One corporate AI chat for everyone is a bad idea

“Is one bot for everyone going to wander into HR and finance again?”

Our controversial take after WikiMind: it’s better to split a corporate AI chat into role-based modes from the start than to try to build one universal assistant for all employees.

Yes, a single chat looks cleaner in the interface. One button, one conversation, fewer explanations for the user. But in a corporate platform, another question quickly comes up: what context is the assistant allowed to see, and who is it supposed to answer in the first place.

In WikiMind, we implemented four AI chat modes:

• a corporate AI assistant for the company knowledge base;
• ChatGPT for general tasks;
• an HR assistant for people policies;
• an FD assistant for finance.

The HR team sees the HR bot. The finance team sees the FD bot. The other modes are available under their own rules. That is less “beautiful” from a product simplicity standpoint, but clearer in terms of accountability.

Why we didn’t build one smart interlocutor

The main argument against splitting is usually this: the user shouldn’t have to think about where to ask the question. Let them ask in one window, and the system will figure out whether it needs HR context, financial context, or the company knowledge base.

That sounds good in a demo. In a production product, an awkward gray area appears.

If an employee asks about vacation, sick leave, or an internal policy, that is an HR domain. If the question is about budgets, payments, or finance rules, that is an FD domain. If the question is general, the company knowledge base or ChatGPT mode is enough. For a person, the boundaries may be fuzzy, but for access control, they have to be strict.

We didn’t want request routing to become a hidden part of security. When the mode is chosen explicitly, it is easier to explain the behavior: this bot works with people policies and is available to HR; this one works with financial context and is available to finance; this one is for the general knowledge base.

A real detail: the longest argument was not about prompts, but about the idea of splitting things up at all. One interface seemed simpler. But once we started mapping questions to content owners, the universal chat started to look like a place where too many decisions are made silently.

Roles matter more than “smart” answers

In corporate AI, there is a temptation to optimize for how polished the response sounds: longer, more confident, more polite. But for internal services, the first constraint is often not the style of the answer, but the right to use that mode at all.

In WikiMind, we tied modes to roles. HR sees the HR assistant, finance sees the FD assistant. That does not replace the other platform checks and settings, but it removes the odd situation where a user sees a bot they shouldn’t have any working scenario with.

Each mode also has its own expected answer type. The HR assistant should stick to people policies. The FD assistant should stick to finance rules. The corporate assistant should use the knowledge base. The ChatGPT mode covers general tasks where no internal specialized context is needed.

This approach is less flashy than “one AI for the company.” But the owner of each area understands which mode they are responsible for. HR can review how the HR bot answers separately. Finance can analyze FD scenarios separately. Errors do not get smeared across one big chat.

Threads and reactions are not just for looks

We added threaded chat history. For a corporate assistant, that matters more than it seems: the user returns to a specific topic instead of scrolling through one long feed of mixed questions.

We also added answer quality ratings with like and dislike buttons. Not as decorative controls. These reactions are there to improve prompts based on real conversations: where the assistant answers too broadly, where it drifts into the wrong context, where users expect a different format.

These ratings are especially useful for role-based modes. A dislike in HR mode and a dislike in FD mode are different signals. In one case, the problem may be how a people policy is phrased; in the other, it may be the finance prompt or the expected answer shape.

We also added AI-generated illustrations for documents in WikiMind. That is a separate scenario, but it fits the same logic well: AI tools inside a corporate platform do not have to live behind one big button. They have different contexts, different users, and different risks.

Where this approach may be unnecessary

We may be wrong for small teams that do not separate HR, finance, and the knowledge base, and where access is simpler. In that case, four modes will look heavyweight: the user will start choosing between similar doors and make mistakes more often.

But if a platform already has roles, content owners, and sensitive sections, a single AI chat starts to require hidden rules. Someone has to decide when to attach HR context, when to attach financial context, and when to use the general knowledge base. We preferred to make that explicit, as separate modes tied to roles.

The next practical step after launching such modes is to look at dislikes for each assistant separately and adjust prompts not “for AI in general,” but for a specific work scenario.