Editorial Team Using AI in customer support workflows is a practical necessity now for any brand wanting to keep a competitive edge. This case study breaks down a recent campaign where we implemented AI customer service and predictive assistance with a conversational AI platform we called “Iris.” The objective was to reduce customer service operational costs while also improving customer satisfaction scores.
Key Takeaways
- By deploying Iris for routine questions, we cut average handle time (AHT) by 22%, which beat our initial 15% target.
- Customer satisfaction (CSAT) scores for chats handled entirely by Iris jumped 8 points from the baseline to hit 87%.
- The 10-month campaign’s $750,000 budget produced a 3.5:1 return on ad spend (ROAS) from cost savings and better customer retention.
- The biggest snag was data sync across legacy systems, which needed an extra $120,000 for integration middleware.
- Future work will be on expanding Iris’s proactive outreach, getting it to do more than just reactive support.
Our client, a mid-sized e-commerce retailer selling home goods, was getting buried under rising customer service costs and inconsistent response times. Their model was almost entirely human agents, creating huge overhead and slow turnarounds for simple things like order tracking, returns, and basic product questions. They decided to pilot an AI-driven solution, “Project Iris,” to automate these high-volume, low-complexity interactions. The goal was to give their agents the breathing room to focus on the complex, high-value problems that actually need a human touch.
Campaign Strategy: Shifting the Support Model
The strategy for Project Iris was built around a phased deployment of the conversational AI. In phase one, we launched Iris as the first line of support on their website and mobile app, where it handled FAQs and pointed customers toward self-service. Phase two got more interesting, integrating Iris with the company’s CRM and order management systems so it could give personalized, data-backed answers like real-time order status updates and product recommendations. The final phase, which is what this teardown is about, introduced predictive assistance, training Iris to spot and solve potential issues before a customer ever had to contact a person.
The whole campaign ran for ten months, from January to October 2026. We had a total budget of $750,000, which covered platform licenses, the integration work, agent training, and a small marketing campaign to let customers know about the new AI assistant. We were laser-focused on a few key performance indicators (KPIs): average handle time (AHT), first contact resolution (FCR), customer satisfaction (CSAT), and of course, the reduction in operational costs. We also kept a close eye on how many people completed a self-service task after Iris pointed them to it.
A huge part of the strategy depended on getting the training data right for Iris. We fed the AI over 50,000 anonymized customer chat transcripts from the previous two years, plus the entire company knowledge base of product specs and policies. This let Iris learn the way customers actually talk and understand their most common pain points. We also built a feedback loop where our human agents would review and correct Iris’s responses, constantly refining its accuracy. Having that human-in-the-loop system was absolutely essential, especially in the first few weeks.
Creative Approach and Targeting
The “creative” here wasn’t an ad. It was the experience of talking to Iris. We designed Iris with a friendly, helpful personality that used clear, simple language. The chat bubble itself was integrated cleanly into the site and app, designed to be there but not annoying. The first message a customer saw was all about speed and convenience: “Need quick help? Our AI assistant, Iris, can get you answers fast!” We focused on making Iris sound helpful and efficient to build trust from the first click.
We targeted every visitor to the website and app, but we built a specific funnel inside the chat. Every chat started with Iris. If the AI couldn’t solve the problem in three to five exchanges, or if the customer typed “talk to a human,” the conversation was passed off to a live agent. Making that handoff feel completely smooth was critical. Any friction there would’ve wiped out all the efficiency we were trying to create.
What Worked: Data-Driven Success
The results from Project Iris blew past our initial projections. The biggest win was the drop in average handle time (AHT). Before Iris, a routine chat took about 6 minutes and 15 seconds. After we launched, that time fell to 4 minutes and 50 seconds, a 22% reduction. This immediately freed up agent capacity, letting the team handle a bigger queue of complex cases without needing to hire more people. This efficiency was the main engine behind the project’s positive ROI.
We also saw a surprising lift in customer satisfaction (CSAT). For interactions that Iris handled from start to finish, CSAT scores climbed from a 79% baseline to 87%. It turns out that for simple problems, customers really value the speed and accuracy of a machine. We think this jump came from Iris being available 24/7 and its ability to pull the right answer from the knowledge base instantly. As HubSpot research has shown for years, instant gratification is a huge factor in customer happiness, and Iris delivered that.
