How to implement agentic commerce workflows: a step-by-step guide with Gwiksoft

TL;DR

An agentic commerce workflow is a job an AI agent finishes end to end — finding the right product, booking a store visit, changing an order — instead of only answering a question. Getting one live safely comes down to five things:

• Start with one workflow that is frequent, rule-based and easy to measure, not with “an AI for everything”.
• Give the agent tools, not freedom: each action is a narrow tool with clear inputs, exposed over MCP or an API.
• Decide who approves what. Reading is automatic, small writes need rules, money needs a person.
• Check every reply before the customer sees it — prices against the catalogue, offers against your list — and hand off to a person when unsure.
• Roll out in stages and measure daily: shadow, then assisted, then autonomous.

What is an agentic commerce workflow?

An agentic commerce workflow is a multi-step business process that an AI agent carries out on its own, using real data and controlled actions, from the customer's first message to a finished outcome. The difference from a chatbot is the last step. A chatbot tells the customer how to change their address. An agent checks the order is not yet shipped, updates the address in the store, and confirms it — within the rules you set.

Which workflows to automate first

Start where volume is high and a mistake is cheap, then move towards workflows that touch money. Most Indian D2C brands and retailers we work with get the fastest return from the first three rows.

How to implement an agentic workflow, step by step

Every workflow we build follows the same eight steps, whether it is visit booking or refunds.

1. 1Pick one workflow and write its success metric in one line — for example, “store-visit requests per 100 enquiries”.
2. 2Map the data and systems it needs: Shopify or your product list, order data, WhatsApp, the CRM where leads go.
3. 3Expose each action as a narrow tool with typed inputs — “create_visit_request(date, time, product)”, not “update database”. We use the Model Context Protocol (MCP) or a plain API.
4. 4Set permissions per tool: what the agent may read, what it may write, and which actions wait for a person's approval.
5. 5Write the rules in plain language the business can edit: how to qualify a buyer, when to offer a visit, what never to promise.
6. 6Add checks that run on every reply before it is sent: prices must match the catalogue, offers must come from your list, products must be in stock.
7. 7Test on real conversations from the last month, replayed against the new agent, and fix what fails before going live.
8. 8Roll out in stages — shadow (agent drafts, people send), assisted (agent sends, people watch) and autonomous — and review the numbers every day.

Tools, not freedom

An agent is only as safe as the actions it can take. Each tool should do one thing, validate its own inputs and refuse anything outside its rules. A “change address” tool that checks dispatch status itself cannot be talked into changing a shipped order, however the conversation goes.

Rules the business can edit

The sales logic belongs to the business, not to the engineers. We keep it as short written skills — “ask for the budget before showing more than two rings” — that the owner can change. Every change runs against a test set of real questions before it goes live, so an edit cannot silently break answers that worked yesterday.

Guardrails every agentic workflow needs

One rule we apply everywhere: the agent never invents urgency. No “only 2 left” unless stock says so, no deadline that is not a real offer end date. Beyond being good practice, false urgency is listed as a dark pattern in India's 2023 consumer protection guidelines.

How to measure an agentic workflow

Measure the outcome, the quality and the cost, every day, from the first week.

• Outcome: leads, bookings, orders or resolved requests per 100 conversations.
• Quality: share of messages the agent could not answer, replies corrected by your team, hand-offs to a person.
• Speed: time to first reply, and how many enquiries arrive outside business hours.
• Cost: AI cost per conversation and per outcome, so the workflow pays for itself.

How we built our own AI Sales Assistant

Gwiksoft's AI Sales Assistant is the first agentic workflow we run in production: it answers WhatsApp and website enquiries, shows the right products and turns conversations into leads and store visits. In brief, it has four layers:

• Channels: WhatsApp through Meta's official WhatsApp Business Platform, a chat button for any website, and a shareable chat page. A website chat can move to WhatsApp with the customer's context.
• Knowledge: the live catalogue (Shopify or a product list, including search across stores with hundreds of products), the business's knowledge base, its live offers and written sales skills.
• Reasoning and checks: a fast, low-cost model answers first; every reply is checked against the catalogue and rules, and is redone on a stronger model when a check fails. Questions it cannot answer go to the team.
• Actions: product cards, store-visit and appointment booking in two taps, call-back requests, buy links, and polite follow-ups when a customer goes quiet, which stop as soon as they reply.

The business dashboard: where the owner stays in control

Everything the agent does is visible and editable by the business, without engineers:

• One inbox for every WhatsApp and website chat; anyone on the team can take over a conversation and the agent steps back.
• Leads and appointments created by the agent, with the customer's name, number and what they asked for.
• Offers with real end dates, which the agent quotes and never goes beyond.
• Follow-up timing, wording and quiet hours, with a preview before switching on.
• Agent policy: when to hand over to a person, by keyword, confidence or opening hours.
• Analytics from chat to lead to booking, plus Google Analytics and Meta pixel events for ads.
• Corrections: mark a wrong reply, write the right one, and the agent follows it from then on — plus weekly improvement suggestions drawn from real chats.

What it does in practice: for a jewellery brand on our platform, over 30 days the assistant answered 91% of 405 customer messages, with a typical reply time of 2.6 seconds. One in three messages arrived between 10 pm and 10 am, and 61 conversations became leads with a WhatsApp number for the team to follow up.

Agentic workflow readiness checklist

• One workflow chosen, with a one-line success metric.
• Product data complete: titles, prices, variants, stock and photos.
• Policies written down: delivery, returns, exchange, warranty, timings.
• Every action listed with who may approve it.
• A WhatsApp Business Platform number, or a plan to get one.
• A person named to handle hand-offs during business hours.
• Last month's real customer questions collected as a test set.
• A daily report agreed: outcome, quality, speed and cost.

Frequently asked questions

Related reading

• Agentic commerce workflows at Gwiksoft
• AI + human customer support
• Agentic commerce on Shopify: Catalog, Universal Cart, Checkout Kit and UCP explained

Gaurav Kumar advises Gwiksoft on AI and commerce engineering, working with Shopify Plus, Hydrogen, GraphQL and agent frameworks.