What an AI assistant for a sales team costs, and where the budget goes
The budget for an AI sales assistant: an MVP in 4–8 weeks, a full project in 3–4 months. Why 60–70% of the estimate has nothing to do with the model, what inflates a project quietly, and how to tell a company is not ready.
In short. A full AI sales assistant project takes 3–4 months to production. An MVP that answers questions about the catalogue and drafts a quote takes 4–8 weeks, provided the company can give access to the catalogue, examples of past quotes and 20–50 typical questions its sales people ask. What moves the number inside that range is the size of the catalogue, the number of integrations and the security requirements, not the choice of model.
We quote in roubles, so the figures below are given as proportions and timelines rather than converted amounts — convert at the rate of the day if you need an absolute number, or ask us for a quote.
Why the model is the cheapest part
One figure explains most of the estimates that went wrong: 60–70% of the budget has nothing to do with the model. The money goes into preparing data, integrations and access rights.
| Line | Share of budget | What is inside |
|---|---|---|
| Preparing the catalogue and documents | 25–35% | Parsing price lists, cleaning descriptions, chunking, metadata |
| Retrieval and answers | 15–20% | Vector index, hybrid search, citing the source |
| Integrations | 15–20% | CRM, accounting system, mail, messenger |
| Access rights and quality | 15–20% | Role model, answer evaluation, logging, handover to a human |
| Rollout and training | 10–15% | Pilot on part of the team, collecting corrections, fixes |
Choosing between hosted models and a local one moves the estimate noticeably less than the state of the catalogue does. The ranges for the retrieval part alone are in what RAG costs.
Where projects actually stall
The demo is convincing: the assistant finds the right item, cites the source, answers coherently. The project moves into trial use and stops a week later. The reason is almost never the quality of the answers.
What stops it is this: a sales person turns out to have access to data they should not see. Purchase prices, another dealer’s terms, the discount matrix for key accounts. In a sales department that is not an abstract risk, it is a direct conflict of interest.
Access rights rarely look impressive on a slide, and they are exactly what separates a prototype from a working system. If every sales person in the demo has the same set of rights, you are not looking at a system.
What moves the estimate most
The state of the catalogue. This is the main factor. A catalogue exported from the accounting system with full descriptions and attributes is almost ready to work with. A catalogue where half the items are described as “pipe 20” and the attributes sit in attached PDFs needs a preparation stage of its own, and that stage easily becomes the most expensive part of the project.
The number of integrations. An assistant that only answers is cheap. An assistant that writes the result into the client record in the CRM, checks stock in the accounting system and emails the quote costs several times more, because each of those places has to be designed and tested against real data.
Security requirements. If the data has to stay inside the perimeter, a local environment is added: your own infrastructure, updates, responsibility. That is justified when a written policy requires it, and almost never when there is only anxiety.
What inflates a project quietly
Nobody to accept the result. You need a working sales person who will say which answer they would send to a client and which they would not. Budget 20–30% of that person’s time for the length of the project. Without them the assistant learns to answer correctly from a developer’s point of view rather than a seller’s.
No examples of what counts as a good answer. Twenty real client conversations are worth more than a month of discussion. If they were not collected before the start, they will be collected during it, and that time goes into the estimate.
Expecting the assistant to replace the sales person. It does not. It removes catalogue lookups, the assembly of routine quotes and the answers to repeated questions. Projects started with a different goal fail for reasons that are not technical.
What to check before you start
Three questions worth asking yourself before you ask a vendor:
- How many hours a week the sales team spends searching the catalogue and assembling routine quotes. If it is less than a few hours per person, the project will not pay for itself.
- Whether a catalogue export exists and can be looked at today.
- Who inside the company will say that an answer is correct.
The full preparation checklist is in what to prepare for an AI sales assistant.
Cost of ownership
The benchmark is 15–20% of the development cost per year: refreshing the index when the catalogue changes, tuning for new questions, paying for the model or the infrastructure, fixes.
Worth remembering separately: the catalogue changes constantly, and a system without a regular index refresh starts answering with old prices. That is the worst kind of failure, because it is invisible until a client receives a wrong offer.
Common questions
How soon is the first result? 4–8 weeks with a ready catalogue and collected examples. That means a working MVP for one scenario, not a demo.
Could a subscription to an off-the-shelf service do the job? Sometimes. If you only need answers from publicly available material and no integrations, a ready service is cheaper. Custom development starts where you need access to internal data with different rights for different employees.
Do we need a local model? Only if the data has to stay inside by policy: critical infrastructure, defence contracts, strict compliance. The numbers are in a local LLM for B2B.
Why is the range so wide? Because the main unknown until the catalogue has been looked at is how much work the data preparation is. That is 25–35% of the budget, and it cannot be estimated from a verbal description.
What decides the number
The position inside the range is set by the state of the catalogue, the number of integrations and the security requirements. The model is not on that list, and that is the main thing worth knowing before any conversation about budget.
If the task is already defined, look at how we build AI assistants for sales and write to us. We will say plainly if there is not enough data for the project yet.
Facing a similar problem?
Tell us what you are building. We will walk through the architecture and give you a budget range in one call.