The assistant

Asking the assistant

When your company has AI on and your role includes it, a Chat button sits in the sidebar (in the menu on a phone). It opens the Chat page, which uses the whole width, with the assistant in it. The assistant answers for the page you opened Chat from, which offers questions it can answer, such as "What needs my attention today?" or "Which RFQs are still open?"; tap one, or write your own.

It can tell you what needs your attention, record by record, the same list as the Home's "Needs your attention" (approvals to decide, RFQs to chase or enter or send for approval, purchase orders to issue or send; the urgent first, each linked to its page), and look at your work and your RFQs, what needs chasing on open RFQs (the same list as Follow up: RFQs due and short of quotations, reminders due, replies to enter, vendors' questions and phone prices waiting; yours, or everyone's if you ask), what is waiting for your approval, purchase orders (with their totals), materials (their unit and category, never a price), vendors (who is active, who is blocked and why, what they supply and which documents they still owe), last purchase prices (what you last paid for a material, from which PO, to which vendor) and spend, each only if your role may open it on its own page: a requester is offered what waits for their decision, what needs their attention and the materials, and purchase orders, last prices and spend are offered only to people who see prices. A vendor's contact details (email, phone, address) are never put in an answer: open the vendor's page for them.

• Answers come with cards. The records it found are shown as numbered cards, drawn from the records themselves: the numbers, prices and dates on a card are exactly the record's. The assistant's own words point to a card ("the second card is due soonest"). Click a card to open that record or page, and click a figure on it (such as the number of open RFQs) to open the list it counts. Chat brings you back to the conversation you were in.
• It only reads. It never changes anything, sends anything or approves anything.
• It can be wrong. Under what you write it says so, and who can read the conversation: check anything important against the record before you act on it. Your company chooses the words of the warning.
• It sees what you may already open, and no more. A requester is never shown quotations' prices, and someone without access to reports gets no spend figures, however the question is worded.
• Thumbs under an answer tell your company whether it helped. A thumbs down asks, once, what was wrong (you may leave that empty).
• Your conversations are listed on the left of the AI tab, the newest first, as the team's chats are: click one to carry on from where you stopped, or New to start another. On a phone the list and the conversation take turns. You cannot delete a conversation. Where your company keeps no words there is no list: each question stands alone, and New conversation (at the top) clears what is on screen.

• It can look in your company's own documents. When your admin has added policies, handbooks or terms of business, ask about them in your own words ("how many days of casual leave do we get?"). The assistant shows the passages that answer as cards, each with the document's name and how close it is: click a card to read the document. If no document answers, it says so rather than guessing. It looks only in documents you may be answered from.

If it cannot answer, it says so in plain words: AI is off for your company, your role does not include it, your company's credits for the month are used up, or a budget was reached. Nothing is charged for a question that got no answer.

Who can read what you ask

Your company's admins can read the conversations in your company, and the panel names them ("Your company's admins, Asha and Ravi, can read these conversations"). Each time an admin opens someone else's conversation it is recorded in the company's audit trail. You cannot delete a conversation: each is kept for the days your company chose (90 at first) and then erased for everybody.

SourcePro's team never reads them. SourcePro sees numbers only: how many answers a company had, what they cost in credits, and how many calls were stopped and why.

A company may choose to keep no words at all. Then each question stands alone, nothing is kept to read, and the panel says "Your company keeps no record of what you ask here".

