How TAC Built an Agentic Chatbot with Haystack to Transform Trade Promotions Workflows

See how TELUS Agriculture & Consumer Goods (TAC) gives users unprecedented access to their data with safety in mind

October 6, 2025

When a leading company like TELUS Agriculture & Consumer Goods (TAC), with a strong presence in agriculture and consumer goods, turns to AI to streamline complex processes, it’s worth taking a closer look.

TELUS Agriculture & Consumer Goods helps businesses optimize everything from supply chains to retail operations. One of their latest innovations: an agentic chatbot powered by Haystack that simplifies how users interact with their trade promotions platform.

We sat down with the team behind this project to learn how they built it, why they chose Haystack, and what advice they have for other teams looking to implement Retrieval-Augmented Generation (RAG) and agent-based AI solutions in production.

The Challenge: Simplifying Complex Workflows

The team’s mission was clear: give users instant, intelligent access to their data without requiring them to dig through documentation or rely solely on the existing UI.

TAC’s trade promotions platform is where large consumer goods companies manage the sales incentives and agreements they set up with retailers and distributors. Think of discounts, special in-store displays, or seasonal promotions. The platform helps teams plan, track, and measure the impact of these promotions across products, customers, and time periods.

Previously, users had to navigate dense manuals to find the information they needed about promotions. This slowed down decision-making and made even simple questions difficult to answer quickly.

The chatbot changes all of this. Instead of manually searching documentation or waiting for new reporting features, users can now ask questions conversationally and get real-time answers. For example, someone managing promotions for a product can quickly check which campaigns are running, how much has been spent, or what results have been achieved — and share that information across their organization. Whether it’s a one-off question or a request for deeper insights, the system delivers immediate results.

As the team put it, this was about “giving users unprecedented access to their data” while removing the bottlenecks of traditional reporting systems.

Choosing Haystack: From Exploration to Production

When the team began exploring options, they came across Haystack while researching RAG workflows. What they needed was a framework that could support retrieval-augmented generation, handle tool calling for tasks like SQL query generation, and provide streaming capabilities for real-time interactions.

Haystack quickly stood out.

“Haystack checked all the boxes and was easy to get up and running. Switching to Haystack gave us exactly what we needed.” says Kelsey.

The team had some experience with other frameworks, but Haystack’s flexibility, modular design, and ease of use allowed them to build a proof of concept almost immediately, earning early buy-in and helping them move fast from experimentation to production.

The Technical Architecture: From Pipelines to Agents

The first version of the system used a pipeline-based architecture with two separate workflows:

One workflow handled knowledge base queries via RAG after ingesting user documentation and converting it from raw HTML into a searchable format
Another generated SQL queries from user inputs using metadata and schema descriptions, then executed them on the SQL database

A topic router decided which workflow to trigger. However, this setup quickly proved too rigid for real-world use.

The solution was to move to an agent-based architecture. Instead of isolated pipelines, the team wrapped each capability—the RAG workflow and the SQL pipeline—into separate tools that an agent could call dynamically.

Switching to an agent-based model was a game-changer. With Haystack Agents, the system gained the ability to:

Retrieve documentation and query the SQL database within a single conversational flow
Retry and rewrite queries based on error messages (self-debugging)
Deliver emergent behaviors like combining insights from multiple sources

Key Components of the Solution

Knowledge Base Tool: Ingests user documentation (HTML → searchable documents)
SQL Tool: Generates queries with metadata awareness, using MS SQL + SQL Alchemy. This tool enables end users to both get data faster and access datasets that may not have been easily reachable before.
Observability & Monitoring:
- OTEL (OpenTelemetry) for observability
- Langfuse + Sentry for monitoring and debugging
ETL & Development: Kedro for ETL pipelines, Gradio for UI testing
Guardrails & Safety:
- Restricting the incoming SQL queries to SELECT statements with LIMIT clauses
- Sanitizing SQL table/column names before output to prevent leakage
- Enforcing user identity constraints on generated queries

The team’s focus on security and reliability ensures sensitive data stays protected while users enjoy a smooth experience.

Evaluating the Performance

To track performance, the team focuses on three main signals: latency, accuracy, and user engagement.

Tool and LLM response times are monitored with Langfuse, while accuracy initially relied on human-in-the-loop evaluation before moving toward automated benchmarking. Feedback from early users guides iterative improvements and feature priorities.

The knowledge base bot was deployed to production this month, while the SQL-generation bot is scheduled for later this year. As adoption grows, the team expects to share more impact stories around productivity gains and user adoption.

Lessons Learned & Advice for Other Teams

The top advice from the team is to start small with a minimal vertical slice of functionality.

“Starting small gave us confidence” says Kelsey. “Our first proof of concept was simple: two pipelines with topic routing, shown through a Gradio demo. But it proved the value and helped us scale up with certainty.”

More advice for anyone considering a similar project:

Invest in observability early for debugging and insights
Use agents for flexibility rather than hardcoded pipelines
Automate evaluation to speed up iteration cycles

What’s Next

The journey doesn’t stop here. TAC plans to automate data ingestion processes, build an AI-based evaluation framework to score chatbot responses, and even white-label the application for other use cases across the organization.

“We see this as a template for the future,” says Kelsey. “Our goal is to bring this capability to more workflows across TELUS Agriculture & Consumer Goods, including internal documentation and SQL tooling.”

The TAC team’s journey shows what’s possible when innovative teams combine Haystack with real-world challenges. From streamlining SQL workflows to instant access to documentation, they turned a complex problem into a powerful, production-ready solution. Get started with Haystack and build reliable, production-ready AI applications for your own team.

We know there are many more stories like this out there. If your team has built something exciting with Haystack, whether it’s a chatbot, a retrieval system, an agent, or an AI-powered internal tool, get in touch with us. We’d love to feature your work and share how you’re shaping the future with Haystack.