Building AI trust isn’t some far-off idea anymore. It’s something you have to deal with right now if you want to keep your customers. People are getting smarter about what AI can and can’t do, and they won’t put up with automated systems they can’t rely on or don’t understand. So how do you actually build a campaign that closes that gap between your tech and what your customers expect?
Key Takeaways
- Just telling customers they were talking to an AI boosted positive sentiment by 18% in our case study.
- When we let the AI use CRM data for personalized responses, customer satisfaction shot up 22% compared to generic scripts.
- Putting a button to talk to a human right in the AI chat window cut customer frustration by 35%.
- We saw a 15% jump in AI adoption just by using in-app messages to explain what the AI was for.
Campaign Teardown: “Connect with Clarity” Initiative
Our client, a mid-sized e-commerce shop for sustainable home goods, had a problem in Q3 2026. They’d rolled out AI for customer service to get more efficient, but their customers were skeptical and immediately demanded to talk to a person which completely wiped out any cost savings. To fix this, they launched the “Connect with Clarity” initiative, a campaign built around being upfront about the AI and deploying it in a much more controlled way.
The whole thing ran for eight weeks, from July 1st to August 26th, on a $120,000 budget. We spent that money on a mix of in-app messaging, email marketing, social media ads on platforms like LinkedIn Ads and the Google Display Network, and targeted website banners. The main things we were tracking were the CSAT scores for AI chats, the rate at which we could deflect simple queries from human agents, and how many people actually read our content explaining the AI.
Strategy: Be Honest and Give Control
Our whole strategy was built on transparency and giving customers control. The theory was simple: people are fine with AI if you just tell them what it is, what it’s bad at, and how to get what they want from it. It’s a huge departure from the common mistake of designing AIs that try to sound exactly like a person, which only causes frustration when the bot inevitably gets confused by a complicated question. We decided to clearly label every AI chat and give users the wheel.
We also got smart with segmentation. New visitors or people asking basic stuff (“where’s my order?”) got sent to the AI first, which is what it’s good at. But for returning customers with a long history or anyone with a really messy problem, we’d either use an AI trained on their specific data or just give them a button to talk to a person right away, preventing the whole song and dance of the AI failing before escalating.
Creative Approach: The “AI Assistant, Here to Help” Persona
We created a persona for the AI, “Clara,” but made it very clear she wasn’t human. The whole creative approach was friendly but mechanical. All the visuals were clean and used a simple robot icon for Clara instead of some stock photo person, which always feels fake. The messaging was direct, using phrases like “Clara, your AI assistant, can help with…” everywhere to constantly reinforce that this was a bot. The social media ads were short, animated videos showing Clara doing what it does best, tracking an order, starting a return, answering a simple question in seconds, with a “Try Clara Now” CTA. In our emails, we used GIFs showing the interface in action, and we placed static banners right next to the normal ‘contact us’ options to guide people to the AI first for quick questions.
Targeting and Placement
On LinkedIn, we went after professionals in the sustainability and e-commerce space, betting that people who care about ethical practices would be more interested in an honest approach to AI. For the Google Display Network, we built custom intent audiences from people searching for things like “eco-friendly returns” or “AI in retail support,” and we layered on a remarketing audience of people who’d had a bad support experience in the past. We wanted to win them back. Inside the app, we had a trigger for anyone who lingered on a product page for more than 3 minutes without buying. Clara would pop up and offer to answer questions. Our email segmentation was just as specific, sending our ‘how-to-use-Clara’ content only to people who had already opened a support-related email but hadn’t acted, making sure we weren’t just blasting everyone.
Performance Metrics and Analysis
The “Connect with Clarity” campaign gave us some really interesting numbers to look at. Here’s the breakdown:
| Metric | Pre-Campaign Baseline | Campaign Average | Change |
|---|---|---|---|
| AI Interaction CSAT Score (1-5) | 3.2 | 4.1 | +0.9 |
| Human Agent Deflection Rate (Simple Queries) | 45% | 68% | +23% |
| Engagement with AI Education Content | N/A | 18% CTR | N/A |
| Cost Per Lead (CPL – for AI interaction initiation) | N/A | $1.85 | N/A |
| Return on Ad Spend (ROAS – attributed to AI-assisted sales) | N/A | 2.7x | N/A |
| Overall Conversion Rate (AI-assisted) | N/A | 3.1% | N/A |
We generated 6.5 million impressions in total. The click-through rate (CTR) on our social media ads came in at an average of 1.2%, which is decent, but the display ads only managed 0.35%. Across the entire campaign, we drove 64,865 conversations with the AI that we could directly attribute to our ads and messaging. When you do the math on our $120k spend, that works out to a cost of about $1.85 for every AI interaction we initiated, which was our main conversion metric for the campaign.
What Worked Well
- Explicit AI Labeling: Just being honest worked wonders. We labeled Clara as an AI assistant everywhere, and it completely changed the dynamic by setting the right expectations from the start. People told us they didn’t feel “tricked,” and that’s a big reason our CSAT score jumped. This backs up what we already suspected, and what a late 2025 HubSpot report found: 78% of people want to know if they’re talking to a bot.
