Building the AI brain for BDC automotive retail
Designed a unified AI agent system for car dealerships that automates sales, service, and parts conversations
DURATION
3 months (Oct 2025 - Jan 2026)
MY ROLE
Design Strategy, Research, Ideation, Interaction Design, Visual Design
COLLABORATORS
Product Design Manager, Product Managers, Developers
BACKGROUND
Tekion's AI BDC is a CRM-native Conversational AI module built for car dealerships - automating lead engagement, nurturing sales and service conversations, and handling repetitive follow-ups across SMS, email, chat, and voice. It operates 24/7, keeping communication consistent and personalised at scale, so dealership teams can focus on the conversations that truly need a human touch - enhancing efficiency.
As Tekion's AI BDC scaled across dealerships, each with its own communication style, brand voice, and workflows -there was a growing need to configure the AI agent to reflect these differences, ensuring every dealership could deliver a personalised, consistent experience to their customers
Who are we designing this for?
The initial requirement: Tip of the iceberg
The project kicked off with a seemingly simple task:
Design an intuitive Knowledge Base (KB) interface for a single BDC AI agent.
The deep dive
My discovery process involved conducting a deep-dive audit across PRDs, BRDs, and competitive landscapes, while synthesizing the stakeholder interviews and BDC call recordings to understand how BDC agents operate in a real world. A few critical insights emerged:
While my initial research defined the "how" of BDC work, a deeper audit revealed three fundamental gaps in our KB strategy that would have prevented us from reaching the business's long-term vision:
Audit showed 50% overlap in content between departments
Call recordings showed 25% of users asked cross-departmental questions
1 KB per agent would create maintenance debt as we scale to 5 agents
After synthesizing these insights, I couldn't shake one recurring observation.
Although we had defined the right ingredients - Personality, Tone, and Capabilities - we weren't building the robust 'brain' needed for a true BDC ecosystem. I recognized that the strategy needed to shift toward a scalable architecture for all future AI agents, aligned with the larger business vision.
How might we give BDC managers the confidence to set up and trust their AI agent - so customers always get timely, accurate responses across every touchpoint, while laying a foundation that scales as new AI agents are introduced?
Shift towards a modular knowledge framework
Through competitive analysis, I identified multi-agent orchestration with unified routing as an emerging industry pattern - seen in Salesforce Agentforce and Google's agent architecture - where specialised agents operate under a single coordination layer leveraging a shared knowledge base for consistent context.
Recognising the complexity of the real BDC environment, I collaborated with the PM team to apply this pattern to our context - scoping a unified, scalable framework that could handle the same edge cases, fallback scenarios, and compliance standards as the human agents it was designed to support.
This Modular Framework allowed us to treat dealership knowledge as a set of plug-and-play components - ensuring a single source of truth for core information while allowing departmental agents to remain highly specialized and scalable.
Let’s understand this with a real world scenario
The Inquiry: Mark calls reporting a shaking engine and asks if a shuttle service is available
Modular Triage: Ana (AI) checks the Universal KB for the dealership's shuttle policy and the Service DMS for live appointment availability
Cross-Department Action: Ana (AI) seamlessly routes to the Sales KB to gather trade-in data (mileage, condition), alerting a Sales Specialist to prepare a valuation before Mark arrives
Intent Pivot: Mark accepts the service slot, but frustrated with the repairs, asks about his truck’s trade-in value
Defining the AI Fallback Logic
I mapped out fallback scenarios based on how real BDC conversations can break down — with the expectation that post-launch conversation data would help us refine and evolve these over time:
Testing AI before going live
For the AI to operate confidently in a real dealership environment, BDC managers needed a way to test it like they would test a new agent - through structured simulations across the channels customers actually use. I designed an end-to-end test workflow within the platform that lets managers run, evaluate, and improve the AI before it ever speaks to a real customer:
Running the Simulation
From any agent's configuration, managers can launch a test simulation in one click - across both channels customers actually use. Each test opens with the AI's real greeting and pre-defined dealership-specific prompts, so managers can quickly run scenarios without scripting from scratch.
Capturing feedback & Closing the Loop
Inline feedback is captured at every AI response - managers can rate accuracy, flag specific issues like missing details or incorrect tone, and instantly trigger updates to the knowledge base. Every flagged response loops back into the platform's configuration, turning each test into an opportunity to refine the AI.
The impact
We launched the Service AI Agent with a pilot cohort of 28 dealerships - with the Sales AI Agent already in the pipeline for the next 3 months, bringing the vision of a fully unified, multi-agent dealership platform one step closer.
79%
Launch readiness
65%
Appointments scheduled by AI
28
Rooftops live for pilot
Whether you're a potential co-worker, a fellow creative, or simply a curious soul, I'd love to hear from you!
Drop me a Hi and let's get the conversation started.
arushi.aa13@gmail.com
Check out more About me
Copyright 2026 - Made with love by Arushi Arora |
You can also find me on