AI Agents for E-commerce & D2C Brands in India

FILE 01 / THE E-COMMERCE OPS PROBLEM

Your ops team is doing work that a well-scoped agent could finish in seconds.

Someone is opening Shiprocket, checking a courier delay, copy-pasting the tracking link into WhatsApp, then updating the ticket. Someone else is comparing prices on Flipkart against margin rules in a Google Sheet. A third person is reviewing risky orders one by one before dispatch. Every task looks small until you count the hours.

Chatbots reduced reply times. They did not remove the work. Agents remove the work. An agent checks the courier, drafts the refund, triggers the Razorpay reversal, updates Shopify, posts to Slack. Human approves the exceptions.

The economics are simple. If a task costs your team 3 minutes and happens 5000 times a month, that's 250 hours. Roughly two full-time ops staff. An agent doing the same work runs at rupees per hour of compute. The math starts favouring agents at about 500 events a month per workflow.

FAILURE LOG What breaks in D2C ops without agents

✗ Refund backlog. 3-5 day SLA slips to 10 days at scale
✗ Cart recovery windows missed because no one sent the WhatsApp in time
✗ Marketplace prices drift and buy-box is lost overnight
✗ Fraud orders ship, chargebacks arrive 45 days later
✗ Warehouse hits zero on hero SKU because no one flagged the trend
✗ Support tickets pile up because L1 has to look up 5 systems per query
✗ Upsell campaigns never launch because marketing bandwidth is elsewhere

New to agents? How they plan, call tools, and stay inside guardrails is covered on the AI Agents & Automation service page

FILE 02 / OUR APPROACH

How we build an agent that actually ships

Not a demo. A running system that touches your store, your logistics, and your payment gateway.

01

Map the workflow

Sit with your ops lead. Walk through the exact steps a human takes today. Identify decision points, escalation triggers, and where policy is fuzzy.

02

Wire the tools

Shopify Admin API, Shiprocket, WhatsApp Business, Razorpay. Each becomes a tool the agent can call. Every call is logged, rate-limited, and reversible.

03

Set the guardrails

Refund cap without human approval. Discount ceiling. Fraud score threshold for auto-hold. Written policy that becomes the agent's system prompt and hard-coded checks.

04

Shadow, then ship

Agent runs in shadow mode for 2 weeks. Drafts actions, logs decisions, human approves. Once accuracy hits 95%+ on the sample, we flip auto-execute for safe cases.

FILE 03 / AGENT PATTERNS WE SHIP

Eight agents that pay for themselves

Same agentic backbone underneath. Different tools, guardrails, and escalation rules per workflow.

AGT-01

Order Status & Tracking Agent

Customer pings on WhatsApp or email. Agent pulls order ID, hits Shiprocket or Delhivery, checks delay against SLA, drafts response with tracking link. Escalates only if courier is stuck 48+ hours.

70% SELF-SERVE RATE

AGT-02

Refund & Return Agent

Verifies policy eligibility (window, category, condition rules), initiates Shiprocket reverse pickup, updates Shopify order, triggers Razorpay refund on receipt confirmation, closes ticket. Escalates high-value or repeat cases.

3-DAY SLA HELD AT SCALE

AGT-03

Cart Abandonment Recovery Agent

Watches cart events, segments buyer (new vs repeat, cart value, price sensitivity from past orders), sends personalized WhatsApp within 30 minutes, decides discount inside your margin rules. Not a template blast.

15-25% RECOVERY LIFT

AGT-04

Inventory Replenishment Agent

Reads stock levels across warehouses, factors in sell-through velocity and lead time, drafts POs for suppliers. Auto-sends for pre-approved SKUs under threshold. Alerts humans for anything above your comfort cap.

STOCKOUTS DOWN 40%

AGT-05

Post-purchase Upsell Agent

Segments buyers by first purchase, browses catalogue for logical add-ons, drafts WhatsApp or email campaigns with copy, offer, and target list. Marketing lead approves in Slack, agent schedules the send.

