Neither Coat Fits

On August 2, 2026, OpenAI announced GPT-5. A one-trillion-parameter model with unified multimodal input. Three days later the news cycle moved on. What did not move on was the question sitting inside every operator's inbox: are we buying, or are we building?

The question is a trap. The trap was set two years ago, and it closed this month.

The report that killed the debate

In late July 2025, MIT's Project NANDA published State of AI in Business. The number that traveled was ninety-five percent. Ninety-five percent of enterprise generative AI pilots produced no measurable P&L impact. Total spend to that point was between thirty and forty billion dollars. That report keeps getting cited through August 2026 because nobody has published a bigger sample since.

The number that did not travel was more useful. Buy strategies succeeded roughly sixty-seven percent of the time. Internal builds succeeded roughly one-third as often. Read one way, that finding says buy. Read another way, it says something more specific: buy at the layer where the market has priced the model, not at the layer where a legacy vendor has bolted an AI feature onto software you already own.

Six months later, Retool published its 2026 Build vs. Buy report on February 17. Thirty-five percent of surveyed enterprises had already replaced at least one SaaS tool with a custom internal build. Seventy-eight percent said they planned to build more in 2026. Read one way, that finding says build. Read against the MIT number, it says the same thing MIT said: the middle layer is failing.

The middle layer is the packaged AI feature. The AI copilot bolted onto the CRM you renewed in 2023. The workflow assistant your ERP vendor rolled out at their user conference. The natural-language query bar your BI vendor added six weeks after their competitor did. That is the tier the numbers are killing.

What "custom AI" actually means in August 2026

The phrase custom AI versus off-the-shelf tools reads like a binary. It is not. It is a question about which floor you buy at in a very tall building.

The bottom floor is compute. GPUs, memory, network. Nobody sane buys here anymore. AWS, Google Cloud, Azure, Oracle Cloud, and Cloudflare rent it to you by the second.

The next floor up is the frontier model. Claude Opus 5 shipped on July 24. GPT-5.6 reached general availability on July 9. Gemini 3.6 Flash landed on July 21. This floor moves every quarter. The market has priced it aggressively. Sol-tier GPT-5.6 runs four dollars in, twenty out per million tokens. Opus 5 runs five in, twenty-five out. Gemini 3.6 Flash sits below both. You cannot beat these prices by training your own model. You cannot even come close. The frontier is now the off-the-shelf product. Read that sentence twice.

The floor above the frontier is where the fight lives. This is where you assemble a system that does one specific thing for one specific business. Retrieval, tools, memory, governance, evaluation, a workflow that a real person actually uses. This floor cannot be bought. Every SaaS vendor selling you an AI feature is trying to sell you their version of your floor. Every consulting firm selling you a two-year data platform is telling you to build your own version of their floor. Both are wrong for a specific reason.

Buying the vendor's floor means you get their model choice, their prompt, their evaluation harness, their memory schema, and their willingness to change any of it. That means you do not get to swap to Opus 5 when it beats their pinned Sonnet 4. That means you do not get to test a cheaper Flash model on the eighty percent of your traffic that does not need reasoning. That means when the vendor's roadmap drifts, your workflow drifts with it.

Building your own floor from scratch means paying a data science team for eighteen months to fine-tune a model that will be less capable than the next frontier release. That is the story the MIT report told about internal builds. Not that they were technically wrong. That they were architecturally wrong. They put custom effort at the model layer, which is the layer the market is repricing every ninety days.

Fine-tuning left the aisle

Something quieter happened this year that closes the trap. Fine-tuning, the classic answer to "how do we make it ours," went off the retail shelf.

In May, OpenAI posted a deprecation notice for its self-serve fine-tuning platform. New organizations could no longer create fine-tuning jobs. In parallel, OpenAI opened an enterprise fine-tuning tier for large customers with dedicated security posture and negotiated data controls. Anthropic went further and did not open general fine-tuning access at all. Their public position is that unrestricted fine-tuning can degrade the constitutional properties the model ships with. Customization runs through prompts, tool interfaces, policy layers, and operator configuration.

