AI Assistant for Company — Internal Chat Over Your Data
An AI assistant for company use is an internal chat interface that lets employees ask natural-language questions across SharePoint, Google Drive, CRM, ERP, contracts and HR policies, and get a cited answer in 4-7 seconds. It runs on your private data via RAG (Retrieval-Augmented Generation), enforces role-based access control from Active Directory, and stays on-prem or in an EU data center — no document ever leaves your infrastructure.
What the internal AI assistant replaces
An average knowledge worker loses 1.8 hours a day searching across Slack, shared drives, email and archived PDFs. We deploy an AI assistant for company workflows that indexes those sources once and answers "what is our B2B return procedure" in 5 seconds — quoting the SOP, not inventing it. Measured outcome: 30-45% fewer internal tickets to IT, HR and finance, and 6-8 hours saved per employee per week.
- ✓RAG over private documents — we index SharePoint, Google Drive, Confluence, Notion, file servers, ERP, CRM, helpdesk and email; the chat answers only from those sources and always returns a link to the original.
- ✓Role-based access control — the assistant inherits the asker's permissions from Active Directory or Google Workspace; finance reports never leak to sales, HR files never leak to marketing, even when they share a vector store.
- ✓EU or on-prem hosting — Frankfurt, Ireland, Sofia or your own server; data and embeddings stay in the chosen region, covering GDPR and internal classification policies without extra DPAs.
- ✓Slack, Teams and web app in one — employees ask where they already work; one backend serves a Slack bot, a Teams tab, mobile chat and an admin web app with no duplicate knowledge.
- ✓Citations and traceability — every answer ships with 2-4 citations down to the page and paragraph; the bot never replies without a source, which kills hallucinations and preserves an audit trail.
- ✓Actions, not just answers — files a Jira ticket, opens a contract in DocuSign, creates a lead in CRM, books a room; the assistant is a front door into business process automation, not a search box.
Who it is for
A technical director asks "what is the tolerance for material X under our 2024 ISO 9001 procedure" and gets the exact paragraph from the QMS archive in 6 seconds. Replaces 40 minutes of SharePoint digging per shift — measured 12-18% drop in line downtime.
A lawyer asks "what does clause 7.4 of the Acme contract from March 2025 say" — the assistant finds the PDF, quotes the clause and compares it to your template. RBAC ensures a junior never sees files outside their matters.
A new developer asks "how do I deploy staging on project Y" and gets the Confluence runbook plus the last three Jira post-mortems. Onboarding drops from 6 weeks to 2, and senior knowledge survives departures.
How we build it
We do not resell a SaaS platform. Each rollout is custom software on our own backend — your embeddings, conversations and prompts stay with you. You can swap the LLM (GPT-4, Claude, Mistral, on-prem Llama) at any time without migration or vendor lock-in.
1. Source mapping and classification
In the first 5 business days we map every place where company knowledge lives — SharePoint, Drive, ERP tables, CRM fields, email archive, Confluence, ticketing — and classify sensitivity (public, internal, confidential, restricted) before anything is indexed.
2. RAG pipeline and vector store
We build the retrieval layer: source connectors, parsers for PDF/DOCX/XLSX/email, chunking with overlap, a multilingual embedding model and a vector store on PostgreSQL with pgvector or Qdrant. Incremental sync runs hourly — a new contract is queryable within an hour.
3. Identity, RBAC and audit
We wire SSO via Azure AD, Google Workspace or Okta. Every query is permission-filtered at vector-search time, before any chunk reaches the model. Each conversation is logged with timestamp, question, cited documents and answer for full audit.
4. UI: Slack, Teams and web app
One backend, three interfaces: a Slack bot with a slash command, a Microsoft Teams tab with adaptive cards, and a self-hosted web app for staff without Slack or Teams. Same answers, same citations, same permissions everywhere.
5. Eval, feedback and continuous learning
Before launch we run 100-200 real questions through an eval suite — accuracy, citation rate, latency. After launch every answer has thumbs up/down; bad ones loop back as new instructions or extra documents. The assistant improves every sprint, no retraining.
Why Saitami
Fixed EUR pricing — no seat subscriptions, no per-conversation fees. For broader scope see AI agent for business processes that acts autonomously, or AI chatbot for customer service for the external channel.
Frequently Asked Questions
Does our data leave the company when using the AI assistant?
Not necessarily. Three options: full on-prem deployment of an open-source model (Llama 3.1, Mistral) on your GPU server with zero outbound traffic; EU-only hosting via OpenAI or Anthropic with a signed DPA and zero-retention mode; or a hybrid where sensitive documents are processed on-prem and the rest go through the EU API. The choice is made before indexing, not after.
How does role-based access really work?
Vector search filters by permission tags before chunks ever reach the LLM — we do not rely on a "please do not share" prompt instruction, which is a known hole. Tags come straight from the source system: if a SharePoint file is HR-only, only HR users see it in the chat, exactly as in SharePoint. Citations are filtered too.
How long from kickoff to production?
Standard timeline is 4-6 weeks for a company with 3-5 knowledge sources and 50-300 employees. Weeks 1-2: source mapping and connectors. Week 3: RAG pipeline, embeddings, RBAC. Week 4: Slack/Teams/web UI and SSO. Weeks 5-6: eval, 10-20 user pilot, prompt tuning, production. On-prem with GPU procurement adds 2-4 weeks.
What does it cost — one-time and monthly?
One-time from €4,200 for the standard package — 3 sources, RBAC via one SSO, Slack or Teams bot, admin web, up to 200 users. Multi-source with 5+ systems, on-prem or multilingual indexes — €7,800 to €18,000. Monthly: API tokens (€80-€450 by volume), hosting €40-€180, no seat fees. For 100 employees that is under €5 per person per month — an order of magnitude cheaper than Microsoft Copilot or Glean.
Can we extend it to autonomous actions later?
Exactly why we build it modular. The same RAG layer and SSO later power an AI agent that not only answers but acts — places an ERP order, replies to a client email, reviews a new contract. Start with an assistant (low risk, fast ROI), in 3-6 months upgrade to an agent on the same infrastructure without throwing anything away.
Ready for an internal AI assistant that knows your company?
Send a list of 3-5 knowledge sources (SharePoint, Drive, CRM, Confluence, ERP) and 10 real questions your team asks every week. Within 5 business days we return a plan, a price and a proof-of-concept trained on a subset of your documents.
Request an AI assistant POC →Related services: AI agent for business processes · AI chatbot for customer service · custom software