Service 02

Custom AI assistants for business

The answer to "what terms did we agree with this client" sits in three files, two chats and the head of someone who is on holiday today. An AI assistant finds it in seconds, in your documents, your CRM and your accounting system, with the source attached.

50+ projects delivered
2–8 weeks to a working version
RAG answers from your documents
GPU private server for models

Who gets hours back

Sales teams

A manager asks about a client's terms, stock on a line, or the detail of a past deal, and gets an answer without opening five tabs. Quotes are assembled in minutes instead of being written from scratch.

Management

Summaries across CRM, ERP and spreadsheets from a single question. We built CRM analytics as a standalone project: sales per manager appear in a dashboard with no manual exports.

Support and service

The assistant finds the right clause, the warranty terms or the client's history and drafts a reply that only needs checking.

New hires

Instead of a month of asking how things are done here, people ask the assistant and get an answer from the current document.

Customers directly

The same assistant answers customer questions that have verified data behind them and hands unusual cases to a person.

Assistant, chatbot and agent: the difference

  1. 01 A chatbot is a conversation interface. It tells a customer what you sell and on what terms.
  2. 02 An AI assistant works with information on a person's request. The manager asks, the assistant searches the CRM, drafts the proposal, pulls the numbers together. The decision stays with the person.
  3. 03 An AI agent is given a goal and carries it through. A new enquiry arrives: check the data, update the CRM, create the task, prepare the next contact, all inside the permissions you set.
  4. 04 These are not exclusive: a chatbot is often the interface to an assistant, and an assistant can hold tools and perform some actions itself.
  5. 05 If you need to answer customers, start with a chatbot. If you need staff to work faster with company knowledge, start with an assistant. If a process has to run across several systems without a person, you need an agent.

Show us where your knowledge lives today

Policies in Google Drive, price lists in Excel, client terms in the CRM, and half the practice in people's heads. We will look at what can be connected to an assistant right now and what needs tidying first.

Talk about your knowledge base
Project

A vector knowledge base for AI

The task was to make the AI answer from the company's real documents rather than general knowledge, and to make every answer checkable.

Sources
Documents, policies, price lists and product descriptions are connected as knowledge sources.
Indexing
Material is split into fragments and stored in a vector database.
Retrieval
Before answering, the system finds the fragments relevant to this specific question.
Answer
The model composes the answer from what it retrieved, not from what it happens to remember.
Access rights
Each person sees only what their role allows.
Own infrastructure
For data that cannot leave the building, we stood up a GPU server: models run on your hardware with no dependency on external APIs.

The AI answers from the company's real documents, and staff can see which source each answer came from.

All projects →

What it connects to

The integration set is decided during the data audit: what has an API, what shape the sources are in, and who is allowed to see what.

How we roll it out

  1. 01

    Reading the task

    We take the questions staff ask most often and find where the answers live.

  2. 02

    Data audit

    We check the sources: what is current, what is duplicated, what contradicts itself.

  3. 03

    Architecture and rights

    Role-based access, logging, and the line where the decision stays with a person.

  4. 04

    MVP on one team

    We launch the assistant for a single department on their real questions.

  5. 05

    Answer testing

    A control set of questions with known correct answers, run before anyone relies on it.

  6. 06

    Expansion

    New sources, new departments, CRM and ERP integrations.

What it costs

The price comes from the volume and state of your sources, the number of integrations, how strictly access has to be separated, and whether the data needs its own infrastructure.

After a scoping call you get an estimate in three shapes: an MVP for one department, a pilot on real questions, or the full solution with integrations. A pilot usually takes 2 to 8 weeks.

Get an implementation estimate

Questions that come up before this decision

Can the assistant work only from our documents?

Yes, that is how we build it: answers are grounded in the corporate sources you connect. We also build in quality checks, a control set of questions with known answers, so you can see where the assistant is wrong before your staff do.

Do we need to train our own model?

No. For most tasks an off-the-shelf model connected to your data through RAG and APIs is enough. A private model makes sense in specific cases, and we then run it on your own GPU server.

Can the assistant perform actions on its own?

Yes, given the tools, permissions and rules. Once actions start chaining without a person in between, that solution is usually called an AI agent, and we build those too.

Can it connect to our ERP?

Yes. We have integrated with Odoo, 1C and BAS, and with in-house accounting systems over their APIs. The assistant reads current data on orders, clients and stock rather than a month-old copy.

Is it safe to give AI access to internal documents?

Access is separated by role, every request is logged, and for data that cannot leave the building the model runs on your own GPU server. We have built and maintained exactly that setup.

How long before we see something working?

A working pilot for one department usually takes 2 to 8 weeks. The first thing you see already answers your real questions from your own documents.

Other services

Not sure which of these you need?That is what the first look at the task is for

Name and phone. A line about the task helps but is optional.

  • We reply within one business day
  • A short scoping call instead of a long sales process
  • We work with your stack; happy to start with a pilot

Rather not leave a number? Telegram or WhatsApp

Get your plan