# Building a custom AI assistant for your company: what it actually takes

> What goes into a private AI assistant or RAG automation built on your own data - the components, the data preparation nobody budgets for, realistic costs, and when custom beats off-the-shelf.

Source: https://olyteck.com/blog/custom-ai-assistant-what-it-takes
Published: 2026-07-20
Author: Oleg Garasym, Olyteck (France, EU-hosted)
Product: Olyteck Studio
Topics: custom ai assistant, rag, ai automation, build vs buy, private ai, internal tools
License: free to quote with attribution to Olyteck and a link to the source URL.

---

At some point every team hits the wall of off-the-shelf AI: the SaaS assistant cannot reach your
ERP, the chatbot cannot follow your approval workflow, the generic tool cannot be allowed near the
data that would make it useful. That is the moment "should we build our own?" stops being a
hypothetical. Here is what building one actually involves, without the vendor gloss.

## What a custom assistant is made of

Every serious build has the same five layers, whatever the use case:

1. **A data pipeline.** Connectors that pull from where your knowledge lives - SharePoint, a file server, the ERP, ticketing - and keep the assistant's index current when documents change. This is plumbing, and it is half the project.
2. **A retrieval layer (the RAG part).** Documents are chunked, indexed and searched so the assistant answers from your content rather than from the model's general training. Quality here decides whether answers cite the right paragraph or the almost-right one.
3. **The model itself - a choice, not a given.** Hosted frontier models, EU-hosted models, or fully local ones, traded off on quality, cost per query and confidentiality. The right answer differs per use case, and a good architecture lets you swap without rebuilding.
4. **Guardrails.** Source citations, abstention when retrieval comes up empty, permission awareness (the assistant must not show a user content they could not open themselves), and logging you can audit. This layer is the difference between a tool and an incident.
5. **The workflow around it.** The assistant that saves real money is rarely a chat box. It is the agent that drafts the questionnaire response into the Excel file, routes it for approval, and files the result - the automation shaped exactly like your process.

## The part nobody budgets: your data

The model is the cheap part. The expensive part is discovering that your knowledge base is three
generations of contradictory policy documents, that the "current" price list exists in four
versions, and that the folder everyone relies on has permissions nobody can explain. Every custom
AI project is partly a data-hygiene project; budget for curation, an owner per source, and a rule
for what happens when documents conflict. Teams that skip this ship an assistant that answers
confidently from the wrong version.

## When custom beats off-the-shelf (and when it does not)

We covered the general decision in our buy-vs-build article; for AI assistants the short version:

- **Buy** when your need matches a well-served pattern - answering questionnaires from past responses, internal Q&A on policies - and the tool meets your data-residency bar. Proven SaaS ships today at per-user prices.
- **Build** when the value depends on YOUR systems and process: data in an ERP or bespoke database, a workflow with approvals and hand-offs, confidentiality that rules out shared platforms, or an assistant that must act (create, update, route), not just answer.
- **The hybrid is often right:** an off-the-shelf RAG core with a thin custom layer for your workflow, which costs a fraction of a ground-up build.

## What it costs, honestly

For an SMB-scale scoped build - one process automated end to end with a RAG layer, guardrails and
handover - realistic 2026 pricing starts around the low five figures. Olyteck Studio prices this as
fixed-scope, fixed-price packages: a Custom Automation Build from €12,000, and AI-Ready Foundations
(making a Microsoft 365 tenant safe for AI rollout first) from €9,600, built by EU engineers, with
code you own at the end - no per-seat rent, no lock-in to us.

> The successful projects we see share one trait: they automate ONE painful, measurable process
> first. The failed ones start with "an AI for everything" and ship a demo.

Start by writing one paragraph: the process, who does it today, how many hours it eats, and where
the source data lives. If you can write that paragraph, the project is scoping-ready - and a free
30-minute discovery call will tell you whether it is a buy, a build, or a hybrid.