Our cost per lead (CPL), or in this case, cost per resolved inquiry, was another clear win. By automating so many interactions, the effective cost to resolve an inquiry with Iris was just $1.50 (calculated by dividing the $750,000 campaign cost by all the AI-resolved tickets). That’s way down from the estimated $4.00 it cost for a human agent to do the same. This operational saving was a major boost to the project’s financial success.
Even though it was early days for the predictive assistance feature, it showed a lot of promise. Iris started flagging customers who browsed a product page multiple times without buying, or who started a return but didn’t finish. Proactive outreach to these groups through in-app messages and targeted emails led to a 5% bump in conversions for the browsers and a 10% drop in abandoned returns. This shows how AI can move from just reacting to problems to proactively driving business outcomes, a real strategic advantage.
What Didn’t Work: Integration Hurdles and Data Silos
Our biggest headache by far was integrating Iris with the client’s patchwork of legacy systems. They were running an old on-premise CRM, a totally separate inventory system, and a third-party shipping platform. Getting data to sync between them was way harder and more expensive than we’d planned. We had budgeted $100,000 for the integration work but ended up spending an extra $120,000 on custom API development and middleware. That mess delayed the full launch of phase two by almost two months. This is a common pitfall in AI projects. The data plumbing is often a bigger monster than the AI itself.
Another spot that needed a lot of work was Iris’s ability to pick up on nuance or emotional language. It was great with factual questions, but it could get tripped up by sarcasm, obvious frustration, or slang. This resulted in a small but vocal group (about 3%) of customers complaining that the AI didn’t “get it.” We tackled this by constantly feeding it new NLP models and lowering the threshold for automatically escalating a chat to a human whenever our sentiment analysis detected strong negative feelings. It’s a good reminder that even advanced AI still has no real empathy.
Optimization Steps and Future Outlook
Based on what we learned, we made a few key adjustments. We pushed for more investment in the cross-system integration, building a unified data layer that Iris (and the human agents) could access for a single source of truth. This didn’t just help Iris’s performance. It made the agents’ jobs easier, too. We also built out the agent training program with modules on “AI collaboration,” teaching them how to take over a chat from Iris gracefully and how to provide the right kind of feedback to make the AI smarter. This model, with AI and humans working in tandem, is where support is heading.
Looking ahead, the client is all-in on expanding Iris’s proactive features. They’re exploring how to use predictive analytics to spot customers who are at risk of churning and then have Iris offer a personalized incentive to stay. There’s also a new pilot to get Iris working on their phone lines, handling the initial triage of calls before routing them to the right specialist. The success of Project Iris cemented AI as a foundational piece of their CX strategy.
This whole journey with Iris shows that while AI can absolutely change the game in customer service, success depends on good planning, solid integration work, and a constant cycle of learning and improvement. When you’re thoughtful about the implementation, focusing on real pain points and getting better over time, the initial investment really pays off.
What is predictive customer service?
It’s using AI and data analytics to anticipate customer needs or potential problems before they happen. This lets a company proactively offer a solution or information, often preventing a customer from ever needing to contact support. Good examples are sending proactive shipping delay alerts or offering troubleshooting tips based on product usage patterns.
How does AI reduce average handle time (AHT) in customer service?
AI slashes AHT by giving instant, accurate answers to common questions and automating routine work like checking an order status. By taking all the simple interactions off the board, it frees up human agents to work on complex problems, which makes the entire support team faster and more efficient.
Can AI improve customer satisfaction (CSAT) scores?
Yes, especially for routine questions. AI improves CSAT by providing instant answers 24/7 with consistent accuracy. When customers get their simple problems solved fast, they’re happier. For anything complex or emotional, though, having a smooth handoff to a human is key to keeping CSAT high.
What are the common challenges when implementing AI customer service?
The biggest challenges are usually integrating the AI with clunky legacy systems, getting enough clean data to train the AI, managing customer expectations about what the AI can do, and tuning the natural language processing to handle real-world conversations. Fixing these things often requires a serious investment in your data infrastructure and a lot of ongoing work on the AI models.
What was the overall return on investment (ROI) for the Project Iris campaign?
Project Iris hit a 3.5:1 return on ad spend (ROAS). We got to that number by comparing the $750,000 total cost against the hard savings from the reduced AHT, higher agent efficiency, and better customer retention that we could directly attribute to Iris’s work.