2-4% INCREMENTAL AOV

AGT-06

Marketplace Repricing Agent

Scans competitor prices on Amazon and Flipkart every 30 minutes, checks your margin floor and buy-box status, adjusts within rules. Logs every change. Alerts if competitor drops below your cost.

BUY-BOX HOLD UP 25%

AGT-07

Fraud Triage Agent

Scores incoming orders on velocity (same card, multiple accounts), geo mismatch (billing vs delivery vs IP), device fingerprint, and payment signals. Auto-holds risky orders before dispatch. Human reviews the flagged pile.

CHARGEBACKS CUT 30-50%

AGT-08

Support Triage Agent

Reads inbound ticket, classifies intent, checks order and customer LTV, drafts L1 response, routes complex or high-value queries to human agents with full context. No more agents starting from scratch on every ticket.

FIRST REPLY UNDER 2 MIN

AGT-09

Multi-agent Orchestration

These agents talk to each other. Refund agent tells fraud agent about repeat return-fraud accounts. Cart agent asks inventory agent before offering a discount on low-stock SKUs. One graph, shared memory.

BUILT ON LANGGRAPH

Which agent should you build first?

Ask Deci or reach us via the contact form. Tell us your platform, monthly order volume, and the top-3 tasks your ops team hates. We come back with a ranked list and a pilot scope with fixed price.

Chat with Deci

FILE 04 / TECH STACK

What we build on

Boring choices, mostly. We pick tools that will still exist in three years.

STACK-01 Agent Frameworks

LangGraph (multi-step, stateful) · CrewAI (role-based) · Vercel AI SDK · Custom Python orchestration for tight-latency cases

STACK-02 Models

Anthropic Claude (Sonnet, Haiku) · OpenAI GPT-4o and mini · Bedrock · Self-hosted Llama and Mistral for DPDP-sensitive flows

STACK-03 Storefront APIs

Shopify Admin & Storefront · WooCommerce REST · Magento 2 · Amazon SP-API · Flipkart Seller API · Custom Laravel/Rails

STACK-04 Messaging

WhatsApp Business API · BookMySMS (our own bulk-comms) · Gupshup · Interakt · Twilio · MSG91 · Postmark for email

STACK-05 Logistics

Shiprocket · Delhivery · Blue Dart · Ecom Express · Xpressbees · DTDC · India Post webhooks

STACK-06 Payments

Razorpay · PayU · Cashfree · CCAvenue · Stripe for international · UPI intent flows · Refund automation via webhooks

STACK-07 Data & Memory

Postgres · Redis for agent state · pgvector or Qdrant for retrieval · ClickHouse for event logs · S3-compatible for artifacts

STACK-08 Ops & Eval

Langfuse · LangSmith · Custom eval harness · Sentry · Prometheus + Grafana · Slack for human-in-loop approvals

FILE 05 / WHY US

We built our own e-commerce comms rails. We know where agents break.

BookMySMS is our own product. Bulk WhatsApp, SMS, and email for Indian businesses. It runs template approvals, delivery reports, opt-out handling, and retry logic at scale. Every quirk of the Meta and Indian telecom stack, we hit first.

That matters because most agent projects fail on integration, not on model quality. The refund logic is fine, the WhatsApp template gets rejected. The repricing math works, the marketplace API rate-limits at 3pm. We ship agents that survive contact with production because we run production ourselves.

Beyond BookMySMS, our team has shipped AI systems for Indian legal tech (VakeelSaathi RAG on 50k+ documents), compliance (dcomply document review), and property ops (Realzent). The pattern is the same. Boring reliable integrations, tight guardrails, monitored evals, human-in-loop where mistakes are expensive.

Visit BookMySMS

ON THE RECORD What we defend on customer calls

30-50%

Ops team hours saved within 90 days on the workflows we automate

15-25%

Cart recovery lift over template-only WhatsApp blasts

3 days

Held refund SLA at 2000+ returns/month for one D2C client

₹0.30-1.20

Blended cost per completed agent run (models + infra)

Numbers vary by catalogue, region, and integration depth. We commit to targets after scoping via the contact form, not before.