For company admins: Settings → AI

• The switch. AI is off until you turn it on. It also needs SourcePro's own switch and a plan with AI; the page says which of them is stopping it when it is on but not answering. If your plan has no AI yet, the page says so and how to get it: ask SourcePro to add AI to your plan, then come back here to switch it on.
• Who can use it. A person's role sets the most they may do; you narrow it, never widen it. A switch sets the default: on, everyone whose role includes AI may use it, and you turn off the people you want to keep it from; off, nobody may until you turn them on (a good way to begin with a few people). A switch on a person beats the default either way, and Use default takes your choice back. A choice can end on a day (End day: a pilot buyer for a month, a trainee until they are trained): after the last day you give, the default applies to them again. The day is by India's calendar, from today up to a year ahead; empty keeps the choice until you change it. Someone you turn off sees "Your company has not given you AI" and cannot ask or open their history; you can still read what they asked before. A role that does not include AI cannot be given it. Each change is in your audit trail.
• Roles. Under "AI roles" in the rules, switch AI off for a whole role (buyers, requesters, technical reviewers, viewers): only the people of that role you turn on one by one then have it. A Company Admin is never switched off by a role, and someone with two roles has AI if either is on.
• Pages. Under "AI where and when", turn the assistant off on the pages it has questions for (home, RFQs, reports, orders, vendors, tasks) or on every other page. It is not offered there, and a question sent from there is refused.
• Quiet hours and days. The hours the assistant is quiet, by India's clock on a 24-hour clock (22 to 6 is the night), and days it is quiet all day (such as SUN). The same hour for both: never quiet. Nobody is above them, an admin included.
• Departments. Not a switch yet: a person does not belong to a department in SourcePro, only a request does. Use the roles and the people list.
• This month's credits. They come with your plan on the 1st; what is left when the month ends expires. A meter shows use, and where you are by your own marks: 70% tells you (a bell and an email), 85% shows everyone that credits are running low, 95% moves answers to a cheaper model, and 100% stops them (you may allow a little more first). Each mark is told once a month.
• Use this month. Credits by day, by person and by feature, and the calls that were stopped and why.
• Budgets. Keep any one person, department or feature from using the month's credits: per hour, day, week or month, in credits or in questions, stopping at the limit or only warning.
• Assistant documents (Settings → Assistant documents). Add the policies and handbooks you want the assistant to answer from: a PDF, Word, Excel, CSV or text file up to 25 MB. A scan or a photo has no words to find: use the text exported from the original. The file is kept privately (it is not in your company's files), checked for viruses, and then read into passages, which uses a few AI credits; the page does it by itself for a document you add, and shows Ready when the assistant can use it. Read it or Try again reads a document that was waiting or could not be read. Remove deletes the document, what was read from it and its file, for good. Two rules sit on the same page: switch looking in documents off, and set how close a passage must be to the question before it is offered (higher: fewer, surer passages). Anyone who may use the assistant can be answered from a document.
• Conversations. Read every conversation in the company, filter by person, page or date, and download the list (never the words). Opening someone else's is recorded.
• Guardrails (in the rules, under "AI guardrails").
• Your wording. Write how the assistant should speak and which terms to use ("say purchase order, not PO"). It is added after SourcePro's own rules for the assistant and can never change them.
• Topics it must not discuss. One word or phrase on each line. A question that contains one is not answered and costs nothing: the person gets your own words (you write them, or keep the default), the question is in their history, and the Use card counts it as stopped. SourcePro's own refusals (AI off, credits used up, not given to this person) still come first.
• Personal data. On at first: email addresses, phone numbers, PAN, Aadhaar, IFSC codes and bank account numbers (when called an account) are hidden in what people ask, before it reaches the model, and in what the assistant answers. What is kept in the history is what was sent. These are checked by pattern, not by the model, so a number that merely looks like one may be hidden too; PO, PR and RFQ numbers, amounts and GSTINs are not touched.
• Flag words. One word or phrase on each line (a competitor's name, a word like "bribe"). A conversation in which a question or an answer contains one is flagged: it is answered as usual, the words it matched show in your list of conversations (filter it to the flagged ones) and on the conversation, and you get a bell and an email the first time, naming who and which words, never what was said. Switch the bell and email off if you only want the list. Opening a flagged conversation is recorded like any other. The person never sees the flag. Flags need the company's words to be kept: where none are kept there is nothing to flag or to read.
• Always: no answer carries a link to another site, whatever you set.
• Rules. The assistant's name and greeting in your own words, the warning that AI can be wrong (reword it as you like; left blank, SourcePro's words are shown, so it is never left out), its longest answer, the marks above, how many days conversations are kept, and whether their words are kept at all.

More credits for the month come from SourcePro: ask them to add some.