The 7 workflows
to automate first.

Most automation advice starts with the technology. That’s backwards. The right first workflow has three properties: it happens often, it follows rules, and you can measure the result. If one of those is missing, the project becomes a science experiment.

We looked at what Indian teams are actually deploying — from WhatsApp agents sold from ₹10,000 to enterprise claims triage — and ranked the workflows by a simple test: can a mid-market team get a payback within 12 months, with a system that can be trusted?

The shortlist

Ranges are directional, based on public vendor pricing and delivery estimates. They exclude GST and third-party usage. The audit replaces them with your scope.

1. WhatsApp front desk

Before: a lead messages on WhatsApp or Instagram. One receptionist replies in four to six hours, or never. Site visits are booked by phone tag. The owner is the escalation path for everything.

Agentic workflow: webhook → intent → qualifying questions (budget, location, timeline) → grounded answers from live inventory → calendar slots → CRM write → three-touch follow-up → human handoff on negotiation → owner digest.

Why it works: the channel already exists, the buyer expects an instant reply, and the system of record is clear. Meta launched a free WhatsApp business agent in India in July 2026, and Haptik sells SMB agents from ₹10,000 — which means the generic chatbot is commoditised. The value is in the qualification, inventory grounding, and CRM write-back, not the reply itself.

Metric: first response under 60 seconds; percentage of enquiries answered; qualified leads; site visits or demos booked.

2. Speed-to-lead + CRM hygiene

Before: leads arrive from IndiaMART, the website, and WhatsApp into a shared inbox. Reps call when free. Half the deals have no next step; the CRM is updated on Sunday night.

Agentic workflow: intake → fit score → territory assignment → drafted first touch in the rep’s voice → call brief → post-call CRM update → stale-deal nudges → duplicate suggestions → weekly data-quality report.

Why it works: the bottleneck is not lead volume; it’s the time between interest and contact. The agent doesn’t need to sell. It needs to make sure every lead gets a relevant, timely first touch and every deal has a next step.

Metric: speed to first touch; percentage of open deals with a next step; field completeness; stage conversion.

3. Invoice-to-Tally / accounts payable

Before: invoices arrive by email and WhatsApp. Clerks retype them into Tally or SAP, match POs by eye, and find GST/ITC mismatches at month-end. Duplicate payments are discovered later, if at all.

Agentic workflow: ingest → extract GSTIN, HSN, tax split, PO reference → vendor/bank check → duplicate check → three-way match → GSTR-2B reconciliation → TDS suggestion → exception queue → ERP posting → vendor query drafts → weekly ITC report.

Why it works: the work is document-heavy, repetitive, and has a clear system of record. Vendors like Cygnet.One and Expenzing are shipping similar AP agents, which validates demand. The guardrail matters: the agent should never release a payment; a human approves.

Metric: touchless invoice rate; days to post; duplicates caught; ITC recovered; cost per invoice.

4. Multilingual support triage

Before: 5,000–6,000 tickets a month across email, app, and WhatsApp. Agents tag manually; 30% are misrouted; macros are English-only; QA samples 2%.

Agentic workflow: language ID (Hindi, Hinglish, Tamil, Telugu) → intent and sentiment → order lookup → L1 resolution via tools → L2 draft with knowledge-base citations → exception routing → auto-tag → 100% QA → weekly knowledge gaps.

Why it works: support is where the ROI is easiest to measure, but it’s also where over-promising is common. A realistic target is 25–40% deflection on eligible Tier-1 contacts, not 90%. The agent should answer only when it has the data and confidence; everything else goes to a human with context.

Metric: first-contact resolution; average handling time; deflection; CSAT; cost per ticket; regional resolution without a human.

5. Collections follow-up

Before: due dates pass. Tele-callers work overdue lists; most calls go unanswered. Payment promises live on paper; payment links are sent manually.

Agentic workflow: risk queue → multichannel cadence (WhatsApp, call, SMS) → intent (will pay, dispute, hardship) → UPI link in the conversation → settlement confirmation → promise-to-pay logged → human escalation → daily QA dashboard.

Why it works: the economics are direct — recovered cash, lower cost per rupee recovered, fewer broken promises. But conduct risk is real. Hours, consent, recording, dispute freezes, and human escalation need to be designed in from day one, especially for RBI-regulated lenders.

Metric: resolution rate; roll rate; cost per rupee recovered; promises kept; complaints.

6. MIS pack + variance commentary

Before: month-end means twelve spreadsheets, a four-day close, and an analyst pasting numbers into a deck and writing commentary from memory.

Agentic workflow: pull ERP/CRM/billing → reconcile definitions → compute variances in code → narrative agent explains the drivers with drill-downs → anomaly detection → draft board commentary labelled known/estimated → human review → scheduled distribution.

Why it works: the model never does arithmetic. Numbers are computed in code; the agent writes the narrative and flags anomalies. That separation is what makes it trustworthy. The hard part is agreeing on metric definitions — which is exactly what the audit surfaces.

Metric: close days; rework; time to board pack; error rate.

7. Resume screening + scheduling

Before: 1,200 applications for 40 roles. Two recruiters screen for five hours a day. Keyword filters miss adjacent skills; 30% of interviews are no-shows.

Agentic workflow: parse resume/JD → fit score with quoted evidence → knockout checks → ranked shortlist with reasons → calendar negotiation → reminders and reschedules → recruiter summary; rejection reasons logged.

Why it works: the admin burden is enormous, and the decision can stay with a human. The guardrail is non-negotiable: no auto-rejection, bias testing by gender/college/city, explainable scoring, and DPDP-compliant consent and retention.

Metric: time to shortlist; recruiter hours per hire; no-show rate; offer acceptance; adverse-impact ratio.

Where not to start

• Meeting notes → CRM. Native Zoom/Teams features already do this; willingness to pay is low.
• Deterministic routing. If a rule covers 80% of cases, use the rule. An agent adds cost and unpredictability.
• High-stakes decisions with no human. Credit approval, medical advice, legal conclusions. The agent can prepare the file; a named human decides.
• Low-volume, one-off work. The build will cost more than the hours it saves.
• A process nobody can describe. Fix the process first. Automation makes a bad process faster, not better.

How to choose

Score each candidate on five dimensions, 1–5: frequency, rule-basedness, data availability, error cost, and measurability. Pick the highest total that has a named owner. If two workflows tie, choose the one with the cheaper mistake — it will be easier to get approval and easier to learn from.

Then run a one-week audit. The output should be a baseline, a workflow map, a feasibility check, and a payback estimate. If the numbers don’t clear 12 months, don’t build. That’s not a failure; it’s the most valuable thing an audit can tell you.