How CarCutter's Head of CX built fair performance feedback for every support rep using Dust

CarCutter is an AI-powered vehicle merchandising solution designed for the automotive industry. The company modernizes vehicle photography workflows for car dealerships, automotive groups, and marketplaces in North America, Europe and beyond.

Using advanced computer vision technology, CarCutter helps dealerships capture, edit, and publish vehicle images at scale. Their platform replaces real-world backgrounds with branded virtual showrooms, corrects image alignment and colorimetry, and delivers finished photos directly into clients' inventory management systems. With 86% of editing automated through AI and human quality control for the remaining work, CarCutter processes tens of millions of vehicle images while maintaining high consistency and quality standards.

Nicolas Harant, Head of Customer Experience and Operations at CarCutter, kept hearing the same frustration from his support reps. The performance feedback system felt arbitrary and unfair.

The management team could only manually evaluate 3% of support tickets that reps handled. The math was brutal: evaluating every ticket would require two full-time employees doing nothing but quality reviews all day. That budget didn't exist.

Support reps received performance feedback based on two or three tickets when they handled more than a hundred each month. Those evaluations directly impacted their monthly bonuses. When a manager happened to review tickets from a difficult customer interaction or a product issue outside the rep's control, the feedback didn't reflect actual performance.

The reps understood the practical constraints, but understanding didn't make the system feel fair. A rep might deliver exceptional service 98% of the time, but if the reviewed sample caught them on a bad day, their bonus suffered.

Beyond individual fairness, the 3% visibility created systemic problems. Knowledge gaps, process failures, and training needs affecting multiple team members could hide in the unexamined 97%. Nicolas needed comprehensive evaluation to improve team performance, but the manual approach didn't scale.

Nicolas had been following the AI conversation online closely. He read LinkedIn articles about operations leaders building impressive automations. He saw people on X talking about building workflows with tools like N8N. The possibilities seemed endless.

But there was a gap between what he saw and what he could actually do.

He didn't have an engineering background. When he started reading about connecting tools, building workflows, and creating automations, it felt out of reach. The technical language, the talk of APIs and integrations, it felt like something that required developers.

Beyond this, he noticed many established SaaS tools were rolling out AI features of their own. He tested these AI features in tools like HubSpot and Notion. They were useful but limited, siloed within each individual platform. The bigger workflows he imagined, connecting multiple systems and automating entire processes, seemed to require resources he didn't have.

Nicolas had notebooks full of automation ideas. Processes that could be streamlined, data that could be connected, repetitive work that could be eliminated. But those ideas stayed in notebooks. The gap between concept and implementation felt insurmountable without engineering support.

CarCutter's CEO, Gaétan Rougevin-Baville, made a decision that shifted the entire organization's relationship with AI. He announced that everyone, regardless of role or technical background, would spend two hours per week working with AI in some capacity. The mandate wasn't about policing how people spent those two hours. The message was about organizational priority. AI was fundamental infrastructure, as essential as email.

Nicolas complemented this top-down mandate with individual conversations. He talked with people about how AI would affect their specific roles and career trajectories. For support reps, he was direct. The repetitive parts of their jobs would be automated in the coming months. The question was whether they would learn to use AI as a multiplier for their expertise or watch it make their current responsibilities obsolete while they stood still.

This framing shifted the conversation. Once people understood that AI could amplify their unique value rather than replace them entirely, curiosity replaced resistance. Teams started discussing processes they wanted to automate and bottlenecks they could eliminate.

Nicolas tested Dust with the support ticket evaluation challenge. The platform removed the technical barrier he'd felt with other tools. He could describe what he wanted to build conversationally, and Dust handled the technical implementation.

Dust offers me this really cool interface where I can just chat with it, and it guides me. It asks me the questions. It enables tools for me, connection with the data that it requires. — Nicolas on using Dust’s Sidekick feature

Nicolas decided to build his support ticket evaluation system around two specialized AI agents, each with a narrow, well-defined purpose.

The Zendesk Ticket Evaluation agent runs every day, identifying and analyzing closed support conversations. When tickets close in Zendesk, the system automatically applies a tag. The agent reads that tag, knows which conversations need review, and gets to work.

The agent uses the same evaluation scorecard that managers previously used. Nicolas embedded the existing standards directly into the agent's instructions. For each tagged ticket, the agent analyzes the conversation and writes results to a Google Sheet with one line per ticket and scores for each criterion.

After completing the evaluation, the agent updates the ticket's tag in Zendesk to mark it as reviewed, preventing duplicate analysis the next day.

