Who approves a €15,000 software purchase? If you ask an AI assistant this question, it will answer with confidence even when it has no basis for knowing. At QTNOVA, we put AI assistants into production in companies, where they are used every day, and we know well that the difference between a convincing answer and the right one comes down entirely to the unwritten rules that govern each individual organization.

The same question, without and with company context. Illustrative example: company, rules, and answers are fictional.
Language models know the internet, but they do not know the operational reality of your business. They do not know that your fiscal year starts in February, that a parent company with three subsidiaries counts as a single customer, or who holds internal spending authority.
On October 30, 2026, the Ramp team published a practical guide explaining how to overcome this structural limitation by building a "company brain"—a shared, AI-readable record of institutional knowledge[1].
In companies, scattered information already takes a heavy toll: according to Atlassian research surveying 12,000 knowledge workers and 200 executives, people waste 25% of their working time just searching for answers[2]. At the same time, a survey commissioned by Microsoft of 2,003 UK employees found that 71% of staff have used unapproved AI tools at work, and 51% continue to do so every week[3]. Knowledge thus remains fragmented in individual chats that end up lost as soon as an employee leaves.
Contents
- Why AI gets it wrong: the tribal knowledge no model can guess
- Common pitfalls: the fine-tuning myth and the permissions trap
- How to start: the one- to two-page context doc
- Connecting live systems: testing, ownership, and the MCP protocol
- From individual prompts to a company brain: turning knowledge into a collective asset
- The QTNOVA approach: knowledge that stays inside processes with Snake Suite AI
- Memory that doesn't disappear
Why AI gets it wrong: the tribal knowledge no model can guess
Without specific configuration, a language model knows nothing about your company. It cannot independently deduce internal conventions, policies, and the concrete steps through which work gets done every day.
This internal knowledge, often taken for granted, touches on critical points where assistants frequently make mistakes:
- Accounting rules and fiscal year. If revenue can mean booked or recognized, or if the fiscal year does not match the calendar year, the model will apply generic assumptions. The same applies to strict close deadlines.
- Customer segmentation. Deciding whether a parent company and its three subsidiaries count as one customer or four requires an internal convention an algorithm cannot invent.
- Approval thresholds and spending limits. Who approves a given purchase, above what threshold, and following which hierarchy.
- Internal policies and regional entities. Travel reimbursement rules, per diems, and holiday calendars that vary by office or country.
- Acronyms and internal terminology. Product code names, company acronyms, and team names that mean something else outside.
When these rules remain unwritten, scattered across informal conversations, or trapped in people's heads, the assistant will formulate answers that sound confident but may turn out to be wrong.
Common pitfalls: the fine-tuning myth and the permissions trap
When faced with inaccurate answers, the route many organizations consider first is fine-tuning—retraining the model on proprietary data. Yet it is the method people reach for first and need least. As OpenAI's official guide on accuracy clarifies, modifying the model serves to make its behavior, tone, or formatting consistent, whereas when proprietary knowledge, up-to-date information, or specific context is missing, you must focus on the context provided to the model[4]. Fine-tuning does not reliably teach facts, and it goes stale as soon as company rules change.
The other pitfall involves permissions management. Connectors to company tools typically inherit the access each person already has. Claude's official documentation states that “Claude inherits each person's permissions from the connected service”[5]. While this prevents an employee from viewing files they would not otherwise have access to, it also means that any file previously shared with "anyone with the link" becomes searchable through AI.
This happens in a context where risk awareness remains limited: data from the survey commissioned by Microsoft show that only 32% of workers say they are concerned about the privacy of company or customer data entered into consumer tools, and just 29% fear for the security of their organization's IT systems[3].
How to start: the one- to two-page context doc
To build a solid foundation without launching complex projects, the recommended practice is to start with a one- to two-page initial document, written as if explaining the company to a bright new hire during their first week. Just like someone joining the company for the first time, AI has no prior history.

The document to start from: one or two pages, four sections, and the date of the last update.
This starter document summarizes what the model cannot deduce on its own:
- Company snapshot and customers. What the business sells, who it sells to, total headcount, corporate entities, fiscal year start date, and criteria for distinguishing small businesses, mid-market companies, or enterprise accounts.
- Glossary and definitions. Product and team names, internal acronyms, and the exact calculation formula for each tracked metric (such as gross margin or churn rate).
- Workflows and authoritative sources. Where the official version of data lives—for example, clarifying that sales pipeline lives in the CRM and revenue lives in the ERP—who approves expenses, and operational calendar deadlines.
- Internal policies and tone. Expense limits, approval thresholds, data-handling rules, and guidelines for drafting customer emails or internal documents.
For any element subject to change over time, such as headcount or price lists, it is important to include the date of the last update to know when information is stale.
Connecting live systems: testing, ownership, and the MCP protocol
A static document sets general definitions, but the information needed for daily work lives in support tickets, meeting notes, and shared drives. The next step is to connect the assistant to the two or three tools where day-to-day decisions are actually made.

