“We want our own AI” is one of the most common openings in this trade, and one of the most expensively misunderstood. It names four different projects with a thousandfold cost range between them. This page is the ladder, climbed honestly: what each rung fixes, what it costs, and where firms should actually stop.
The ladder at a glance:
| Rung | What it fixes | Typical cost | Most firms should |
|---|---|---|---|
| 1. Careful prompting | Tone, format, rules | A day of thought | Always start here |
| 2. RAG | “It doesn’t know our stuff” | A real build, modest running | Stop here, happily |
| 3. Fine-tuning | Consistent style/format at volume | Cheap runs, costly datasets | Use for specific niches |
| 4. From scratch | Almost nothing a business has | Lab budgets | Never |
Rung one: careful prompting
A written brief, good examples, your terminology and red lines, maintained like the document it is. This is “your own AI” in the sense that matters most often: the model behaves your way because it is told to, verifiably, on every call. Cost: a day of thought. It is where every project below starts anyway, and a surprising share end here happily. The internal assistant playbook shows the pattern.
Rung two: RAG, your knowledge attached
Retrieval-augmented generation: the system finds relevant passages from your own documents and hands them to the model at answer time, so answers come from your knowledge base, current and citable, with refusal when nothing matches. This is the correct rung for the most-wanted outcome of all: a bot that knows your policies, your products and your history without hallucinating them. It is how grounded customer-facing chatbots are built, and updating it is filing a document, not retraining anything. Cost: a real build, modest running costs, and the discipline of keeping the source documents true.
Rung three: fine-tuning, behaviour adjusted
Fine-tuning re-trains an existing model on your examples so behaviour you previously prompted for becomes default: house style held exactly, a fixed output format at high volume, a narrow classification done cheaply by a small model. The frontier labs and cloud platforms sell it hosted; the open-weight families take it via efficient adapter methods, giving you a specialised model you can run privately. The honest gates: you need hundreds of clean examples, an evaluation you trust, and a reason RAG and prompting have demonstrably failed, because facts are not what fine-tuning is for. Where those gates pass, it is a genuinely good buy; where they do not, it is an expensive way to avoid writing a better prompt.
Rung four: training from scratch
Building a model from raw data is a lab exercise with lab budgets, and for a business the case almost never closes: an adapted open-weight model beats a small bespoke one on cost, capability and upkeep in essentially every commercial scenario. If a supplier proposes it, ask what rung three failed to deliver and watch the answer carefully.
Choosing your rung
Work down from the symptom:
- Wrong tone or format → rung one, then rung three at volume
- Doesn’t know our stuff, or makes it up → rung two, almost always
- Too expensive per call at scale → a smaller model plus rung three
- Data cannot leave the building → rung two or three on open weights you host
Nobody’s symptom says rung four.
Where I fit in
The ladder is my daily terrain: prompted systems, grounded RAG builds and routed model stacks run in production across my own estate, and telling clients which rung their problem lives on, including when it is a cheaper rung than they came in wanting, is half the value of an Automation Audit. Bring the “we want our own AI” sentence and I will translate it into its rung, its price and its timeline on the call.