Domain-specific model or LLM (ChatGPT)? Cost, accuracy, speed and data security
A private, dedicated domain-specific model wins in repetitive company tasks such as classification, data extraction or scoring. According to Coblit’s internal research, in the tasks it is trained for, it reaches 95–100% accuracy and responds in about 100 ms, has a fixed cost and can run fully on-premise. A general LLM (e.g. ChatGPT) is better for free-form conversations and texts on any topic — but it charges per token, responds in 2–3 s and gets narrow tasks right in 60–70% of cases.
The short answer
If AI is to perform one repetitive task in your company — assess a tender, read a document, classify a request, assign a score — a private, dedicated domain-specific model trained on company data is the better choice. If you need an assistant for free-form texts and conversations on any topic, a general LLM such as ChatGPT will do.
Comparison: Coblit’s private, dedicated domain-specific model vs a general LLM
The figures below come from Coblit’s internal research and apply to the tasks the domain model is built and trained for. They are not universal values for every AI use case.
| Coblit’s domain-specific model | General LLM | |
|---|---|---|
| Cost | Low | High |
| Cost type | Fixed | Variable — usage-based (tokens) |
| Accuracy | High: 95–100% | Low: 60–70% |
| Response time | Low: ~100 ms | High: 2–3 s |
| Margin on an AI-based process / product | High | Low |
| Infrastructure | Fully on-premise | Data transferred to the cloud |
| Data | No transfer to the cloud | Transferred to the cloud |
| Errors and hallucinations | Doesn’t hallucinate in its domain | Errors and hallucinations |
Why doesn’t a domain-specific model hallucinate in its domain?
A general LLM generates free-form text, so when it doesn’t know the answer, it can “fill in” one. A domain model works differently: it answers only within the task it was built for — classification, data extraction or scoring on the client’s data — and doesn’t generate free-form text outside it. So there is no room for it to make up an answer beyond its domain.
Fixed cost vs per-token cost — what it means for margin
A general LLM charges for every query in tokens. The more documents, customers or decisions AI handles, the bigger the bill — and the smaller the margin on the process or product built on that AI. A domain model has a fixed cost, so growing volume doesn’t eat into the margin.
When a general LLM is enough
- you write texts, emails and summaries and need ideas,
- tasks are varied and hard to fit into a single category,
- the scale is small, so token fees don’t matter,
- the data can leave the company.
When a domain-specific model wins
- the same task is repeated hundreds or thousands of times (classification, extraction, scoring),
- accuracy and no made-up answers matter,
- speed matters — an answer in about 100 ms instead of 2–3 s,
- AI is part of a product or process where margin matters,
- the data cannot go to the cloud — then the model runs fully on-premise.
How Coblit does it
- Analysis of the data and the task — we agree exactly what the model is to do and how we will measure accuracy.
- A private, dedicated domain-specific model — we build it on the client’s data for that one task.
- Deployment — on the client’s own servers and resources (on-premise) or, if the client prefers, in the cloud. We have specialists for this.
- Maintenance — monitoring, retraining on new data and extending the system with further tasks.
We quote the price and deployment time individually. The easiest way to start is an AI pilot on a single process.
Example: a dedicated tender model
For an international pharmaceutical company, we built an AI tender analysis system with a model trained exclusively on the client’s data. The model assesses how well tenders fit the company’s offer, and work that required more than 10 people is now done by 3.
More on how our models work: Private AI models and Private AI model or ChatGPT?.