Why the team was wary of AI
At mCaffeine & Hyphen, even a simple answer took work: agents switched between several platforms to find one order's status, and customers waited. When the customer experience team looked for a faster way to answer them, AI wasn't the plan. "I wasn't looking for an AI tool. I was looking for workflow automation," says Rakesh Yadav, SVP Business at mCaffeine & Hyphen.
It was the team's first AI rollout, and the team wasn't sure AI could handle most of its queries. The team had also seen what can go wrong: an AI tool Rakesh was asked to try got confused a few questions in, and didn't pass the customer to an agent. The worry was reputation: what would a wrong answer from an AI do to the brands?
What the team needed from the AI
The team needed two things. When the AI knew the answer, the answer had to be right. When the AI didn't know, it had to pass the customer to an agent instead of making something up.
The team answered its worry with evidence, not reassurance. Once CSAT was collected on conversations, the team could compare how customers rated the AI with how they rated agents, and audit the conversations rated badly.
Rakesh also tests the AI the way a customer meets it. When a colleague has a problem with an order, he doesn't fix it for them. He asks them to raise a ticket on WhatsApp, and tells them why: "I really want to know what your experience is as a customer." One colleague who received a spoiled product did exactly that, and rated the experience five stars.
Connecting every system so the AI has the right data
The AI can only give the right answer if it can see the data, and a D2C brand's data sits in many third-party tools. The first job was to connect them all to one system. In Rakesh's words:
"My team used to work across three or four different platforms to find the correct status of an order. I truly believe that the fewer systems agents have to switch between, the more productive they are and the less complex the work is. The main problem with D2C is that every system is third party, and I don't think any company has integrated the entire D2C tech stack yet, except Flowcall."
— Rakesh Yadav, SVP Business, mCaffeine & Hyphen
The team connected its tools to Flowcall one at a time: the store, then the order platform, then the shipping partner. Each new connection was easier to add than the last. Other vendors ask for a requirements document and then wait on their own developers; here, the team adds an integration itself with a short instruction. The AI answers from all of these tools.
Fresh data is what keeps the AI's answers right. The order platform can take about four hours to show what the courier already knows. If the AI read only the order platform, a customer could write in to say their parcel arrived damaged and be told it hasn't been delivered yet. That is why Rakesh values the flexibility to connect any tool: the AI can check the courier's own data instead of waiting for the order platform to catch up.
Rolling out the AI: answers first, actions later
The team rolled out the AI in phases, and each phase was earned by the one before.
Answers only. The AI Agent started on WhatsApp with the basics: where is my order, and has my refund gone through. It answered from connected order data and took no actions.
A month of tuning. In the first month, the team shaped how the AI talks to customers: short replies that get straight to the answer.
First actions, under the team's rules. Only then did the AI get its first actions, each bounded by the team's rules: sharing a customer's refund ID, and following the team's logic for when an order can and can't be cancelled.
The CX team, under the guidance of CX Lead Vikash Pathak, set the rules, decided when to give the AI each new action and reviewed the results. Today the team changes and tests the AI's responses itself, so a wrong reply gets fixed fast. In Vikash's words: "If a customer says a reply was wrong, I know exactly where that conversation was picked up from, where to go and what to change, and what the right response should be if the customer comes back. For me, it's a two-minute task."
Answering questions and taking actions, the AI now handles nearly eight in ten customer conversations with no agent involved.
When the AI doesn't know the answer, it hands the customer to an agent
For questions the AI hasn't been given an answer to, the team set one rule. Rakesh puts it simply: "If your AI doesn't know the answer, don't give a wrong one. Pass the question to our agent instead."
So instead of guessing, the AI tells the customer an agent will help, and hands the conversation to the AI Helpdesk as a ticket, with the order and product details it has already collected. The agent starts from the case instead of asking the customer the same questions again, and solves it. The team then decides whether the answer belongs in its FAQ. If it does, the AI answers that question itself next time.
The AI Helpdesk replaced the team's previous CRM and is now the one system agents work in. Some decisions stay with agents by choice: if a product has caused a reaction, an agent calls the customer. Flowcall also built the ticketing features the team asked for. Speed of change was one of the reasons the team chose Flowcall: "Other companies say they need 15 days to develop something. Flowcall does it in about half an hour," Rakesh says.
From replying to customers to solving their problems
The team doesn't compare tickets per agent before and after, because the work itself has changed. In Rakesh's words: "There's a major difference between replying to a problem and solving it. Today we solve problems. If 100 tickets come in, 90 of them are solved the same day."
The reputation risk the team worried about has also shrunk. By Vikash's count, complaints that reach social media or the founders used to come in once or twice a day; now there are two or three a month.
The AI has also absorbed the brands' growth. In one brand, query volume is about three times what it was when Vikash joined. Customers can now reach the team on WhatsApp and Instagram as well as email, which used to be the only option. Without the AI, Vikash estimates, the team would need 100 to 150 people or would have to close channels.
The team has also changed how the AI sounds, from formal to playful. A customer who writes in anger can end up calming down and leaving a good rating.
About mCaffeine & Hyphen
mCaffeine and Hyphen are Indian D2C beauty and personal care brands. Hyphen is a celebrity-driven brand whose offers bring in many first-time buyers. Customers reach the brands' support team on WhatsApp, Instagram and email.
Start with the questions you'd trust AI with least
mCaffeine & Hyphen started with one test: would the AI get the answer right, and what would it do when it couldn't? Run the same test on your own hardest cases, such as refunds, returns or cancellations. Share those workflows or SOPs, and Flowcall sets up the AI Agent on them and connects the systems it needs. Your team asks the AI the questions it should know and the ones it hasn't been told, and sees whether it answers correctly or hands the customer to an agent. You give feedback, Flowcall makes the changes, and you go live when you're confident. The POC is free. Talk to Flowcall about a POC