AI Mode

A second way to use Bohari on the web, built around the assistant: an inbox of what needs you, a chat docked beside every page, a search palette, and proposals you approve with one click. The Classic interface stays exactly as it was.

Bohari has two interfaces on the web, and each person picks the one they prefer. Classic is the sidebar-and-pages layout that most of this documentation describes. AI mode puts the assistant at the centre: you open on an Inbox of what is waiting for you, you ask questions in a chat that follows you from page to page, and when the assistant wants to change something it shows you a card to approve or reject. Both interfaces run on the same data and the same permissions. Nothing you can do in one is hidden from the other, and the assistant can never do more than you are allowed to.

This guide covers how to get into AI mode, what is on the screen, how actions are approved, and the controls around it. For what you can ask, see Using the AI Assistant; for what is masked before a prompt leaves Bohari, see AI Privacy & PII Handling.

Switching between AI and Classic

If AI mode is open to you, a Classic / AI mode switch appears in the top bar. On narrower screens it shows icons only, with the words as tooltips. Your choice is saved against your account, not your browser, so it follows you to every device. Each person chooses for themselves; switching does not affect anyone else.

AI mode is open to a person in a company only when all of these are true:

• The company has the AI module on its plan.
• AI is switched on for the company (see AI Privacy & PII Handling).
• Bohari has opened AI mode to the company. Bohari can open it to every company with AI or only to a list of companies, so a company that has AI may still see Classic only.
• You hold the Use AI chat and AI-powered features permission (use_ai_chat).

If any of these is missing, you see no switch and you stay on Classic. Nothing is lost: if the company is added later, the switch appears. If a person who chose AI mode later loses access, they are shown Classic. The phone app has no AI mode and no switch (see the end of this guide).

What is on the screen

The rail. In AI mode the sidebar is replaced by a short rail of the places you use most: Inbox, Ask, and links such as Stock, Counts and Buying, showing only the ones you are permitted to open. All pages opens the search palette, which lists everything.

The Inbox. Your starting point. Under Needs you it lists the things waiting for your decision, including AI proposals, approvals and undecided items, and under Done today the proposals decided today. When nothing is waiting it says so.

The Ask page. The full chat, with your past conversations. This is the same chat that Classic has, with the same tools and the same controls.

The dock. A chat docked beside every other page. It is the same conversation as the Ask page, so an answer can send you to a page and the conversation is still there. On a record page it tells the assistant which record you are looking at, unless you switch that off, and its starter questions suit the page. Ask Bohari in a record's header opens it about that record.

The search bar and palette

AI mode has a search field in the top bar and a keyboard shortcut (Ctrl K, or Cmd K on a Mac) that opens the same palette. It is AI mode's list of every page, and the one place quick actions live. Type to filter pages and actions, and press Enter to go.

The palette follows the same rules as the rest of the app. Every page and action it shows is one you may open or do, and an action that leads to a list page also needs that page's permission, so it never lists something you would be refused. Recent choices are remembered per person and per company, and a recent only appears while it is still available to you. Classic has no palette.

Proposals: Approve and Reject

When the assistant wants to change something, it does not just do it. It proposes. A proposal is a card that says what will happen, with a real preview computed from your data, and two buttons: Approve and Reject. Nothing is committed until a person clicks Approve. If you reject, the change is dropped.

Proposals appear in the chat where you asked for them, and they also wait in your Inbox under Needs you, so one you did not get to is still there later. Each card shows who proposed it (you, through the chat, or a connected agent by name). The Inbox also shows what was decided today.

Approving uses your permissions. If you are not allowed to do the thing the card describes, the approval is refused, and any separation-of-duties rule that normally applies (the person who entered a supplier invoice cannot approve it, for example) applies exactly the same way.

Agentic mode

By default, the assistant can only run safe reads: looking things up, answering questions and finding pages. To let it propose changes, a person has to turn on Agentic mode with the Agentic switch in the chat's options, which only works while Tools is on. Two things are needed:

1. You hold the Enable agentic mode (write/action tools) in AI chat permission (use_ai_agentic). Without it the switch is not shown at all.
2. You flip the switch, in the chat you are using. It is a deliberate step for that conversation.

