How Much Does an Internal AI Copilot Cost? Real Pricing by Scope

An internal AI copilot costs between €4,000 and €20,000 as a project for an SMB, plus a monthly operating cost of €100-600 that combines inference (the tokens the model consumes) and maintenance. The range is wide because "copilot" covers everything from a conversational search layer over your documents to an assistant that also executes actions in your systems — and the jump between those extremes is one of scope, not of model brand.

This guide breaks that range down by scope level, with the numbers we use in real quotes: what each level includes, which part of the cost is project and which is recurring, and the three decisions that move the final figure the most. We run internal copilots inside our own businesses, so these ranges come from invoices, not brochures.

What is an internal copilot, and how is it different from "just using ChatGPT"?

An internal copilot is an AI assistant that knows your company: it answers from your documents, your data and your policies, and in its advanced version it acts on your systems. The difference versus handing ChatGPT or Copilot licences to staff is context: the generic tool knows about everything except your company; the internal copilot answers "what's our maximum authorised distributor discount?" with the figure from your current commercial policy, citing the document.

Technically, the core piece is a RAG system (retrieval-augmented generation): your documents get indexed and the model answers grounded in the relevant fragments, with citations. On that base you add, if the case demands it, read connections to your systems (CRM, ERP, database) and later scoped actions. Each layer adds cost and adds value — which is why the price is best understood by levels.

How much does it cost by scope? The three levels

Three caveats on the table. First: the ranges assume an SMB of 10-100 employees with a reasonable document volume (hundreds to a few thousand documents); huge or very messy corpora make level 1 more expensive. Second: the jump from level 2 to 3 isn't just technical, it's governance — defining what the copilot may do, within which limits and with what logging — and those design hours are part of the bill. Third: almost nobody should start at level 3; the sensible path is to deploy level 1, measure real usage and expand where the data says so.

What makes up the monthly operating cost?

The recurring cost has three lines, and they behave differently:

• Inference (API tokens). The most elastic line: it depends on how many people use it and how much. For a team of 20-50 with moderate daily use, expect roughly €50-300/month on current mid-range models. Good design (a small model for simple queries, a large one only for complex ones, caching of frequent answers) can cut this line in half.
• Infrastructure. Vector database, hosting, periodic indexing of new documents. For an SMB, €20-100/month — often it fits inside the cloud plan you already pay for.
• Evolutionary maintenance. Prompt tuning, re-indexing, changes when a connected system updates its API, answer-quality reviews. As with any agent, budget 10-20% of the project per year; on an €8,000 copilot, roughly €70-130/month.

The classic trap is comparing project prices and ignoring the recurring cost: a €6,000 copilot with €400/month of operation costs more over three years than a €10,000 one at €150/month. Always ask for the 36-month cost.

The three decisions that move the price most

Inside each level, three choices explain most of the spread between the low and high end of the range:

• The state of your corpus. Clean, current, deduplicated documentation makes level 1 land near €4,000; a shared drive with fifteen years of sediment pushes it toward €8,000, because the clean-up is part of the project whether it appears as a line item or not.
• How many systems get connected, and how. Each level-2 integration through a decent API adds days; each integration against a system with no API (screen exports, email-based flows) adds weeks. This is the single biggest cost driver at levels 2 and 3.
• Permission granularity. If everyone may see everything, retrieval is simple. If sales must not see HR documents and country teams have separate policies, the copilot needs per-user access control mirrored from your identity system — very buildable, but it's real engineering and it shows in the quote.

Custom copilot or Copilot/ChatGPT Enterprise licences?

They are different things, and often complementary. Per-seat licences (Microsoft 365 Copilot runs around $30/user/month; ChatGPT business plans sit in a similar per-user range) buy generalist personal assistance: drafting, summarising, searching the documents that user already has access to. For 50 employees that's on the order of €15,000-18,000 per year, every year.

A custom internal copilot solves something else: a single source of answers about YOUR operations, the same for everyone, with fine-grained control over which documents go in, traceability of what gets answered and, at levels 2-3, connection to your business systems. It doesn't compete with assisted office software; it competes with the hours your team burns hunting for internal information and asking "where was that thing again?".

The practical rule we apply: if your pain is personal productivity in Office/Google, buy licences. If your pain is scattered knowledge, slow onboarding, saturated internal support or inaccessible business data, build the copilot. Many companies end up with both, and it isn't redundant.

What return can you expect, and how do you measure it?

The business case rests on search time and internal support load. Our conservative reference math: if 30 employees save 15 minutes a day on finding information and interrupting colleagues, that's about 180 hours a month; at an average loaded cost of €25/hour, roughly €4,500/month of freed capacity. Against a level-1 build of €6,000 plus €200/month, payback lands within the first quarter even after discounting for optimism.

To keep that math honest, measure it: copilot-resolved queries per week, useful-answer rate (thumbs up/down), questions escalated to humans and to which department. Copilots that fail don't fail because of the model — they fail because of stale corpora and missing ownership: someone has to curate the documents the way someone waters the plants.

How to do it right: the process in 5 steps

1. Knowledge inventory. Which documents, in what state, with what duplicates and dead versions. This is where the real cost of level 1 reveals itself.
2. Prioritised use cases. The 20-30 questions most repeated across your company, collected from the teams themselves. That's your test bench.
3. Pilot with one department (3-5 weeks). Level 1, one team, metrics from day one.
4. Honest evaluation at 4-6 weeks. Usage and usefulness rates against the test bench. Below 70% useful answers, fix the corpus before expanding.
5. Data-driven expansion. More departments, system connections (level 2) or actions (level 3) according to what real usage demands.

To place the copilot on the full automation map — what else an agent can do for each area — see our overview of AI agent use cases by department; and to understand how any agent's price is structured (the copilot being a particular case), the reference is how much a custom AI agent costs.

Frequently asked questions

Can I build an internal copilot with no-code tools for less?

You can build a level-1 prototype for €1,000-3,000 with packaged RAG platforms, and as a way to validate internal interest it's legitimate. Its limits show up with per-user permissions, citation quality, complex documents (tables, drawings, versions) and system connections — exactly where a copilot stops being a toy and becomes a tool.

Will my company's data be used to train models?

With the major providers' enterprise APIs, no: API data is contractually excluded from training. Serious design still adds layers on top: data-processing agreements, PII filtering before indexing and, if your sector demands it, deployment options in your own cloud or on-premise.

What does inference cost if the whole staff uses it?

It scales fairly linearly with usage: as an order of magnitude, a typical RAG query costs from fractions of a cent to a few cents depending on the model. A hundred heavy users can land at €300-800/month unoptimised, and considerably less with model routing and caching. It's a line to monitor monthly, not to fear.

What about outdated documents?

They are enemy number one: a copilot that cites the 2023 policy as current destroys trust in two incidents. The fix is process, not technology: a corpus owner, automatic re-indexing of live sources and explicit expiry dates on critical documents.

How fast can it be up and running?

A serious level 1 is in a pilot department's hands within 3-5 weeks. What delays projects isn't the tech — it's the prior document clean-up, which, incidentally, is an investment your company amortises with or without the copilot.

If you want a figure for your specific case rather than a range, our AI consulting service sizes the copilot on your actual document volume and use cases, with a closed quote per level.