Customer Support AI

It varies enormously with what you are building. Integrating an existing model API into a product sits at the lower end; a custom-trained model with a data pipeline, evaluation harness and governance documentation sits considerably higher. The bigger surprise is usually running cost rather than build cost, so we model cost per inference at expected volume during architecture.

A prototype with evaluation typically runs six to ten weeks. A production system with integration, guardrails and monitoring usually runs four to eight months. We deliberately front-load a prototype and evaluation phase so you get a real accuracy number before committing to the full build.

That is the first question we answer, and sometimes the answer is no. Data audit covers volume, quality, coverage of edge cases, labelling and whether consent covers the intended use. If the data is not there, we will say so rather than building something that will underperform quietly.

API models cover most use cases and are usually where to start. Fine-tuning helps when you need a consistent format or domain tone that prompting cannot reach. Training from scratch is rarely justified outside specialised domains with proprietary data. We recommend based on your case, and the cheapest option that works is usually the right one.

Retrieval grounding so answers come from approved sources, citations so users can verify, confidence thresholds that trigger a refusal or escalation instead of a guess, and an evaluation set run on every change. Hallucination cannot be eliminated, and any vendor claiming otherwise is overselling.

Yes. Options include enterprise API tiers with no-training guarantees, deployment inside your own cloud tenancy, or open-weight models self-hosted entirely within your infrastructure. Which is right depends on your data residency requirements and cost at volume.