Let's talk
All posts

What I Built at Aircall: AI Slack Agents, Self-Service Data and Data Pipelines

· 4 min read

At Aircall I worked in the data team on three main projects, plus a number of smaller tasks for other teams. Most of the work sat where data engineering meets AI: getting company data to people through conversational interfaces they already use.

1. AI-powered Slack applications

My first two projects were Slack applications. I designed and built Slack bots that use Slack Block Kit for the interface and connect to AI agents working over predefined company data.

Customer and Sales Insights

Customer and Sales Insights brings key account data and AI insights into Slack. Users could:

  • Look up core metrics for any company, such as MRR and user count.
  • Get AI insights on companies at risk of churning.
  • Ask questions about past conversations between the customer and Aircall team members.
  • Get a summary of the customer’s whole journey with the company.
  • Take follow-up actions from the same view.

The churn risk came from a churn prediction model built by a colleague. My part was the application layer: the Slack integration, the AI agent, and the data plumbing behind them.

Monitor Pro Subscription

Monitor Pro Subscription helps the company manage its spend on internal tools, especially paid SaaS licences such as Hex and Monte Carlo. It shows who actually uses each tool, so seats can be right-sized and unused licences removed.

On Hex alone, we cut the number of paid seats from 60 to 20, which saves about $80K per year.

2. Self-service data with AI: FastData

The second area was data self-service across the company, built on AI-driven SQL generation.

We evaluated existing options, including the CastorDoc and Hex agents. I also built our own agent, FastData:

  • A Custom GPT application backed by an MCP server that uses function calling.
  • A large, curated knowledge base about our data models, so the agent understands the business context.
  • An API layer built with Flask and running on AWS Lambda (Python), which runs the generated queries against Amazon Redshift.

Anyone in the company can ask questions in plain language, for example:

How many users did company X have last year?

How many calls did user X make in the last 10 days, and how long were they?

FastData reached about 65% accuracy in end-to-end SQL generation and execution across our whole warehouse. Next on my list for it are more specialised agents and a full chatbot experience.

3. Data pipelines with Airflow and dbt

Our warehouse is modelled with dbt and orchestrated with Apache Airflow. I built many DAGs (directed acyclic graphs), mainly for:

  • Event-driven Slack notifications for go-to-market teams, mostly about Voice Agent (VA) customers (see below).
  • Historizing tables, for example keeping a history of the churn prediction model’s output so it can be analysed over time.

Voice Agent notifications

Voice Agent is Aircall’s AI voice assistant, which can answer calls on a customer’s behalf. To support its adoption, I built pipelines that watch for key customer events in the warehouse and push alerts to dedicated Slack channels.

The main audience was the customer representative and sales teams. They used the alerts to reach out to customers at the right moment: to welcome a new VA user, help one who was struggling, or pitch Voice Agent to an account that clearly needed it. Managers used the same channels to track opportunities, especially customers missing many calls who had not yet signed up for Voice Agent.

The events we tracked included:

Event Why it matters Who acts on it
Customer signed up for Voice Agent Welcome and onboarding outreach Customer representatives
Customer made their first VA call Confirms setup worked; good time to check in Customer representatives
Customer reached their first hour of VA calls Adoption milestone; chance to expand usage Customer representatives
Customer missed most of their calls Clear sign they need Voice Agent Sales, managers
High missed-call rate but not signed up for VA Strong upsell opportunity Sales, managers
Customer signed up but hasn’t used VA after N days Stuck in onboarding; needs help Customer representatives
VA usage dropped sharply week over week Early churn signal for the feature Customer representatives
VA trial ending soon Conversion conversation Sales
Customer disabled or cancelled Voice Agent Find out why; try to win back Customer representatives, managers
Weekly VA adoption digest Sign-ups, usage, and top missed-call accounts Managers

Each alert included the company, the key numbers behind the event (such as missed-call rate, MRR and user count), and a link to the account. Reps could act straight from Slack without opening a dashboard.

Tech stack

Area Tools
Languages Python, Node.js, SQL (Redshift / PostgreSQL dialect), Bash
AI & agents Custom GPTs, MCP servers, function calling, Claude
Data dbt, Apache Airflow, Amazon Redshift
Cloud & infra AWS Lambda, Terraform, AWS YAML templates (IaC)
Observability Datadog, Amazon CloudWatch
Automation & tooling Zapier, Slack Block Kit, Cursor

Wrapping up

Across these projects I connected machine learning outputs, a well-modelled warehouse, and LLM-based agents to tools people already use every day, like Slack and ChatGPT. Not everything went as fast as planned, and infrastructure blockers slowed some work down. Still, the work showed how much faster teams can get answers when data is a question away.