Context in two layers: a curated document for rules and definitions, connectors for live data.
To keep the system robust, follow these practical steps:
- Audit shared folders. Before enabling connectors, review broadly shared folders containing HR records, accounting files, or customer information, deciding who can turn on integrations and separating read access from write access.
- Open standards. To connect AI applications to databases and external tools, open standards exist such as the Model Context Protocol (MCP): an open-source standard for connecting AI applications to external systems, functioning much like a USB-C port for software[6].
- Test questions. It helps to prepare 10 to 20 test questions with known answers—from the margin formula to who approves a specific expense—to run every time context is updated[1]. The assistant should cite the references it draws from and be configured to state clearly "I don't know" when information is missing, rather than guessing.
- Context owners. Each section of company knowledge must have an owner responsible for keeping it current (for example, finance for accounting metrics and HR for internal policies). If the AI makes a mistake, log it and categorize it as either missing information to add or stale information to update.
From individual prompts to a company brain: turning knowledge into a collective asset
In companies, AI adoption often passes through three stages: isolated use by individual employees, shared team projects, and a shared company brain. In the first stage, the benefit remains personal and vanishes when the employee closes their browser tab or leaves the company. In the second, uploaded documents go stale if no one maintains them.

The three stages: from the individual employee to the entire company.
The transition to a "company brain" happens when workflows developed by individuals during daily work become processes accessible across the entire organization. Ramp illustrates this with its own workspace, Glass[1]: an account manager had saved a prompt that extracts feature requests directly from customer calls, requests previously copied into Salesforce by hand. That prompt was shared: now, the guide explains, it captures requests for the entire team, a second prompt prioritizes them, and a third emails customers when the requested feature ships.
Operational knowledge ceases to be an individual note and becomes a reliable procedure on which colleagues and agents can collaborate.
The QTNOVA approach: knowledge that stays inside processes with Snake Suite AI
In business, information almost always exists: PDF manuals, intranet procedures, emails, shared folders. The challenge is reaching it quickly without manually copying data and without turning experienced technical staff into a bottleneck for questions already answered dozens of times. With Snake Suite AI we built a modular architecture in which knowledge uploaded once serves all connected modules: the assistant works within a closed boundary, queries only internal documentation, always cites its source, and never touches the open web; documents remain on EU infrastructure and are not used to train public models.
Furthermore, Snake is not locked into a single AI provider: it connects to leading market models—Anthropic's Claude, OpenAI's GPT, and Google's Gemini—and for each project the best-suited one is chosen based on response quality, cost per interaction, or privacy requirements. If switching models makes sense down the road, your company knowledge base stays right where it is.
Data measured by clients using Snake every day illustrates the impact of this approach:
- Pelazza Peppino S.r.l., an elevator component specialist operating since 1974, faced a technical helpdesk overwhelmed with calls from field technicians requesting wiring diagrams. By integrating Snake Document and Snake AI Helpdesk into the Odoo ERP, 11,600+ technical drawings and 50+ manuals were indexed. The assistant provides 24-hour first-level coverage, identifies codes and warranty status, and applies an explicit safety rule: if the issue remains unresolved after 5 exchanges, it automatically opens a structured ticket for an engineer, complete with the chat transcript. The project won the ANACAM Product Award 2026 in the "Services supporting the elevator industry" category. In addition, the voice explanation recorded by the technician upon closing a ticket is transcribed and stored, enabling the AI to resolve that issue independently next time.
- Radici Products S.r.l., a benchmark in engineering polymers for over 25 years handling more than 15,000 custom orders per year, spent an average of 60 minutes processing a single quote between file analysis and manual ERP data entry. By adopting Snake Document and Snake Quote integrated into their ERP, the average time per quote dropped from 60 to 10 minutes, reducing quoting time by 83%, measured on a quarterly average.
With existing documentation already organized, an assistant is operational in days, and a working prototype on your real documents is ready within the first week; through the dashboard, you see everything it does, and it hands off to a human whenever needed.
Memory that doesn't disappear
Letting each employee query AI from isolated chat tabs means relying on answers that sound convincing but lack your company's internal rules—risking the loss of that institutional knowledge as soon as someone moves on. Building a shared company brain, on the other hand, turns policies, data, and daily workflows into a resource that is accessible to everyone, verifiable, and aligned with real decisions.
What about your company: does operational knowledge still live trapped in personal notes, or is it ready to become a shared asset? Thirty minutes, a sample of your real documents, and you can see live what answers Snake delivers: no obligation, and if it doesn't convince you, you'll know right away. If you'd rather start with a conversation, let's talk.
Want to know more? Leave us your email below and we'll get back to you as soon as possible.
Sources:
- Ramp — “How to give AI company context and build a company brain” (October 30, 2026) – ramp.com
- Atlassian — “State of Teams 2025” – atlassian.com
- Microsoft UK Stories — “Rise in ‘Shadow AI’ tools raising security concerns for UK organisations” (October 13, 2025) – ukstories.microsoft.com
- OpenAI — “Optimizing LLM Accuracy” – developers.openai.com
- Claude Help Center — “connectors” – support.claude.com
- Model Context Protocol — “What is the Model Context Protocol (MCP)?” – modelcontextprotocol.io