Custom AI agents, RAG systems and private LLMs

The price moves with six things: how many systems the agent touches, whether it acts or only drafts, the state of the data it reads, how many cases the evaluation set needs, the security and approval steps, and where the model runs. A drafting agent with one integration is a small project. An agent that updates records across several systems behind approval gates is a larger one. Running costs are separate: model usage, hosting and monitoring. Every proposal states the build price before work begins.

Often a workflow is enough. If the steps are the same every time, a fixed workflow with a model at one or two points is cheaper to run and easier to test. We build an agent when the path changes from case to case and the model has to choose the next tool. We will tell you which one the job needs before we quote.

Sometimes, and then you should use one. No-code tools are quick for linking apps and suit simple, low-risk flows. A custom build pays off when the agent needs your permission model, an evaluation set, audit logs, approval steps or private deployment, or when it becomes part of a product you sell.

It can only call the tools we give it, and each tool has the narrowest access that does the job. Anything expensive or irreversible, such as sending an email, issuing a refund or changing a record, waits for a person to approve it. Every call is logged and traced. Content from documents and the web is treated as untrusted, and prompt-injection attempts are part of the red-team checks before launch.

Usually not because of the model. Nobody wrote down what a correct result is, the pilot ran on clean sample data, integrations and permissions were left for later, or the running cost was never modelled. We start with an evaluation set from real cases, connect to the real systems early and set a cost per request in the design.

Train: fine-tune or build the retrieval index on your data, in your account. Test: run a held-out evaluation set from real cases, with a baseline and a regression run on every change. Validate: guardrails, red-team checks, cost and latency budgets and a human sign-off. Deploy: serve it on your cloud or hardware with tracing, monitoring, cost limits and a rollback path.

RAG when the model needs facts that change or must be cited. Fine-tuning when it needs a consistent style, format or behaviour that prompting cannot hold at volume. Many products use both. We measure the gap on your test set before recommending either.

We score retrieval and answers separately. For retrieval, we check that the right passages came back for each question in the evaluation set. For the answer, we check that it is correct, that every claim is supported by a cited source, and that it says so when the sources do not contain the answer. Automatic scores run on every change, and people review a sample where judgement matters.

Yes. Permissions are applied at retrieval time, using the same roles as the source system, so a user is only shown passages they could already open.

Not by default. OpenAI, Anthropic and Google Cloud each state that business and API data is not used to train their models unless you opt in or give permission. We use only the providers you approve in the proposal, under their API data terms, and serve an open-weight model in your network when no outside provider is acceptable.

Wherever your data policy allows: hosted models on Google Cloud or OCI for most products, or open-weight models served inside your network with vLLM when data cannot leave it. Both get the same evaluation and monitoring.

Usually not. Hosted models cover most products. We set up GPU compute when you need private inference, high volume or a custom model.

It depends on the job. We use LangGraph for long-running agents that need checkpoints and human approval, a provider's own SDK when a product is built around one model family, plain code when a workflow is enough, and the Model Context Protocol to give agents audited access to your systems. The evaluation set, not the framework, decides what ships.

Agentic AI describes systems in which a language model plans the steps of a task and calls tools to carry them out, rather than returning a single answer. It ranges from one agent with three tools to several agents passing work between them. The more autonomy a system has, the more testing and approval it needs.