From recommendation to action — with per-role gates on every autonomous step
WhatsApp debtor chase, Helpdesk ticket triage, and CRM deal coaching that respect your access model. Every agent is off by default and lit up only after you approve the confidence threshold and target scope.
Autonomous ops agents are the acting layer of the AI Copilot — the components that turn a decision the copilot recommended into a real business action. Three agents ship in early access: a WhatsApp debtor-chase agent that sends language-matched reminders on your aging schedule and logs replies back into the ERP; an ERPNext Helpdesk triage agent that classifies incoming HD Tickets and drafts or auto-sends the first response above a per-DocType confidence threshold; a Zoho CRM deal-stage coaching agent that posts next-best-action notes on deals stuck too long against your own historical won-deal patterns. Every agent is off by default and enabled per user, per DocType, and per confidence threshold.
Last updated 2026-07-02 · Part of the AI Copilot for ERPs early access
What this capability does
The gap between knowing an invoice is overdue and actually sending the reminder is where working capital leaks. The gap between a Helpdesk ticket arriving and an agent responding is where NPS drops. The gap between a CRM deal stalling in Proposal and someone noticing is where pipeline evaporates. Autonomous ops agents close those gaps. Three agents ship in the AI Copilot early access. The WhatsApp debtor-chase agent reads your aged AR from TallyPrime, Zoho Books, ERPNext, or a custom ERP; picks the correct WhatsApp template per party (language auto-selected from state code — Hindi for the Hindi belt, Hinglish for metros, English otherwise); sends reminders on the aging schedule you configured; and logs replies back into the ERP as voucher narrations, Note attachments, or Communication records. The ERPNext Helpdesk triage agent classifies every incoming HD Ticket by DocType and topic against your knowledge base, produces a confidence-scored first response, and either sends it or drafts it depending on the threshold you set per DocType. The Zoho CRM deal-stage coaching agent watches for deals dwelling in a stage longer than your own historical baseline for won deals in that segment and posts a next-best-action Note on the deal or, if configured, fires a WhatsApp reminder to the assigned rep. The unifying discipline: every agent is off by default. Turning one on requires an explicit decision — which users can trigger it, which DocTypes it applies to, at what confidence threshold. Every autonomous action writes an audit record to your storage naming the user, the trigger, the confidence, and the action taken.
How it works
End to end, from question to grounded action.
Configure per agent, per scope
Each agent is configured in the ERPPlugs dashboard. You choose the target scope (which companies, which DocTypes, which user segments), the confidence threshold (for the triage agent), the aging schedule (for the debtor-chase agent), and the language matrix (for the coaching agent). Nothing runs until you explicitly enable it.
Shadow mode first
For the first 1–2 weeks after enabling, the agent runs in shadow mode — actions are queued and visible for your review but not sent. This is your calibration window. Pilot customers routinely find that the initial confidence threshold needs adjustment for their specific customer language before going live.
Live actions with audit
Once shadow mode ends, the agent takes real actions. Every WhatsApp message sent, every HD Ticket first response, every CRM Note posted is logged with the user identity, the triggering ERP record, the confidence score at the time of action, and the outcome (delivered, replied, resolved).
Kill-switch and per-instance opt-out
A dashboard-level kill-switch pauses all autonomous actions immediately. Individual customers can be excluded from the debtor-chase agent (a strategic account you would rather chase personally). Individual HD Ticket DocTypes can be moved back to draft-only after go-live if the confidence baseline drifts.
How this shows up on each ERP
Same capability, tuned to each ERP's data model.
Debtor-chase agent reads aged AR through the on-prem bridge and dispatches through your existing Tally-WhatsApp connector. Replies get logged as voucher narrations against the corresponding Sales Invoice — visible to the accountant inside TallyPrime.
The strongest surface for autonomous ops because all three agents run natively — debtor chase reads Books AR, CRM deal coaching reads Zoho CRM, and if you use Zoho Desk the triage layer applies there too. Replies log as Zoho Notes on the corresponding record.
Helpdesk triage is the flagship agent for ERPNext customers — classifies HD Tickets against your Wiki and past resolved tickets, drafts a first response, and auto-sends above the per-DocType confidence threshold. Every action writes a Communication record against the acted-on Doc.
Debtor-chase runs against your mapped AR ingestion. Reply logging depends on your custom ERP's supported write-back mode (REST callback preferred; CSV export otherwise). Coaching and triage agents scope to whichever CRM or Helpdesk you have wired through Zapier / n8n / Make.
What this looks like in practice
Real question-and-answer patterns from pilot deployments.