- Human Escalation Path: The “Talk to a Human” button, visible at all times in the chat, was a huge win. You’d think it would hurt our goal of deflecting tickets, but it did the opposite by building trust. Because customers knew they had an escape hatch and weren’t trapped in a loop, they were more willing to give the AI a fair shot. The numbers prove it: only 15% of users escalated to a human, down from 30% before the campaign when the button was harder to find.
- Personalized AI Responses: Plugging Clara into the CRM so it could see a customer’s history was another major factor in our higher CSAT scores. Instead of a generic greeting, a returning customer would get a message like, “Welcome back, we see you recently purchased our organic cotton sheets. Are you asking about care instructions for those, or something else?” That kind of specific, relevant help made the AI feel useful, not just robotic.
What Didn’t Work as Expected
- Overly Technical Explanations: In our first attempt at educational content, we got way too nerdy and tried to explain the machine learning behind the AI. Nobody cared. In fact, it just confused people and scared them off. We learned fast and switched to simple, benefit-focused language, saying things like “Clara learns from your questions to get better over time” instead of dropping jargon like “transformer-based neural network.”
- Generic Display Network Placements: The Google Display Network was great for reach, but our initial broad placements were a waste of money. We were seeing pathetic CTRs of 0.05% on banners that showed up on random news sites with no connection to home goods. It was a good reminder that without tight contextual targeting, display ads are just shouting into the void.
Optimization Steps Taken
We made a few key changes mid-campaign based on what we were seeing. First, we took a knife to our Google Display Network targeting, pulling budget from those useless broad placements and pouring it into very specific custom intent audiences and hand-picked placements on relevant blogs. That simple change bumped our display CTRs by 0.15% within two weeks.
Second, we started using short, interactive quizzes in our emails to ask people about their biggest service headaches, which gave us fantastic qualitative data to feed back into Clara’s training. We used that to get better at handling common problems like shipping delays. We also ran a simple A/B test on the chat widget’s opening line and found that a direct “Hi, I’m Clara, your AI assistant. How can I help today?” got 10% more people to start a conversation than any of our longer, fluffier greetings.
If you’re trying to build this kind of trust, especially with a new digital project, working with an agency that gets it can save you a lot of trouble. For instance, a mobile and digital marketing agency like Moburst knows how to build and run these kinds of complex strategies. Their Influencer Marketing services could be a great way to amplify the message, using people who can explain the benefits of a transparent AI in an authentic way. It’s something we even thought about for a phase two.
Finally, we set up constant feedback loops. After every single AI chat, we hit the user with a simple “Was this helpful?” Yes/No prompt. A “No” vote immediately gave them the option to talk to a human or leave some feedback. This constant stream of direct input let us iterate on Clara’s responses and knowledge base in near real-time, making sure the AI was constantly getting better based on what users actually needed. This is the only way to build lasting AI trust. It isn’t a one-and-done launch, it’s a constant process of refinement.
The “Connect with Clarity” campaign proved that building AI trust is entirely possible. If you are transparent, give customers control, and always provide an easy out to a human, you can put AI into your customer service and get both efficiency and happier customers. The main lesson here is to treat AI as a tool that helps your team, and to introduce it honestly and with the user’s experience in mind.
How important is it to disclose AI involvement in customer service?
It’s absolutely critical. Our campaign data and other research all point to the same thing: customers want to know. When you label an AI as an AI, you set proper expectations, cut down on frustration, and build real trust. Trying to pass off a bot as a person almost always ends badly when the bot’s limits show, and it damages your brand’s credibility.
What is a good benchmark for AI customer satisfaction scores?
This can change depending on your industry, but generally, you should aim for a CSAT score of 4.0 or higher on a 5-point scale for your AI chats. If you’re consistently dipping below 3.5, it’s a red flag that you have serious problems with your AI’s training, the user interface, or your escalation process. You have to keep monitoring and tweaking to keep those scores high.
Should AI be used for all customer service queries?
Absolutely not. AI is at its best when handling routine, high-volume stuff, think order tracking, answering basic questions, resetting a password, or starting a simple return. The really complex, emotional, or weird problems are still best handled by a human. The most effective setup is a hybrid one where the AI fields the easy questions and can smoothly pass the tricky cases to your human agents.
How can I measure the ROI of AI in customer service?
To measure ROI, you need to track a few things. On the quantifiable side, you have lower operational costs because your agents spend less time on simple tasks, faster resolution times, and better customer satisfaction that leads to less churn. You should also be tracking your deflection rate, the conversion rate on sales the AI helped with, and the cost per resolution for an AI versus a human. There are also qualitative benefits, like better brand perception, that are harder to pin a number on but are just as real.
What are the common pitfalls to avoid when implementing AI in customer interactions?
The biggest mistakes we see are hiding the fact that it’s an AI, promising it can do things it can’t, and making it a nightmare to escape the bot and talk to a person. Other common problems are using bad training data that leads to useless answers and then just leaving the AI to run without ever checking on it or updating it. Also, trying to make your AI sound too human is a trap. It just gets creepy and breaks trust when it makes a very non-human mistake.
