From a convincing prototype to a disappointing service
Generative AI in the enterprise almost always begins with a successful demonstration. A prototype answers questions about a set of documents, and the audience is convinced. The difficulties come later, when real users ask real questions.
The prototype answered well because the questions were chosen by those who had built it. In real use, requests are ambiguous, incomplete, sometimes out of scope. The system answers anyway, with the same confidence, including when it is wrong. A user who spots two serious errors stops trusting the tool, then stops using it.
Then there are the questions the demonstration had left aside. Does the assistant respect access rights to documents? What becomes of an answer when the source document changes? What does a request cost, and what will a whole department’s daily use cost? Who is alerted if quality deteriorates after a change of model version? These are engineering questions, not model questions.
LLM, RAG, fine-tuning: the options and what they involve
A large language model, or LLM, produces text from the patterns learnt during its training. It does not know your internal documents, it is unaware of facts after that training, and it can produce a plausible but false statement. The architecture of a business assistant consists of compensating for these three limitations.
| Approach | Principle | Suitable for | Limitation |
|---|---|---|---|
| Prompting alone | The model receives an instruction and, if needed, the text to process | Rewording, summarising, extraction from a supplied document | No knowledge specific to the organisation |
| Retrieval-augmented generation (RAG) | The system retrieves the relevant passages from your document base and passes them to the model | Questions about an internal corpus, sourced answers | The answer is no better than the passages retrieved |
| Fine-tuning the model | The model is retrained on your examples | Highly specialised format, vocabulary or task | Maintenance burden, work to redo when the base model evolves |
RAG is often the starting point for a document assistant, because it makes it possible to cite sources and update knowledge without touching the model. Its quality is decided before generation: document chunking, metadata, version management, combining keyword search with semantic search, reranking of results.
The choice of model comes afterwards. It is made on your tasks, not on a general ranking: quality measured on your own questions, place of data processing, cost per request, response time, the possibility of changing provider. A model operated by a provider and an open model you host yourselves entail neither the same confidentiality constraints nor the same operating burden.
Our approach to generative AI in the enterprise: starting from evaluation
Our consultants work within your teams; decisions remain yours.
Defining the task and its limits. An assistant ‘that answers everything’ cannot be evaluated. We start from a precise task, an audience and a bounded corpus. We also write down what the assistant must refuse: an out-of-scope question, a request for which no source exists.
Building the reference set before the system. With the domain experts, we put together a set of real questions and expected answers, trick cases included. This set serves as the acceptance criterion. It separates two measurements: does the system retrieve the right passage, and is the answer faithful to that passage?
Preparing the corpus. Duplicates, outdated versions, contradictory documents, poorly extracted tables and diagrams: document preparation often weighs more on the result than the choice of model. It also raises the question of who owns each source and how often it is updated.
Treating security as a functional requirement. Document access rights must be carried through into the index: a user must not obtain through the assistant what they are not entitled to read. Indexed content may also contain malicious instructions (prompt injection), which makes it necessary to filter inputs and to limit what the system can do.
Industrialising. Logging of exchanges, tracking of cost and quality, regression tests rerun at every change of prompt, index or model, collection of user feedback. These practices extend those of MLOps, adapted to language models.
When the need goes beyond answering a question and involves chaining actions in your applications, the subject becomes that of AI agents, with additional control requirements.
Common mistakes in generative AI projects
- Choosing the model before defining the task. Comparing models only makes sense on your questions. Without a reference set, it boils down to an impression.
- Indexing the entire document estate at once. A broad, unsorted corpus lowers the precision of answers. A narrow, clean scope gives better results, and is then widened.
- Evaluating by demonstration. A few questions asked in a meeting say nothing about behaviour across all real requests. Only a reference set rerun systematically makes it possible to compare two versions.
- Not providing for refusal. An assistant that cannot say ‘I cannot find this information’ invents. A reasoned refusal is a feature.
- Ignoring cost at scale. The length of the context passed to the model and the number of calls per request determine the bill. They can be measured from the pilot onwards.
- Delivering without an operator. An assistant whose corpus, evaluations and changes no one maintains deteriorates without anyone noticing.
How to measure the success of a business assistant
We track four families of indicators, set at framing and recorded both before and after going into service.
- Quality. Share of answers judged correct and faithful to the sources on the reference set, tracked version after version. Justified refusals and wrongful refusals.
- Usage. Share of target users who come back once the discovery phase is over, frequency of rephrasing, negative feedback reported.
- Operations. Cost per request, response time, confidentiality incidents, time taken to reflect an updated document.
- Business effect. The indicator of the process concerned, chosen with its owner: time to process a case, rework rate, workload of second-level support.
A heavily used assistant whose business indicator does not move raises a question of process design, not of technology. It is at this stage that work with teams on their practices takes over from engineering.