“Aging bucket: 45–60 days. Debtor: Rajarshi Industries Ltd. Outstanding: ₹1.28 Cr across 3 invoices.”
Sent Hindi WhatsApp reminder from your business number to the primary contact (auto-selected from party master) at 10:30 AM on the aging schedule. Reply received at 11:14 AM: "Payment processing next Tuesday, will share UTR." Logged as voucher narration against TSA/2026/1108. Debtor moved to "responded — payment committed" status.
“HD Ticket #HDT-2026-4472 arrived: "Cannot generate GSTR-1 JSON export — button greyed out."”
Classified: DocType "Sales Invoice", topic "GST export". Confidence 0.91 (above the 0.85 threshold for GST tickets). First response auto-sent: "This usually means the Fiscal Year for the export is not set. In your ERPNext, go to Accounts → GST Settings → verify the Fiscal Year field is set to 2026-27 and the reload the report." Ticket status: "responded — awaiting user reply."
“Deal "Rajarshi Industries — 250 units" at Proposal stage for 34 days; historical median at Proposal for Enterprise deals is 18 days.”
Note posted on the deal: "Deal dwelling in Proposal 88% longer than historical Enterprise median. In your last 6 won Enterprise deals at this stage, the successful next action was: a follow-up call with the CFO within 5 days of proposal share. Consider scheduling one." Rep notified via Zoho notification.
Why this matters
Off by default
Every agent disabled until you explicitly enable it.
Shadow mode first
Actions queued for your review before going live.
Per-DocType confidence gates
Different thresholds for different record types.
Kill-switch always available
Dashboard-level pause halts all agents instantly.
Frequently asked questions
What stops an autonomous agent from doing something embarrassing?+
Four layers of control. First, every agent is off by default — turning it on is an explicit act. Second, shadow mode runs for 1–2 weeks so you calibrate before live actions. Third, per-DocType confidence thresholds mean high-risk records (a large customer's Sales Invoice, a legal dispute HD Ticket) can require human review at any confidence, while low-risk records (a routine password-reset ticket) auto-respond. Fourth, the dashboard kill-switch pauses all agents immediately if something drifts. Pilot customers have used all four in practice — most commonly the per-DocType threshold to hold strategic accounts back from auto-chasing.
How does the debtor-chase agent pick the language?+
Same language-matrix logic used by Tally-WhatsApp, Busy-WhatsApp and Marg-WhatsApp connectors — buyer's state code drives default language (Hindi for the Hindi belt states, Hinglish for metros, English elsewhere), overridable per party. Voice notes are transcribed and answered in the same language the buyer spoke. This matters because the accepted collection cadence and phrasing differs materially between a Delhi trader and a Chennai enterprise buyer.
Can I configure the debtor-chase agent to skip specific customers?+
Yes. Any party can be excluded through the ERPPlugs dashboard — the classic case is a strategic account you would rather chase personally, or a customer in an active dispute where an automated reminder would be counter-productive. Excluded parties still surface in the AR view but do not receive automated messages. Exclusion can be per-agent (skip debtor chase but include in cash-flow forecast) or global.
How does the Helpdesk triage agent decide auto-send vs draft?+
A per-DocType confidence threshold. Above the threshold, the first response is auto-sent and the ticket status moves to "responded." Below the threshold, the response is queued as a draft for an agent to review, edit if needed, and send manually. Confidence is a function of the classification match strength against your knowledge base plus the length of the input ticket (shorter tickets are easier to misclassify and score lower).
Does the deal coaching agent send reminders directly to my reps?+
Only if you configure it that way. Default behaviour is to post a Note on the Zoho CRM deal and rely on the rep's normal CRM notification flow. Optionally, a WhatsApp reminder can be dispatched to the assigned rep for deals above a value threshold or at a specific dwell time. Both channels are configurable per team.
What is the audit trail for autonomous actions?+
Every action writes an entry to your audit storage (S3 or on-prem) with the user identity that authorised the agent, the triggering ERP record, the confidence score at the time of action, the action taken, and the outcome (delivered / replied / responded / resolved). The same information is written into the ERP itself as a voucher narration (Tally), a Note (Zoho), or a Communication record (ERPNext) — so a statutory auditor can review the trail without leaving the ERP.
Related AI Copilot capabilities
Every capability composes with the others.
Available today on these ERPs
Each edition uses the same connectivity ERPPlugs already ships.
Want this capability running on your ERP?
Join the AI Copilot early-access pilot. We calibrate against your own data during the first 2–4 weeks and lock in early-adopter terms before general availability.
Join Early Access