FILE 06 / ENGAGEMENT MODEL

Pilot, build, operate. Fixed pricing per phase.

WEEK 1

Discovery

Workflow map, API audit, ranked list of which agents pay back fastest for your business.

FREE

WEEKS 2-4

Pilot

One agent, single workflow. Shadow mode for 2 weeks, then auto-execute on safe cases. Written eval report.

₹1.5-3 L

WEEKS 5-10

Production Build

3-6 agents across order, refund, cart, fraud, support. Multi-agent orchestration, guardrails, human-in-loop, dashboards.

₹5-15 L

ONGOING

Managed Ops

Model API costs, eval runs, policy updates, incident response, new agent additions. We stay on-call.

FROM ₹40K/MO

REF-A Parent & Sibling Capability

REF-B Agents by Industry

REF-C Related Work

FILE 07 / FIELD MANUAL — FAQ

Questions ops heads ask on every call

What is the difference between an AI chatbot and an AI agent for e-commerce?

A chatbot answers questions. An agent takes action. A chatbot tells the customer their order status. An agent checks the courier API, confirms the delay, drafts a refund per your policy, initiates the Shiprocket return pickup, and posts to your ops Slack. Chatbots reduce ticket volume. Agents reduce headcount hours.

Should I build or buy an AI agent stack?

Off-the-shelf tools like Bland, Yellow.ai or Zapier AI cover 60-70% of standard workflows. Build custom when your policies are non-obvious, your margins are tight enough that a wrong discount hurts, or you integrate with Indian logistics and payment APIs those tools do not support well. Most brands run hybrid. Buy for standard flows, build the 3-4 agents that are core to your economics.

How long before I see ROI on an AI agent?

A single agent (say refund automation) pays back in 60-90 days if you handle 500+ refunds a month. Cart recovery agents pay back in 30-45 days once WhatsApp templates are approved. Full multi-agent ops stack takes 4-6 months to show clean ROI numbers, mostly because you need clean baseline data first.

Is customer data safe? What about DPDP compliance?

We build with DPDP Act 2023 in mind. Customer PII stays in your database. We use provider APIs (OpenAI, Anthropic, Bedrock) that support zero-retention endpoints. For sensitive flows like refund verification, we route through Indian-hosted models or self-hosted Llama variants. Full data flow doc gets shared before build starts. No customer data leaves your VPC unless you explicitly approve the route.

How much engineering effort do we need on our side?

For a Shopify or WooCommerce store, minimal. We work with API keys and existing webhooks. For custom stacks (Laravel, Rails, Node), we need 4-6 hours of your dev team's time per week during the build phase to review API contracts and staging tests. Post-launch, near zero effort unless your business logic changes.

How much does an AI agent build cost?

Pilot phase (one agent, single workflow) is ₹1.5-3 lakh, 3-4 weeks. Multi-agent production build across order, refund, cart, and support routing is ₹5-15 lakh, 6-10 weeks. Managed ops from ₹40,000/month, includes model API cost pass-through, eval runs, and policy updates.

Which model should we use, GPT, Claude, or open source?

Depends on the agent. Refund and fraud triage agents need reliable structured output, we default to Claude Sonnet or GPT-4o. Cart recovery messaging works well on cheaper models like Haiku or Mistral. Order status agents rarely need frontier models. We route each step to the cheapest model that meets the quality bar. Typical blended cost lands at ₹0.30-1.20 per completed agent run.

When should we stick with a chatbot instead of an agent?

If your main problem is repeated questions (sizing, restock, policy), a chatbot solves it faster and cheaper. Move to agents when the work is not answering but doing. Processing refunds, updating prices, chasing carts, flagging fraud. Chatbots reduce reply time. Agents remove the task from the human queue entirely.

Ready to give your ops team their evenings back?

Have questions? Ask Deci (our AI assistant, bottom-right) or drop us a line via the contact form. Tell us your platform, order volume, and the tasks eating your team's hours. We come back with a pilot scope and fixed price within 48 hours.

Chat with Deci Send Requirements →