AI Command Center: AI agents that draft the work, and you approve it
Most AI in business software is a chatbot bolted onto a dashboard. Dravyaa's AI Command Center does three things instead: it tells you what needs attention, answers questions you type or speak, and prepares the next document for you to approve. Here is how it works and where it stops.
In short
- The AI Command Center is part of Dravyaa, a cloud manufacturing ERP for Indian SMEs. It has exactly three parts: Needs attention, Ask and Approve.
- Ask by typing or by voice, in English, Hindi or Hinglish, on the web and in the Dravyaa mobile app (hold the mic and speak).
- Five specialist AI agents answer from live ERP data: Manufacturing Planning, Inventory, Procurement, Sales and Orders, and Business Performance.
- Agents never write to your books. They queue a purchase order, production release, finished-goods receipt, sales invoice, goods receipt or vendor bill, and nothing happens until someone with the right permission approves it.
- Shortage and planning numbers come from a deterministic planning engine (BOM explosion and due-date allocation). The AI explains those numbers; it does not calculate them.
Dravyaa's AI Command Center is an AI agent workspace inside Dravyaa ERP that reads a manufacturer's live orders, stock, production and accounts, shows what needs attention today, answers questions typed or spoken in English, Hindi or Hinglish, and drafts the next document (a purchase order, invoice, goods receipt or bill) for a person to approve. It runs on the web and in the Dravyaa mobile app, uses Google Gemini models, and is built so that the AI proposes and a person decides.
It went live in November 2026. This page explains what it does, how it differs from an ERP chatbot or a dashboard, and the guardrails around it.
What is the AI Command Center?
It is one screen with three parts, and deliberately nothing else: no KPI tiles, no charts, no funnels. Those stay on Dravyaa's Finance and Operations dashboards. The Command Center is for acting, not for watching.
| Part | What it does | Example |
|---|---|---|
| Needs attention | A short list of cards built from live data, each pointing at something to do | “Orders at risk”, “Reorder SC-14”, “Ready to invoice”, “E-way bills lapsing in 48 hours” |
| Ask | A chat you can type into or speak to; a router sends each question to the right specialist agent | “Aaj kya banana hai?” or “Which customer owes us the most?” |
| Approve | A queue of documents the AI has prepared, each waiting for a person to approve or reject | Approve a purchase order for 1.2 t of steel coil from the last vendor at the last price |
How is it different from an ERP chatbot or a dashboard?
| ERP chatbot | Dashboard | AI Command Center | |
|---|---|---|---|
| Tells you what to look at | No, you have to ask | Shows numbers, you interpret | Yes, as attention cards |
| Answers a new question | Yes | Only if a chart already exists | Yes, by text or voice |
| Prepares the next document | Rarely | No | Yes, into an approval queue |
| Can change your books by itself | Sometimes | No | No, a person approves every write |
| Hindi and Hinglish voice | Rarely | No | Yes, on web and mobile |
What shows up under Needs attention?
Each card comes from rules over live ERP data, not from the AI guessing. The current cards are:
- Orders at risk: open sales orders that will slip because material is short.
- Reorder: raw materials short against open demand, after counting purchase orders already on the way.
- Ready to start and Waiting for material: production orders that can be released now, and those that cannot yet.
- Ready to invoice: sales orders shipped but not yet billed.
- Goods due from vendors and Vendor bills to record: purchase orders whose goods should have arrived, and goods received that have no bill yet.
- Finished goods: completed production waiting to be received into stock.
- Compliance: e-way bills lapsing within 48 hours and GST e-invoice or e-way bill pushes that failed in the last seven days.
Who are the five AI agents?
When you ask a question, a router picks one specialist. Each agent has its own instructions and its own set of read-only tools, so the procurement agent cannot wander into payroll and the sales agent does not plan production.
| Agent | What it answers | What it can queue for approval |
|---|---|---|
| Manufacturing Planning | What to make today and in what order, by due date and material readiness; what to hold and why | Production plan, release of a ready production order, finished-goods receipt |
| Inventory Intelligence | Stock in hand, what is running low against open demand, what is short for upcoming orders | Goods receipt against a purchase order, finished-goods receipt |
| Procurement Intelligence | What to buy, how much and from whom, using the last vendor and last price per item | Purchase order, goods receipt (GRN), vendor bill from a GRN |
| Sales & Order Intelligence | Open orders, what is due, what is at risk and why, dispatch status, who owes money | Sales invoice for an order that is ready to bill |
| Business Performance | Sales, purchases, margins, receivables and payables ageing, dispatch performance, inventory value | Nothing; it only answers |
Can I talk to it in Hindi or Hinglish?
Yes. Press the mic on the web or hold to talk in the Dravyaa mobile app and speak the way you would to your production manager. You choose the reply language: English, Hindi in Devanagari script, or Hinglish (Hindi written in Latin script, the way most owners text). Business words such as order, stock and PO stay in English because that is how people say them on the shop floor.
