How 12 years, 5 startups & 500k tickets shaped Mirage's AI-first CX
Our AI-first support model improved customer experience & product quality and changed our team design
In the summer of 2024, our customer experience team at Mirage (formerly Captions) hit a breaking point.
For months, we had received a string of one star reviews on the App Store with customers complaining that customer support was a dead end and that no one ever replied. I was horrified to discover that one of our remote, contracted agents had quietly deleted hundreds of tickets per week from her queue to reduce her workload.
Inbound volume was increasing 30% month-over-month, our time to first response was suffering, the team was stretched thin (especially after letting that remote agent go), and morale was slipping. Despite working at an AI-first company, our CX team was set up no different than the way I might have done so a decade earlier. I was hesitant to relinquish the most human-centric part of our customer experience to AI. Although it’s easy to forget, just a year ago, everyone was still concerned with hallucinations.
Determined to fix things, I cleared my calendar and set meetings with every vendor who was offering an AI Customer Support solution. I recognized an opportunity to improve life for my team as well as provide a better customer experience for our rapidly growing customer base and dove in head first.
Why the old support model is broken
You are probably offering support to your customers the same way you were in 2015.
Ironically, in this traditional model, the person who can help the least talks to your customer first. While your front line agents may know a lot about your product and truly care, they are most likely lacking necessary permissions and are kept out of the loop (no fault of their own). I’ve seen this time and time again, especially in older, larger, and regulated industries.
As a customer, when the agent you're speaking to or chatting with can’t make it right, they escalate the issue again and again until it finally lands with someone who might be able to do something about it. By the time your team has closed the loop with a complex workaround or bad news, your customers are long gone. And in many cases, seeking out your competition.
Hiring our first AI agent
The AI agent we chose wasn’t a chatbot (notice how no one says chatbot anymore). It was a fully autonomous agent from Parahelp that worked with our existing messaging platform, Intercom.
I assumed it could only handle the simplest issues. But on day one we were resolving close to 50% of all inbound volume. Within weeks, that rose to 65%, and we are now at a 75% resolution rate. Customers are getting instant, accurate, and thoughtful responses.
One conversation that stood out: A customer asked how credits worked by using an analogy. Our AI Agent understood the analogy, extended it, and then tied up the conversation in a bow. It did a better job communicating than I could have. Here is the actual exchange pasted below:
If you are considering deploying an AI Support Agent, I’d suggest taking a look at Parahelp (for fast-growing, software companies), Decagon (for digital-first brands), Sierra (for enterprise-level B2B companies), Siena (for e-commerce), and Fin (for nearly any use case).
I chose Parahelp for the following reasons:
I was incredibly impressed with their founding team. I met Anker and Mads while they were still in Y Combinator and was blown away with their vision and focus. In full disclosure, after our go-live I was so excited about what they were building that I wrote my first ever angel check.
As their first customer, they were willing to build their product specifically for my particular use case.
Lastly, and perhaps most importantly, their solution didn’t require any engineering support, which meant I could deploy quickly and without the risk of burning eng resources.
After just a week, our AI Agent's customer satisfaction (CSAT) scores were higher than with our human agents. I called an all-hands with my team and shared that, if they hadn’t already realized, AI was changing customer support more quickly than we thought and that our jobs all needed to change.
Building an AI first customer experience team
At Mirage, the Customer Experience team’s motto is #C4L (Customers for Life). We ask ourselves every day: how can we create experiences inside and outside of the app that drive exceptional customer outcomes? How do we ensure that our customers, no matter what new competitors pop up, will continue to choose us?
Today, our AI agent works 24/7, never burns out, and never loses patience. You might be asking how our human agents feel about this. In general, the reaction has been incredibly positive because we’ve been able to upskill our team to focus on higher-value, less repetitive work. They can now:
Build rapport and go the extra mile with customers
Handle sensitive trust and safety escalations
Act as a SDR by recognizing sales opportunities
Have more impact on the product by working more closely with the product team
In place of the vicious cycle of scaling headcount linearly to keep pace with growth, we now see growth as giving us resources to scale customer impact.
