An AI SDR that showed what was missing.

A machine gun, not a sniper rifle.

I proposed a discovery interview with the manager to understand what the operation needed. He summed up one problem like this: "the market asks me to fire a machine gun, and I'm shooting like a sniper".

The proposal, built together with management, was an AI SDR on WhatsApp that could reach leads at scale and hand qualified ones to a salesperson.

A reviewer before every message.

I designed the agent as a graph in Python and LangGraph. One node plans the goal of the turn, another answers with operational context retrieved through RAG, a reviewer checks whether the answer is grounded and meets the goal, and a final step adapts the text to the rhythm of WhatsApp.

Without the reviewer, tests showed the agent drifting from the planned intent. Smaller models classify and decide the handoff, and the larger one only runs where reasoning is needed.

Cost per conversation before scaling.

An agent that talks to hundreds of people a week can get out of control if nobody knows what each message costs and what it brings back. I logged the cost of every node on every turn, next to the outcome of each conversation.

Almost all of the cost sat in three steps and came from context size, not from the answer. That is where the lever is to grow without losing quality.

Good at exploring, not yet at selling.

In the August 2026 pilot, IARA reached out to 144 contacts from an old cold list. 59 replied, about 4 in 10. Of those 59 conversations, 10 ended with a decision maker found or a lead handed to a salesperson, about 1 in 6. The conversation flow ran with no execution errors.

IARA explored well when the lead was open to talking, but she didn't store everything she learned and couldn't pick up a stalled conversation. She lacked tools, clear goals for each stage and a follow-up cadence.

A conversation from the pilot.

During the pilot, we followed every conversation closely. In this one, the list was out of date: the contact had left the company almost two years earlier. IARA didn't stop there. She asked who had taken over, explained what we do when he asked, and found out he was now maintenance manager at a plant that runs the equipment we serve.

After reading the conversation, we handed the contact to a sales rep. Nineteen days later she visited the company, and we are now arranging a talk on heat exchangers at the plant.

Right after this excerpt, he asked whether he was talking to an AI, and IARA claimed to be a person. The cause was the persona, which told her not to identify as an AI. That same day, I replaced the instruction with an explicit rule: when asked, she says she is a virtual assistant. I reinforced the rule in code, so it doesn't depend only on the document, and added a test with variations of the question.

The system marked this lead as an invalid contact, because the number no longer belonged to the old company. The most valuable information in the conversation, the new company, stayed only in the text. It is one of the reasons the SDR depends on the CRM to record the lead's cycle.

Fewer questions, more conversation.

BANT qualification turned into a string of questions that pushed leads away. I replaced it, along with the rule of always closing on a quote request, with closed goals: role, company, industry and need.

I built a sandbox where the team talked to the agent as if they were a lead. There, back-to-back questions dropped and the agent handled leads who went straight to the point better. At the manager's request, the way the sales reps talk became the reference for tone.

Measure the cycle before scaling.

Without a CRM, there was no way to know how much IARA contributed to a lead's path to purchase. The SDR moved to a parallel track while the CRM took shape, and the CRM has already started to record that contribution.

The next batch focuses on what the pilot did well: qualifying cold leads with little information, finding decision makers and sending catalogs.