Skip to content
Develozy

AI integration

Assistants, search and automation added to an existing product — scoped, grounded and honest about limits.

Overview

We have shipped a branded assistant on a live client site: suggested prompts, a handoff to a human on WhatsApp, and a visible note that answers may be approximate. That last part matters. An assistant that confidently invents your pricing is worse than no assistant, so we ground responses in your actual content and design the escalation path before the chat window.

What you get

  • Assistant grounded in your real content, not general knowledge
  • Retrieval over your pages, docs or catalogue
  • Human handoff — WhatsApp, email or your inbox
  • Visible AI disclosure and approximation notice
  • Guardrails on topics, tone and what it must never claim
  • Usage and cost monitoring so spend cannot run away

What this isn't

  • We will not ship an assistant that can invent prices or availability
  • No AI-generated content published without a human in the loop

The problem

The problem AI integration solves

The failure mode is not that the assistant cannot answer. It is that it answers anyway. A model with no grounding will quote a price you do not charge, promise a date you cannot meet, and do it in fluent, confident sentences your customer has no reason to doubt. You then honour it or argue with them.

How it runs

Four phases, and what you get from each.

3–6 weeks

  1. Decide what it may say

    Before any code: the questions it should handle, the questions it must refuse, and where it hands to a human. The refusal list is the more important half.

    Scope document — answer, refuse, escalate

    3–5 days

  2. Ground it

    Retrieval over your real content — pages, catalogue, docs, pricing — so responses cite what you actually publish. Nothing is answered from general model knowledge.

    Indexed content source with a refresh path

    1–2 weeks

  3. Build the interface

    Suggested prompts so people know where to start, streaming responses, a visible AI disclosure, and a handoff to WhatsApp, email or your inbox that carries the conversation with it.

    Assistant live on staging with handoff working

    1–2 weeks

  4. Guardrail and watch

    Adversarial testing against the refusal list, then usage and cost monitoring with a hard ceiling so a bad week cannot produce a surprising invoice.

    Test results, spend caps, monitoring dashboard

    3–5 days

Levels

Three builds, not one price.

  • Assistant

    Answering the questions your inbox receives twenty times a week.

    • Grounded in your existing site content
    • Suggested prompts and streaming replies
    • Visible AI disclosure
    • Email or WhatsApp handoff

  • Grounded

    A catalogue, pricing or availability that must never be invented.

    • Everything in Assistant
    • Retrieval over structured data, not just pages
    • Refusal rules on price and availability claims
    • Conversation history and handoff context
    • Usage and cost monitoring with caps

  • Automation

    AI doing work inside your product, not only talking about it.

    • Everything in Grounded
    • Semantic search across your content
    • Classification, routing and summarisation in workflows
    • Human-in-the-loop review before anything publishes
    • Evaluation suite run against every prompt change

Where we've done this

Shipped, not theoretical.

Questions

Asked before, answered here.

Also relevant

Rarely bought alone.

Have a project in mind? Let's get to work.

Our team is here to deliver tailored solutions that drive results. Tell us what you are building and we will come back with a route to it.

Let's talk