AI work that helps a business is usually a narrow feature: answer a class of support questions from your own docs, pull fields out of a PDF, draft a first reply, or move a lead from one tool to another when a form is submitted. It is not a custom-trained model, and it is not a chatbot glued to the homepage with no fallback.
Who this is for
- You want a support assistant that answers from your content and hands off when it should not guess.
- You want a model inside a product you already have: summarise, classify, extract.
- You have a repetitive process across email, sheets, and a CRM that should run without a person copying fields.
- You tried a no-code bot and it cannot see your real data or your real permissions.
If you cannot describe the job in a paragraph, we start there. We do not sell an “AI transformation.”
What we build
- A defined feature with inputs, outputs, and a failure path. If the provider times out, the user sees a useful state, not a spinner.
- Prompts, temperature, and tool limits stored where you can edit them.
- Retrieval from your documents only when the answers must come from your material. We say what is indexed and what is not.
- Workflow automation in n8n or in your app’s job queue, with a log so you can see why a run did or did not fire.
Your API key, your provider account. Usage bills are not part of our invoice.
Process and timeline
One job per phase. We decide whether a model is even required. A webhook plus a template email is sometimes the whole solution, and we will quote that instead of an LLM feature.
A single extraction or a single automation is often a short fixed-price engagement once the sample inputs exist. A support bot with your documents, roles, and a human handoff is larger because the handoff is the hard part. Figures you want shown publicly go on pricing after you confirm them.
What changes the price
How clean the source documents are, whether answers must respect per-user permissions, how many systems the workflow touches, and whether we are embedding this in an existing app or standing up a separate tool.
Related services
APIs and backends when the model sits behind your API. Custom web applications when users need a screen. Maintenance when prompts and workflows need a monthly owner.