Read this together with the frontier price curve, and the message is unambiguous. The labs will keep the model. They will keep the training data pipeline. They will keep the alignment harness. What they will sell you is inference, a very large context window (Opus 5 is one million tokens), and a set of primitives you compose above.

The strategy is no longer whether to fine-tune. The strategy is what to build in the layer above the model, given that you cannot get inside the model.

The shadow AI signal is data, not embarrassment

The MIT report identified a second number worth taking seriously. Ninety percent of employees regularly use personal AI tools for work. Only about forty percent of their employers have an official AI subscription. When employees run the same tasks through their personal Claude or ChatGPT accounts, those consumer tools reach production output about forty percent of the time. The custom enterprise systems the same companies procured succeed about five percent of the time.

Executives read that gap as a security problem. It is, but that is not what it tells you strategically. It tells you that people are using the frontier directly because the middle layer their employer bought does not work. The middle layer does not work because it is pinned to old models, tuned for a generic use case, wrapped in an approval flow, and rate-limited by a procurement contract that was signed before Opus 5 existed.

The correct response is to build a governed version of what your employees are already doing on their phones during their commute. Route their queries to the current frontier model. Attach the retrieval store they need. Attach the tools they need. Log what they do so you have a corpus for evaluation. That workflow is a two-week engineering project, not a two-year platform.

The dying middle: three concrete failure modes

Take three real product categories the numbers indict.

Salesforce Einstein. The generative AI features bolted onto a CRM that was priced and shaped for the pre-agent era. Every workflow assumes a human clicks a button, reads a suggestion, and edits. The frontier now runs agentic sequences of dozens of steps in a single session. Einstein's shape does not match what the model is capable of. You paid for both the seat and the shape.

Microsoft Copilot for Enterprise. General office suite integration where the surface area is spreadsheets, decks, and email. The average adoption rate reported through 2026 sits below fifteen percent per seat per month across most large deployments. The seat cost is thirty dollars. Consumer ChatGPT Plus is twenty. Your employee is already paying the twenty out of their own pocket. That is what the shadow AI number is telling you when you decode it.

ERP vendor AI copilots. SAP Joule, Oracle Fusion AI Agents, Workday Illuminate. Sold as inside-the-suite productivity. In practice they cannot see data outside their suite, which is where the actual work sits. The frontier model your competitor ran directly on the data lake, with retrieval and tools their team wrote in a week, moved faster and covered more surface.

None of these vendors is stupid. They are structurally constrained. They must protect their contract shape, their data model, their release cadence, and their support commitments. Those constraints are exactly what prevents the middle layer from tracking a frontier that resets every quarter.

What the winners actually built

Look at the five percent of pilots MIT flagged as extracting real value. The pattern is consistent.

They chose a single workflow with a measurable output. Insurance underwriting notes. Legal contract redlines. Claims triage. Customer support first-response drafts. They picked one because you cannot evaluate three at once.

They wrote a small system on top of the frontier. Not a platform. Not a data lake redesign. A set of prompts, a retrieval store, a tool interface, a logging layer, and a review loop. Total original code counted in the low thousands of lines. Total original glue counted in the tens.

They bought the model at the tier that matched the task. Flash for volume, Sonnet for the middle, Opus for the hard one. They routed by task, not by vendor loyalty. Every request carried a tag that said which tier answered it and how much it cost. Every dashboard showed that number to the person who owned the budget.

They kept the model swappable. When Opus 5 shipped on July 24, they were running it in staging by July 25 and in production the week after. When GPT-5.6 Sol shipped on July 9, they benchmarked it against their eval harness the same day. The model was a dependency, not a marriage.

They wrote the eval harness first. Before the workflow shipped, they had a set of paired inputs and expected outputs the model had to handle. That harness is the actual moat. It captures what "good" means at your company for this task. It is not fine-tuning. It is judgment, encoded.

Why the vendor cannot do this for you

The eval harness is the answer to why you cannot buy this. A vendor's evaluation harness reflects what their average customer wants the model to do. Your evaluation harness reflects what your customer wants and what your regulator will accept and what your senior operator would sign. That specificity does not survive translation into a generic product. It never will.

