Delvops

Custom AI chatbot development

An assistant that answers from your real documents, and says “I don't know” rather than inventing.

Typical timeline : 2 to 3 weeks

Built with

  • OpenAI
  • Anthropic
  • Next.js
  • PostgreSQL

Pricing

Setup

from€2,500

Ingesting your documents, tuning the behaviour, testing on real cases and installing it on your site.

Maintenance

€150/ month

Hosting, keeping the knowledge base current, reviewing conversations and adjusting.

Free · No commitment · Reply within 24 hours

The problem

Your team answers the same questions ten times a day, and your clients wait for an answer you have already written down somewhere.

Who it's for

  • Support teams handling a high volume of repetitive questions
  • Sites with rich documentation that is hard to navigate
  • Internal teams hunting for information scattered across dozens of documents

Scope

What's included, and what isn't

Stating the limits up front spares both sides an unpleasant surprise later.

Included in the setup

  • Ingesting your sources: documents, site, knowledge base, product sheets
  • Answers grounded in your documents, with the source cited
  • Guardrails: what the assistant must never assert or promise
  • Handover to a human when the question falls outside its remit
  • A chat bubble embedded in your site, in your brand's colours
  • A dashboard of questions asked — what your clients are really looking for

Included in the maintenance

  • Hosting and inference costs covered up to 1,000 conversations a month
  • Knowledge base updated whenever your content changes
  • Monthly review of conversations and correction of weak answers
  • Email support, answered within one working day

Not included

  • Writing or reworking your source documentation
  • Beyond 1,000 monthly conversations, billed on usage
  • Health or banking data without a dedicated scoping phase
  • A voice version — see the AI voice agent service

FAQ

Frequently asked questions

Can the assistant make things up?

That is the main risk, and it can be handled. The assistant answers only from your documents and cites its sources; when the information is not there, it says so and offers to pass the question on. I test that behaviour against deliberately awkward cases before going live.

Are my documents used to train a public model?

No. Your content stays in your own space and is not used for model training. I verify that point with whichever provider we choose and document it in writing at delivery.

What happens when the assistant cannot answer?

It offers to pass the request on, collects the contact details and emails you the conversation. You get a qualified lead instead of a visitor who closes the tab.