A regional insurer's claims lead opens her laptop and unpins the ChatGPT tab she's kept open for two years. The tool still works fine. The problem is that her team kept hitting the same wall: the model couldn't see the policy language, adjuster notes, and state-by-state endorsements that make her job her job. Every answer had to be hand-fed context, then double-checked.
The subscription was being paid. The work wasn't getting done.
That claims lead is a composite of a pattern showing up across industries right now. The generic AI subscription — one seat, one chat window, one model trained on the open internet — is losing ground to something a lot more specific: a private model that reads the company's own documents, follows the company's own workflows, and lives inside the company's own security perimeter. This week DEV.co launched cstm.ai, a custom AI development practice built around private, public, and enterprise-grade LLM deployments wired into a client's own data and workflows.
The Tab Hits a Wall the Vendor Can't See
The first six months with a general-purpose assistant look great. People draft faster, summarize meetings, clean up emails. Then the questions get harder.
What did we agree to in the master services agreement with that vendor in 2023? Which of our 14 underwriting rules apply to this submission? What's the exact wording our compliance team approved for this disclosure?
The model has no way to know. Nothing in its training set includes the company's proprietary contracts, internal rulebooks, or the three Slack channels where the real decisions get made.
So the human on the other end copies and pastes, redacts, re-prompts, and eventually gives up and does the task the old way. The subscription becomes a writing aid instead of a decision tool. For most of the money, that's a bad trade.
The Replacement Is a Model That Reads Your Documents
The architecture most teams land on isn't a fine-tuned foundation model built from scratch — that's expensive and rarely necessary. It's retrieval-augmented generation: a pipeline that indexes the company's own documents, pulls the relevant passages at query time, and feeds them to the model as grounded context. IBM's explainer frames RAG as a way to connect a model to authoritative sources it wasn't trained on, including proprietary enterprise data, so answers reflect what the business actually knows rather than what the public web happened to contain.
That shift changes what the claims lead sees on screen. The question is the same. The answer now cites the specific policy section, the specific endorsement, the specific adjuster note. If a passage doesn't exist, the system says so rather than inventing one.
Hallucination doesn't vanish, but it gets a much shorter leash. This is the capability a custom build delivers that a generic subscription cannot.
Custom Software Engineering Is Where This Lives or Dies
A RAG pipeline is easy to demo and hard to run. The messy work is everything around the model: ingesting PDFs that were scanned sideways in 2011, chunking contracts so clauses don't get split in half, keeping permissions intact so a junior analyst can't retrieve the CEO's compensation memo, logging every query for audit, and wiring the whole thing into the systems people already use. That's custom software work, not a plugin.
That missing integration layer is what the market is responding to. The firms that already do custom software engineering are the ones best positioned to make AI fit a specific business, because most of the hard problems are integration problems rather than model problems.
The Economics Change When the Model Starts Doing Real Work
Per-seat subscriptions make sense when the tool is a writing aid used occasionally by individuals. They stop making sense when the model is handling a step in an actual business process — triaging claims, drafting first-pass responses, pulling the right contract clause into a renewal. At that point the question isn't how many seats to buy. It's what the system is worth per decision it supports, and whether the company owns the thing doing the work.
Ownership is the quiet reason the generic subscription is losing ground. A custom build gives the business the model configuration, the retrieval pipeline, the prompts, the evaluation harness, and the code around all of it. The subscription gives the business a login.
Back to the claims lead: she doesn't need a better chat window. She needs a system that knows what her company already knows, keeps it inside the walls, and gives her a straight answer with a citation. The ChatGPT tab was unlikely to ever be that. Something built for her finally can be.


