Bohari AI is an in-app assistant. It answers questions from your own company's documents, helps you find your way around the product, and, where you allow it and a tool exists, carries out actions you approve.
The assistant lives in the chat screen as Bohari AI. You type a question the way you'd ask a colleague, and it answers using the documents and data already in your Bohari account. It doesn't pull from the public internet, and it only ever sees data from your own company. What it can do for any given person depends on that person's permissions, whether AI is enabled for your company, and whether tools and agentic mode are switched on for the conversation.
This guide covers what you can ask today, how answers are grounded, the controls admins set, and the sign-off step for actions that change anything.
Getting in
The chat is available to anyone with the Use AI chat and AI-powered features permission (use_ai_chat). Open it from the AI area of the app. If your company hasn't enabled AI, or has turned off the specific feature you're using, the assistant tells you it isn't available rather than guessing at an answer.
The feature selector in the chat header currently offers one option, General Chat. That's deliberate, not a placeholder. It's the honest list of what's built today.
What you can ask
There are three kinds of questions the assistant handles well today.
Questions it answers from your documents
The assistant searches your company's Knowledge Base (policies, procedures, FAQs and any other reference material you or an admin have uploaded) and answers from the most relevant excerpts, citing them as Source 1, Source 2, and so on. If nothing in your documents covers the question, it says so rather than filling the gap. Things like:
- "What's our procedure for a rejected delivery?"
- "Summarise our waste policy."
- "What does our compliance calendar say about our annual health certificate?"
"Where do I..." questions
You can ask the assistant to point you to the right part of the app: "Where do I record a stocktake?", "How do I raise a purchase order?", "Where's the invoice queue?" It answers with the actual page and a short description. If the thing you're asking about is behind a permission you don't hold, it tells you that too, instead of pretending the page doesn't exist.
Questions it answers from your live records
With tools on, the assistant can look up the records the app itself shows. Every lookup is filtered by your company and your own permissions, and every one is shown as a tool card you can inspect:
- Stock: stock on hand by location, balances, movements and transfers.
- Counting and loss: stocktakes, waste notes and the waste summary.
- Variance: the theoretical-vs-actual report for a period.
- Procurement: purchase orders, requisitions, goods receipts, supplier invoices, the three-way match and outstanding payables.
- Suppliers and items: the item catalogue, supplier register and negotiated prices.
- Recipes: recipe definitions and their current cost.
- Compliance: documents and upcoming obligations.
So "What's my variance this week?" or "Which invoices are waiting on me?" get an actual answer, with the tool call shown beside it. In agentic mode (see below) the assistant can also carry out the day-to-day actions those records support: raising a purchase order or requisition, recording waste or a transfer, receiving goods, entering an invoice, deciding documents. It also covers company administration such as settings, users and roles, exports and webhooks. Every write follows the sign-off rules below.
How answers are grounded
Document retrieval is what backs most useful answers today. When you ask a policy or general question, the assistant searches your knowledge base and pulls the most relevant excerpts into its answer, citing its sources. Admins manage what's in the knowledge base on the Knowledge Base page.
Tool calls, when they're used, show up in the conversation as a tool card with the tool's name, a status, and an expandable view of exactly what it sent and what came back. That's the assistant showing its work, rather than asserting an answer with nothing behind it.
You: Where do I record a stocktake?
[ find_feature ] Completed
Input: "record a stocktake"
Output: { page: "/inventory/stock-counts", description: "Open and manage physical stock counts" }
Bohari AI: You can start a stocktake from Stocktakes in the Inventory
section of the sidebar. Open a new count there.
Tools and agentic mode
Two switches in the chat header control how much the assistant can do.
| Switch | What it does | Who sees it |
|---|---|---|
| Tools | Lets the assistant call read-only, low-risk tools (wayfinding and the like). With tools off, it can still answer from retrieved document excerpts, but it won't attempt a lookup. | Everyone with chat access |
| Agentic | Adds higher-risk read tools and any write/action tools on top of the read-only ones. | Only users with the use_ai_agentic permission |
Every tool the assistant can call is tagged with a risk level (low, medium, high or critical) and with whether it reads or writes. Without agentic mode, only low- and medium-risk read tools are ever offered to the model at all. Anything higher, and any write, needs use_ai_agentic on top of your own permission for that specific action.
Sign-off on actions
Actions never run silently. Two things stand between a request and a change being made:
- The assistant explains first. Before any write, it tells you what it's about to do.
- Higher-risk actions need your approval. For anything above low risk, the assistant shows an approval banner with a preview of exactly what will happen and Approve / Deny buttons. Nothing is committed until you click Approve.
[ update_company_settings ] Approval needed
This tool requires your approval to execute.
Review the input below and decide. [ Approve ] [ Deny ]
If you deny, the action is dropped and the assistant moves on.
Per-company AI controls
Admins (with manage_ai) set how the assistant behaves for the whole company on the AI Settings page. The settings that shape your experience:
| Setting | What it controls |
|---|---|
| AI enabled | Master switch for the assistant; off until enabled |
| Provider | Which AI provider runs the assistant |
| Model | The specific model used |
| Monthly token budget | A usage cap; once reached, the assistant pauses for the month |
| Redacted fields | Which sensitive-looking field types are stripped before anything is sent to the provider (on by default) |
| Custom instructions | Extra guidance the assistant follows in every reply |
Because the provider is configurable, the assistant always presents itself as Bohari AI rather than naming the underlying model. Your company chooses the engine; the experience stays consistent.
See AI Privacy & PII Handling for the full detail on redaction, what's logged, and who can see it.
FAQ
Does the assistant search the internet? No. It answers from your company's own documents and, where a tool exists, your company's own records.
Why does it say it can't look something up? Usually because tools are off for the conversation, the feature isn't enabled for your company, or your own permissions don't cover that record. The assistant can never read past what you yourself may see.
Can it see other companies' data? No. Document search and every tool call are filtered by your company, and tools run with your own permissions.
Can it change things by mistake?
With tools on but agentic mode off, it's limited to low- and medium-risk reads. Any write needs the use_ai_agentic permission, your own permission for that specific action, and your explicit approval.
How do I know where an answer came from? Document-based answers cite their sources (Source 1, Source 2). Tool-based answers show a tool card with the exact input and output, which you can expand and inspect.
What happens when we hit the token budget? The assistant pauses AI requests for the rest of the month and tells you the budget is exceeded. An admin can raise the monthly token budget in AI Settings.