STRATEGY · Jul 3, 2026 · 7 min read
You probably don't need a custom model
Small businesses are routinely quoted six figures for a bespoke model when the underlying problem is retrieval, routing, or a missing integration. The decision is easier than vendors make it sound.
Level 1 — Buy the workflow
If a mature product already does the job — scheduling, invoice capture, support triage, call transcription — buy it. You are paying for someone else's maintenance, security review, and roadmap. The only real questions are whether it exports your data cleanly and whether it integrates with what you already run.
Level 2 — Ground a general model in your own content
Most 'the AI doesn't know our business' problems are retrieval problems, not model problems. Putting your policies, product data, and past tickets into a searchable index and giving a general-purpose model access to it solves the majority of internal-knowledge use cases at a fraction of the cost.
This approach has a second advantage: when a policy changes, you update a document rather than retraining anything.
Level 3 — Build a workflow around the model
The value in most engagements is not the model at all. It is the plumbing around it: pulling the record, validating the input, applying your business rules, writing the result back to the system of record, and escalating what it cannot handle.
That plumbing is ordinary software. It is testable, auditable, and it keeps working when you change model providers.
Level 4 — When custom is actually justified
Training or fine-tuning your own model makes sense in a narrow set of conditions, and you should be able to tick most of them:
- You have a large volume of proprietary, well-labelled examples that no general model has seen.
- The task is narrow, repetitive, and high-volume enough that unit cost or latency matters.
- Off-the-shelf models have been measured on your data and demonstrably fall short — not assumed to.
- You have somewhere to put the ongoing evaluation and retraining work after launch.
Decide with a measurement, not a demo
Before committing to a build, assemble fifty real examples from your own operation with known correct answers, and run the cheapest option against them. That test costs a day and settles the argument that the vendor deck cannot.
Got a Priority 1 problem?
Let's find out what it is — and what it's costing you.