The workflow on top of the model is the same. It has to know your data model, your permission scheme, your escalation ladder, your naming conventions, your peculiar rules about what a valid quote looks like. The vendor cannot know that. The vendor's incentive is to sell you a version of that logic that also works for their next hundred customers, which is by definition a version that does not do the specific thing you need well.

The frontier model is the opposite. The frontier model works well precisely because it was trained on the union of the whole public internet and has no idiosyncratic view of your business. That neutrality is the feature. You pay for capability and get to keep the shape of your operation intact around it.

The build that actually stays small

The build most executives fear is the wrong build. The build most vendors sell them into is the wrong build. What actually works is smaller than either.

It is a repo. It is a service that receives a request, calls the frontier model with the right prompt and the right retrieval context, invokes a small set of tools, logs the trace, and returns an answer. It is an evaluation harness that runs every night on a held-out set. It is a review dashboard where a human corrects the model's errors and those corrections flow back into the eval set. It is a switch that lets you point at a different model tomorrow.

Two engineers can hold this system in their heads. That is the test.

If you cannot describe your AI system in a paragraph, it is too big. If you cannot swap the model in an afternoon, you have accidentally built a fine-tune of a specific vendor's assumptions and you will pay for that decision in 2027 when the assumptions age out.

The architectural principle

Rent the model. Own the wiring.

The rentable part of the stack is the piece the market has already commoditized to a level no single company can beat. The frontier moves too fast for capex to keep up. Every hyperscaler has already lost this fight to the labs and now pays them to host their inference. The labs will keep repricing every quarter. Take the discount.

The ownable part of the stack is the piece that reflects your business. Your data. Your rules. Your definition of correct. Your review process. Your eval harness. Nobody sells that to you because nobody else knows what it is.

Buy versus build is the wrong question because it treats the stack as one thing. Split the stack. Buy the parts that reprice every quarter. Build the parts that reflect who you are and what "done" means in your building.

The two-year contract as an anti-architecture

One more piece of the trap deserves naming. The middle-layer vendor sells on a two-year or three-year commit. That commit is the point. It locks you to their model choice for a period during which the frontier will reprice at least eight times. Every quarter you carry the deal, you are further from the price and the capability the market is offering. The commit becomes a debt. The vendor knows this. That is why the pricing looks attractive at year one.

The architecture that survives contracts differently. It signs short. It buys inference on usage. It keeps the switch at the router, not at the vendor. If Opus 5 gets cheaper, the workflow gets cheaper. If Gemini 3.7 beats the Sol tier on your eval harness, the workflow gets faster. The contract shape mirrors the model shape. Both are quarterly.

Why this needs an architect, not a procurement decision

The failure mode the MIT report captured is not a technology failure. It is an architecture failure. Enterprises put their custom effort at the wrong floor. They tried to build the model. They should have been building the wiring. They bought the middle layer. They should have been buying the frontier.

The shift is a design decision that cuts across procurement, engineering, security, compliance, and operations. No single team inside a normal company owns it. Procurement writes the vendor contract. Engineering picks the framework. Security scopes the data access. Compliance rules on the audit trail. Operations decides who sees the output. Without an architect stitching those decisions into one shape, the default is to buy the middle layer because someone in one of those five rooms wants a vendor to blame when it fails.

Agor AI Advisory does that stitching. We come in, look at the stack you already have, name the floor you should be renting and the floor you should be building, and produce the concrete architectural spec that names the model, names the tools, names the eval harness, and names the review loop. We write the paragraph two engineers can hold in their heads. We shape the contract so you keep the swap right. We build the eval harness first, before anyone writes a prompt, because that is where the moat actually sits.

The frontier will keep repricing every ninety days. The next Opus, the next GPT, the next Gemini will each be a step change. If your architecture cannot ride the curve, you are paying full fare for the middle layer while your competitor pays wholesale for the frontier. That gap compounds. Every quarter you carry the wrong shape costs you the delta between what the model can now do and what your bolted-on feature will let it do.

Neither coat off the rack will fit. The winners are already at the sewing machine.