,000 of consulting and roughly $60,000 a year in inference, with the entire model spend on a variable line that shrinks as the labs commoditize.
The executive team's first reaction was that the number felt wrong. It was too small. Neither the internal team's capital request nor the vendor quotes had prepared them to spend under six figures on this problem. But that was the number. The composition was the whole product. Everything else was rent.
What Deep Composition Looks Like In Practice
A deep composition layer is an organizational asset that gets more valuable over time, independent of what the model providers do. Its parts are boring individually and essential together.
Start with data adapters. Not vendor connectors. Adapters that translate your specific tables, columns, and business logic into a shape a model can reason over. This work looks like ontology work, because it is ontology work. You are defining the vocabulary the model uses to talk about your business. When the vocabulary is precise, the model's outputs become usable. When the vocabulary is loose, no amount of prompting saves you.
Then retrieval pipelines. Deciding which fifty rows out of a fifty-million-row warehouse belong in the model's context window for a specific decision. Vector databases might be part of the answer. The underlying decision is a business rules question that only your team can answer.
Then an evaluation harness. A regression suite of real cases with graded correct answers, run against every model version and every prompt change. The harness is what lets you swap the underlying model with confidence.
Then governance. Logs, redaction, refusal, escalation. Every regulated industry needs this and every unregulated one will soon. Governance is the layer where you say what the agent may and may not do, and prove it later.
Then orchestration. The plumbing that decides which model handles which subtask. Some tasks want Claude Opus. Some want Haiku. Some want a Gemini call for a specific tool. The orchestration is where cost meets capability meets latency, and it is the layer where you keep the underlying model market competitive against you.
Every layer above sits outside the shrink-wrapped product category. The composition is architecture. Every part appreciates in value as the underlying models improve, because the composition routes the improvement into your business.
Half-Lives Of The Two Bets
An enterprise that buys the wrapper is making a bet on the vendor. The bet pays if the vendor's product improves faster than the base models improve, which happens roughly never, because the vendor is priced against effort and the base model is priced against inference. The vendor has a structural incentive to keep you dependent on their wrapper. The base model provider has a structural incentive to make every wrapper unnecessary.
An enterprise that builds the composition is making a bet on their own architecture. The bet pays if the composition is deep enough that the underlying model becomes a commodity input inside eighteen months. Every quarter this year has confirmed the direction. Anthropic quintupling in a year and turning its first profit is the substrate bet paying off, and the customers doing the winning are composing on top of a rented brain.
The half-life of the wrapper bet is one model generation. The half-life of the composition bet is however long your business exists.
What Agor AI Advisory Does In This World
We architect the composition layer. That is the entire engagement.
That work means sitting with your team and designing the data adapters. Building the evaluation harness that grades output against your senior operators. Wiring the governance layer against your regulatory environment. Standing up the orchestration that keeps the model layer competitive against your budget. And then handing you an asset that appreciates every time Anthropic, OpenAI, Google or xAI ships a better base model, because your composition is what routes the improvement into your business.
The when-to-build-vs-buy-AI question is priced against a market that closed on May 7. The composition layer is where enterprise durability now lives, and no vendor sells it off a price sheet. Every quarter spent running vendor pilots without an evaluation harness, a governance layer, and an orchestration plane is a quarter of paying a wrapper tax that trains a competitor.
Architecture is the moat. Architects are the necessary partners.
Sources
- OpenAI is shutting down self-serve fine-tuning, Tessl blog, 2026
- Salesforce and Anthropic Announce Claudeforce, Salesforce Press Release, August 26, 2026
- Salesforce Launches 7 Named AI Agents Before Dreamforce, Enterprise DNA, September 2026
- Salesforce's Enterprise Harness Just Became a Product, SalesforceDevops, September 10, 2026
- Anthropic says it hit a $30 billion revenue run rate, VentureBeat, 2026
- Anthropic: $65B Run Rate, First Profit 2026, Value Add VC
- Build vs Buy for Enterprise AI in 2026, Just Think AI