AI Assistant

The AI Assistant is a chat panel on every page of HARi. Ask it a question about your workspace — how many contacts you have, which deals are still open, how a feature works — and it looks the answer up in your own records before replying. It reads; it does not change anything. And it can only see the records your own account can already open.

Why it looks things up instead of answering from memory

Because a language model asked a specific, checkable question about facts it has not been shown will usually answer anyway — confidently, and wrongly. Researchers at Stanford measured exactly that on legal questions and reported that “hallucination rates range from 69% to 88% in response to specific legal queries for state-of-the-art language models”, adding that “these models often lack self-awareness about their errors and tend to reinforce incorrect legal assumptions and beliefs” (Dahl, Magesh, Suzgun & Ho, Stanford Law School / Stanford HAI, 2024).

“How many contacts do we have?” is that kind of question. The number is a fact about your workspace, and no model can know it unless it goes and counts. So the assistant is not asked to remember your data — it is given a way to read it, and every figure it states has to come back from a real lookup in the same conversation. If nothing came back, it says so instead of producing a number.

What the assistant can do today

The assistant is read-only. Everything it does with your data is a way of looking something up:

• Count and list your records — “how many contacts do we have”, “show me the open opportunities”, “which invoices are unpaid”
• Search your contacts and companies by any criteria, and filter and sort across all your data
• Open a specific record — pull up one contact, company or deal and tell you what is on it, including its field values and the records related to it
• Summarise recent activity
• Answer how-to questions — how to set up a workflow, add a custom field, fix a pipeline stage
• Suggest email copy for an outreach or follow-up message, for you to send yourself
• Reply in the language you write in

It knows the entity names your own workspace uses, so on a coworking or field-service workspace it talks about the records you actually have rather than guessing at a set of sales terms you never created.

If you ask it “what can you do?”, it answers with the list for your workspace — the things it is switched on for, and the things it is not — rather than a general description of the product.

What it cannot do

It does not create records. It does not edit or archive them. It does not send email, run a bulk update, or change any setting. There is nothing to undo, because nothing was done. It also cannot build a dashboard or set up new entity types and fields for you — those you do yourself in Settings.

If you ask it to do one of those things, it says so plainly rather than claiming it happened — for example: “I’m not allowed to change records in this workspace, so nothing has been changed.” An administrator sees an extra line pointing at Settings → AI Permissions, which is where what the assistant may do is controlled. Everyone else is told to ask their workspace administrator.

What happens when it cannot look something up

Two different things, and both are deliberate.

If a lookup fails or the data does not exist, you get a short, honest sentence and no figure — “I could not look that up just now.” A wrong number you act on is worse than no number.

And if the assistant does state a figure about your workspace that no lookup actually returned, HARi removes it before you see it and replaces the reply with: “I could not verify that against your workspace, so I have removed the figures rather than risk giving you a wrong one. Please ask me again in a moment.” This is a backstop, not the normal path — but it is the difference between an assistant that is usually right and one you can act on.

Numbers that are not about your data are untouched by this. A step number in an instruction, or a figure quoted from a help page, is not a claim about what your workspace holds, and it is not removed.

What it can see, and what it cannot

The assistant reads your workspace through the same door you do. When it looks a record up, it does so as you — your login, your security profile, your record ownership. It cannot see a record you could not open by hand, and it gains no access by being an AI.

That means two people asking the assistant the same question in the same workspace can correctly get different answers, because they can see different records. If a figure looks lower than you expect, check what your profile is allowed to see — see Permissions for how visibility is decided, and Record ownership for who owns what.

Who controls what it may do

An administrator, in Settings → AI Permissions. That screen lists what the assistant is allowed to do in your workspace and lets an administrator switch each one on or off. Your workspace is set up with reading switched on and nothing else, and that is the state this page describes.

Only administrators can open that screen or change it. If the assistant tells you it is not allowed to do something, that screen is the answer, and it is the person who runs your workspace who decides.

Where to find it

The sparkle icon in the bottom-right corner of every page — dashboard, record, list, settings. Click it, ask, close it when you are done. Your conversation stays open for the rest of the session.

Each conversation uses credits from your workspace’s monthly allowance. The panel shows how many are left underneath the message box.

A long conversation eventually reaches a length limit. When it does, HARi tells you rather than quietly forgetting the beginning, and offers to carry a summary forward or start a fresh conversation.

Related pages

• Permissions — how HARi decides which records you can see
• Search — the search the assistant uses, and how to use it yourself
• Entities and fields — the record types your workspace has, and how to add more
• Quick start — if you are setting a workspace up for the first time