AI Assistant, Agents & AutomationModule 2210 apps
Ask the business a question — and let the routine work run itself At launch
Last for the same reason Dashboard & BI is late: something that answers questions about the whole business can only be built once the whole business is in one place. Three different things live here and the difference between them matters. An ASSISTANT answers a question you asked, from the records, with the records attached. A CHATBOT holds the same conversation with your customer instead of you. An AGENT is given a job rather than a question and works out the steps itself. That last one is what separates this module from Automation Studio in Module 20, where a person draws the steps in advance and the rule runs the same way every time; and from Automation in Module 17, which fires marketing campaigns and nothing else. Both of those stay exactly as they are — this module sits above them and calls them, rather than replacing either. Module 18 writes content; this module answers and acts.
Writes toProjects & CollaborationCRMMarketing
Part of Marketing & AI
10 apps in AI Assistant, Agents & Automation
AI Assistant
Ask in your own words — “what did Myntra actually pay us last week, and what is still short?” — and get the answer with the rows it came from sitting underneath it, each one clicking through to the record. It reads the ledger, the stock table and the settlement lines the same way a report does, so the figure it gives is the figure the books give. When it cannot find the answer it says so and shows what it looked at; it never estimates a number and presents it as a fact, because a plausible wrong figure is far more expensive than an honest blank. It answers only from records the person asking is already allowed to open, so it can never become a way around permissions.
AI Chatbot
The same engine turned to face the customer, on your own storefront and on WhatsApp: where is my order, will this size fit me, I want to return this. It reads the real order and the real size chart rather than a script written six months ago, and it will say “let me get someone” instead of guessing at anything about money, a refund or a complaint. The handover goes into the Module 04 Helpdesk queue with the whole conversation already attached, so the person picking it up starts where the customer left off instead of asking them to explain again. It never asks a customer for a card number, a bank detail or a password — that promise does not get a chatbot-shaped exception.
AI Agents
A job rather than a question: “chase every unreconciled settlement line from last week and draft the claim for each.” The agent works out the steps, does them, and stops at the point where a person has to decide. It runs inside a scope you set — which records it may read, which it may write, and how much it may spend through the Module 01 Provider Router — and it cannot quietly widen that scope mid-run. Anything that moves money, files a claim, changes a price or sends a customer a message waits for a human yes.
Agent Guardrails & Run Log
What each agent is allowed to touch, written down as a scope rather than trusted to a prompt, and every run recorded step by step: what started it, what it read, what it proposed, what a person approved, what it actually changed. Kept in the same audit trail as everything else, with the same absence of an off switch. An agent whose working nobody can inspect afterwards is not a colleague, it is an unexplained entry in the books.
Knowledge & Retrieval
The index that makes the answers grounded: your own designs, rate cards, settlement files, standard procedures and past decisions, searchable so a reply quotes what is actually on file instead of what a model remembers about the trade in general. Permission-scoped at the row, so two people asking the same question get answers drawn only from what each of them may already see.
Voice In and Out
The same assistant asked out loud and answering out loud, because the people who most need a figure are often holding something in both hands. A question spoken in the language the speaker actually uses becomes the same query the typed assistant would have run, and the answer comes back as speech with the rows still attached for anyone who wants to look at them. It refuses exactly what the typed one refuses: it will not read out a figure it could not find, and it will not take a password, a card number or a bank detail by voice — a spoken channel does not get an exception to a promise the rest of the system keeps.
Corrections & What It Has Learned
When the assistant gets something wrong, the correction is a record rather than a conversation that scrolls away. Somebody says what the right answer was, that sits against the question it belongs to, and the next person asking gets the corrected version with a note of who corrected it and when. The whole list is readable and editable by the business — not a hidden memory quietly steering answers — so a correction made in haste can be found and undone. Nothing here changes the model; it changes what the model is shown, which is the only kind of learning anybody can audit afterwards.
Agent Studio
Build an agent by writing what it should do in plain words: what it may read, what it may do, and what must wait for a person. It is tested on the business’s own records before it is switched on.
Agent Store
Ready-made agents for common work in each kind of business, installed in one step and changed like any agent the business built itself. Each one lists exactly what it reads and does before it is installed.
AI Workspace
A place to work with the AI over time: projects, the files put into them, the documents it writes, and what it should remember about the business. Everything in it stays inside the business’s own sealed data.