An illustrative exchange on a steel fabricator's demo data:
You (voice): Aaj kya banana hai?
Command Center: Aaj WO-1043 pehle banao, uska material ready hai aur order Friday ko due hai. WO-1044 hold karo: SC-14 steel coil 1.2 t short hai. Maine 1.2 t ka PO last vendor ke rate pe approval ke liye rakh diya hai.
The 1.2 t shortfall in that reply is computed by the planning engine from the bills of materials and open orders. The model's job is to say it clearly in the language you asked in.
Is it available on mobile?
Yes. The Dravyaa mobile app has the same three parts as the web, with a hold-to-talk mic button for voice questions, and approvals you can action from the phone. A purchase order the AI queued in the morning can be approved from the factory floor. The Android app is coming soon on Google Play, with iOS to follow; on the web the Command Center is available now.
What does the AI actually do when I approve?
Approval runs the same code path as the form your team already uses. An approved invoice is created by the normal invoice service with its usual numbering and GST split from the item master; an approved GRN posts stock exactly as a manual GRN would. Three checks happen at the moment of approval:
- Permission: the approver must hold the permission for that document, for example purchase order create or invoice create. A staff user without it gets a refusal, even if the AI proposed the action.
- Still open: quantities are capped at what is still open when you approve, not when the AI proposed. If someone received half the goods in between, the GRN is for the other half.
- Once only: a second approval of the same action is refused, so a double tap cannot create two purchase orders.
A Find actions sweep also scans for routine work (shipped but not invoiced, goods overdue from vendors, received but not billed, finished goods to receive, drafts ready to release) and queues it, skipping anything already approved or rejected.
What guardrails stop the AI from making mistakes in my books?
| Risk | Guardrail |
|---|---|
| The AI writes a wrong entry | Agents have read-only tools. Every write goes through the approval queue and a person's permission. |
| The AI calculates a wrong shortage | Material planning (BOM explosion, earliest-due allocation, POs arriving in time) is deterministic code. The model explains the result and never computes it. |
| One company sees another's data | Every lookup is scoped to the signed-in company; another company's action id simply does not exist for you. |
| Instructions hidden in data | Your question is fenced off from system instructions, and connected Google Sheets are passed to the model as data, never as instructions. |
| Runaway loops or cost | Each answer stops after six tool steps, and there is a per-user limit on questions per minute and per day. Every AI call is metered. |
| Nobody can tell what happened | Every tool call the AI makes, including refused ones, is logged. |
Does it send a daily summary?
Yes, a morning brief at 9:00 IST lists what is waiting for each approver. It arrives in the notification bell inside Dravyaa on web and mobile, not on WhatsApp, Telegram or email, and only when notifications are switched on for your company.
Can it read data that lives outside the ERP?
Yes, read-only, from Google Sheets. Many factories keep a machine plan or a dispatch tracker in a sheet. You can connect a sheet through a small Apps Script that Dravyaa gives you, or by pasting "Publish to web" CSV links, up to 20 tabs per sheet. The agents can read those tabs alongside the ERP data; they cannot write to them.
Who is it for?
Owners and managers of small and mid-sized Indian manufacturers who already run, or are moving to, one system for stock, purchase, production, sales and GST, and who do not have time to read reports every morning. It is most useful where the daily question is “what should we make, what is short, and what do I need to sign off?”
Frequently asked questions
What is Dravyaa's AI Command Center?
It is an AI agent workspace inside Dravyaa ERP with three parts: Needs attention (cards for what to act on), Ask (questions by text or voice) and Approve (documents the AI has prepared for a person to approve). It reads live orders, stock, production and accounts.
Can the Dravyaa AI create purchase orders or invoices on its own?
No. The agents only read data and propose. A purchase order, production release, finished-goods receipt, sales invoice, goods receipt or vendor bill is created only when a user with the matching permission approves it, and then through the same service the normal form uses.
Does the Dravyaa AI understand Hindi and Hinglish voice commands?
Yes. You can speak or type in English, Hindi or Hinglish, and choose whether replies come in English, Hindi (Devanagari) or Hinglish (Hindi in Latin script). Voice works on the web and in the Dravyaa mobile app.
Does the AI voice command work on mobile?
Yes. The Dravyaa mobile app has the same Command Center with a hold-to-talk mic, attention cards and the approval queue. The Android app is coming soon on Google Play; the web version is available now.
Which AI model does Dravyaa use?
The Command Center runs on Google Gemini models, with automatic fail-over between models if one is busy. Planning numbers such as material shortages come from Dravyaa's own deterministic planning engine, not from the model.
Is the AI Command Center a dashboard?
No. It deliberately has no KPI tiles or charts. Figures and trends stay on Dravyaa's Finance and Operations dashboards; the Command Center is only for what needs attention, questions, and approvals.
Ask your factory what to make today
See the AI Command Center answer in Hinglish and queue a purchase order for approval, on your own data, in a 20-minute demo.