The Monthly QA Agent runs once per month, reading all evaluation data from the Google Sheet. It calculates aggregate scores for each support rep, identifies performance patterns, and generates specific, actionable feedback.

The feedback goes beyond numbers. The agent identifies two or three example tickets where a rep's response could have been stronger, shows exactly what they wrote, and explains what would have been more effective. This specificity makes the feedback genuinely useful for professional development.

Nicolas built a custom skill to ensure the output frame maintains consistent formatting every month. Managers see the same dashboard structure with information always in the same locations.

Nicolas created an Apps Script in Google Sheets that automatically archives the previous month's evaluation data once the Monthly QA Agent completes its analysis. This prevents the sheet from growing indefinitely.

He used Dust to write the script even though he'd never worked with JSON or Apps Script before. He described what the script needed to accomplish, Dust generated the code, and he copied it into Google Sheets.

Nicolas initially tried building a single agent to handle everything. The results were inconsistent. When one agent tried to handle multiple distinct tasks with different expected outputs, the quality varied too much.

The solution was splitting responsibilities cleanly. Each agent has clear boundaries and a single area of focus. This design principle has carried through to other workflows Nicolas has built. He recently created an agent in 30 minutes that gathers customer context for churn conversations, something that would have required weeks of coordination with the data team through traditional development.

CarCutter now evaluates 100% of support tickets every single day, up from the 3% they could handle manually. The care manager who previously spent two hours weekly on manual evaluation now uses that time for coaching conversations and strategic work. Nicolas calculated that evaluating every ticket manually would have required two full-time QA employees, positions the automation made unnecessary to create. Nicolas calculated the savings from avoiding two full-time QA hires paid back the Dust investment in seven months for their India-based team, or about one month for companies with higher European or US labor costs.

Beyond the efficiency gains, the system solved the fairness problem that had frustrated the support team. Every rep now receives performance feedback based on their complete body of work rather than a handful of cherry-picked examples. The monthly reports show patterns across all their conversations, giving a truly representative view of performance.

The specific examples the Monthly QA Agent provides make the feedback more actionable than what managers could deliver with manual reviews. Instead of general comments about tone or completeness, reps see exactly which tickets had issues, what they wrote, and what would have been more effective. This concrete guidance helps people improve specific skills rather than leaving them guessing about what to change.

The automated system applies the same standards the same way to every ticket, eliminating the subjective differences that inevitably creep in with human reviewers. One manager's score of 5 out of 10 might be another's 6, and as stated earlier, those differences affect bonuses and career progression. The consistency matters for both fairness and reliability.

Support reps appreciate the fairness improvement. They understand they're being evaluated on their full performance rather than a small sample that might not represent their typical work. And Nicolas has continued building additional automations using the same principles of focused, single-purpose agents that connect multiple systems without requiring engineering resources.

The automated workflows deliver clear ROI, but Nicolas also relies on Dust in a different way for his day-to-day leadership responsibilities.

As Head of Customer Experience and Operations, he oversees teams with deep expertise in domains where he doesn't have specialist knowledge himself. He can't take a traditional hands-on coaching approach in every area, but he needs consistent reporting across all of them. Dust helps him collect information from different sources like dashboards, Slack messages, and email updates, then format everything in a standardized way that makes patterns visible.

The real value is how AI fits into the fragmented reality of leadership schedules. Securing two uninterrupted hours for focused work is difficult when the day fills with meetings, urgent issues, and requests from multiple teams. Building the support evaluation automation would have taken months in the past, assuming he could even get it prioritized on the development roadmap. With Dust, Nicolas can start processes running in the background while handling other responsibilities throughout the day.

He uses the platform to summarize long email threads, research specific legal questions about customer contracts, analyze cost structures for different service tiers, and capture ideas that emerge during conversations. When a thought strikes during a meeting or while commuting, he can speak it aloud to Dust, have it written out with the logic clarified, then filed in the appropriate Notion page for later development.

This captures ideas that would otherwise be lost in the daily rush of operational leadership. Instead of maintaining mental lists of things to explore when time allows, Nicolas externalizes those thoughts immediately and organizes them systematically.

For Nicolas, the transformation happened because the barrier to building disappeared. He doesn't need to be an engineer or learn to code. He describes his processes clearly, and Dust handles the technical implementation. The only thing standing between a notebook sketch and a working automation was deciding to describe it out loud to Dust. If you can explain your process, you can automate it.