Without both, only safe reads run, whatever the assistant is asked. With both, the assistant can propose writes, but each one still waits for Approve. Agentic mode does not give the assistant any access you do not have, and it does not skip approvals; it only lets it propose. A conversation that used agentic mode is marked with a small bolt in the conversation list.

Files in chat

If your company allows it, you can attach files to a message so the assistant can read them, for example a supplier's price list or a count sheet, and propose importing it. This is controlled per company, by an administrator with the AI settings permission, under File uploads in chat on the AI settings page.

• It is off by default, and turning it on asks you to confirm, because file text is sent to the AI provider.
• It only works while privacy redaction is on. The file's text is scanned and sensitive-looking values are masked before the assistant reads it, the same as for anything you type. If redaction is off, uploads are refused.
• A company that asks to hide names or addresses needs the deployment's stronger privacy check for files. If that check is not available, files are refused rather than stored unmasked.
• Accepted: PDF, Word, Excel and CSV files, and plain text files. A message can carry up to three files, up to 10 MB for PDF and Word, 5 MB for Excel and CSV, and 2 MB for text.
• An unsent file is removed after 24 hours. A sent file's content is removed after 90 days.

Files in chat are available on the web only, not on the phone.

AI cost alerts

Running an AI model costs money, and a few people using it heavily can cost more than expected. The platform watches this for you: each morning it checks the previous day's estimated AI cost per active user for every company that has AI mode, and when a company goes over the platform's limit, the platform team is emailed so someone can look at it. This is a platform-level check, not something your team receives. What your team does receive is an alert to the people who manage AI settings when the company's own monthly token budget is spent. See Alerts & Notifications. While your company uses its own provider key, your admins set that budget on the AI settings page; while Bohari's key runs your AI, Bohari sets it.

Connected agents

An outside AI tool, such as Claude Code or Claude Desktop, can connect to Bohari with a connection token and work on your behalf. You make the token yourself under Profile, then Connected agents:

1. Give it a name (so you can tell your agents apart).
2. Choose its reach: Read only (look up stock, counts, orders and reports), or Read and propose changes (each one waits in your Inbox for you to approve). The second option needs the use_ai_agentic permission.
3. Optionally set when the token expires.
4. Copy the token when it is shown. Bohari keeps only a fingerprint, so it is not shown again.

You can Revoke any token from the same list, and it stops working at once. On the phone, Profile also lists your connected agents and can revoke one, but a new token is made on the web only. The tokens belong to you and to one company.

The key rule: an agent can only propose, and a person approves. A token can never approve or reject anything, and the agent has no way to pass its own approval. Its proposals arrive in your Inbox with the agent's name on them, and you decide. A token also never gives the agent more access than you have. Do not give your own sign-in session to a third-party tool; use a token.

Creating a token tells you and the people who manage users, by an alert that cannot be muted. When an agent proposes something, you are alerted too. See Alerts & Notifications.

On the phone

The Bohari phone app has no AI mode switch and keeps its own layout. Two parts of the AI work are on the phone, both on the Services screen (the centre tab), under AI & Tools:

• AI Inbox. The proposals waiting for your decision, each with Approve and Reject, and what was decided today. This is where you approve a change that was proposed on the web or by a connected agent.
• AI Assistant. The chat. It streams answers and shows proposal cards, but it reads only: the phone chat does not have the agentic switch, so it cannot make a proposal of its own. To make a change, propose it on the web and approve it from the phone Inbox.

Both need the AI module and the use_ai_chat permission. Chat file attachments are not on the phone.

FAQ

I do not see the Classic / AI mode switch. AI mode is not open to you in this company: the AI module, company AI setting, platform list or your use_ai_chat permission is missing. Ask your administrator.

Can the assistant change anything without my click? No. It can propose only with agentic mode, and every proposal needs a person's Approve.

Does an agent I connected see everything? It sees what you can see, through a token you can revoke. Without the propose option it cannot propose anything.

Do the two interfaces show different data? No. They are two views of the same company, with the same permissions.