The assistant
It shows its sources and asks first
Answers link to what it read; actions wait for approval.
What did we agree on Alborz's working capital adjustment?
The normalised working capital peg was set at 364.9 B T, based on the twelve-month average and excluding the related-party receivable of 41.2 B T.
it asked first — and the yes stays a person's
How it works
Retrieve, then answer — never the other way round
Hybrid search
A keyword index verified against Persian text, with semantic embeddings fused in when an embedding model is configured.
Ten kinds of source
Documents, feeds, clients, findings, engagements, contacts, meetings, records, tasks and indicators — each result carrying the canonical link to open it. Restricted records are not among them.
Answers with citations
The answer names what it read. If the sources do not support an answer, saying so is the correct output.
Plans, not surprises
When work is required, you get a plan. Each step is an action from a fixed whitelist, and each waits for approval.
Runs as you
An approved action executes with your permissions, inside your workspace. It cannot reach anything you could not.
Metered and logged
Every model call writes a row to an append-only, per-workspace usage ledger — visible to workspace admins.
It writes records too
It can log a decision, a blocker or a risk — through the same door people use, only after you approve, and carrying the machine mark permanently.
Access
It sees with your eyes, not the firm's
Every retrieval arm reads with the asker's own permissions — engagements, contacts, meetings and records alike, each filtered to what you personally may see. That filter lives inside the query, not after it.
The difference is practical: filter after retrieval and someone with narrow access loses their results to rows they were never allowed to see, and comes back empty-handed. Here those slots go to them in the first place.
Writing works the same way: an approved action runs with the approver's eyes, through the same door a person uses — and the assistant can never create a confidential record.
- An engagement can be team-private; to everyone else it is the same "not found" a nonexistent id gets
- A new contact belongs to its author until it is shared or visibly widened
- Firm-wide counts and totals stay open — the row hides, the number does not
What did we agree on Alborz's working capital adjustment?
The normalised working capital peg was set at 364.9 B T, based on the twelve-month average and excluding the related-party receivable of 41.2 B T.
it asked first — and the yes stays a person's
The boundaries
What it will not do
The limits are the feature.
- It will not act without an approval, even for a step it proposed itself
- It will not reach data outside your workspace, or outside your own permissions
- It will not answer from memory when the sources do not support it
- It will not silently overwrite a validated page, a reviewed spread or a stage history
- It will not hide what it cost — usage is ledgered per workspace
Why it earns its place
The firm's memory, queryable
Three years in, the useful question is rarely "what is the current ratio". It is "what did we conclude about this client's related-party rent, and where did that come from?"
That answer exists in your documents, your notes and your findings. The assistant's job is to find it and show you the page — not to reword it convincingly.
- Search that spans documents and conversations
- Answers that link to the exact source
- Actions that stay inside the audit trail
What did we agree on Alborz's working capital adjustment?
The normalised working capital peg was set at 364.9 B T, based on the twelve-month average and excluding the related-party receivable of 41.2 B T.
it asked first — and the yes stays a person's
About the assistant
Does our data train someone's model?
No. Documents and text are sent to the model endpoint you configured for inference only. If your policy requires it, point the deployment at a model inside your own network and nothing leaves at all.What if we do not configure embeddings?
Search degrades gracefully to the keyword index, which is verified against Persian text. Answers stay cited; recall on paraphrased questions is lower.Can it create records?
Only through a fixed whitelist of actions, only as the person who approved the step, and only inside that person's workspace and permissions.
See it read your own document
Bring one file you already know well. Thirty minutes is enough to tell whether this belongs in your firm.