This has transformed our org structure, which used to look like a pyramid. This included many frontline agents at the bottom, fewer tier 2 and escalation agents above them to handle more complex and technical issues, and a single manager or director at the top.
At Mirage today, the traditional pyramid has transformed into something that more closely resembles a diamond. At the bottom is a single AI agent carrying the majority of frontline support. In the middle sits a group of cross-functional (human) Customer Experience agents focused on higher-value, complex objectives. This includes customer success, product management, and sales in addition to customer support work; with the unifying mission of providing exceptional experiences that create customers for life. At the top of the diamond is a single Head of Support / CX managing both the human team and the AI agent.
Providing the four essentials to every agent
Just like humans, your AI agents require four things when communicating with your customers:
Knowledge: Your agent needs to know as much as possible about your product and company. Your product managers and engineers need to communicate accurate, up-to-date information about new releases.
Tip: You need accurate documentation to train your agent; so it is critical that product, engineering, and design share knowledge openly so your agents always know as much or more than your customers. Our product team now supports multiple releases per day, so this isn’t optional.
Context: Agents need to understand who each particular customer is. This includes knowing the customer’s subscription type, usage, support history, and commercial information such as credits remaining, and annual spend. All of this content should inform how your agent replies.
Tip: Engineers can easily send this data directly to your support platform so AI and human agents don’t have to guess (or ask).
Access: Agents need the ability to take action. If an agent has to open a new tab or log into another system to issue a refund or lookup a user account, you're already behind.
Tip: AI Agents already live inside many tools (Intercom, Linear, Notion, etc.), and the breakthrough comes when they can talk to each other.
Rules: Agents require guardrails for what they can share (ex, a feature that is still in alpha), and what they can do (ex, provide refunds for customers that used all of their credits). You cover this on day 1 of new hire training and on an ongoing basis.
Tip: Train your AI on common support scenarios, including acceptable guardrails on behavior. Teach them how to respond to questions such as “Are you AI?”, or “Can you refund my last 50 transactions?”. Treat these scenarios as part of your QA process; document, test and refine them constantly.
Automating knowledge management
Over the years, one of my biggest issues with updating help centers was knowing that the second I hit “Publish,” the information was already out of date. This is not an insignificant problem: knowledge isn’t just one of the four pillars of support, it’s the foundation that holds everything else up.
In the world of AI support agents, proper documentation is more critical than ever. Unlike human agents, AI cannot listen to hallway conversations to tap your team’s tribal knowledge. Remember, your docs are now the source of truth for both human and AI consumption.
So in the next few months, Mirage will have self-updating documentation, powered by Mintlify. Product changes, bug fixes, and new features will flow directly into our knowledge base through automated feedback loops. AI agents are absorbing those changes, and suggesting edits that in-the-loop humans approve and refine as needed.
This will not only accelerate the pace by which our support agents can train and assimilate new information, but enable employees and customers to retrieve accurate documentation whenever and wherever it is needed: be it Slack, community forums, support chats, or in-app. I believe that companies that master this living, self-updating knowledge system will set the new standard for customer experience and Mirage CX’s team aims to be at the forefront.
Building a real time customer feedback loop
At Mirage, we are also focused on eliminating the time between when a customer reports an issue and we ship a fix. It used to be that only engineers could write code, but with tools like Cursor anyone can ship changes quickly. Today our team can rapidly respond to deliver what customers need without unnecessary bottlenecks (mostly, engineering bandwidth).
We are still early in this journey, but imagine this:
Customer reports a bug or feature request via an in-app messenger
AI Agent triages, including
PM or Engineer approves the PR (as a human-in-the-loop) and it gets pushed to prod
AI Agent automatically gets back to the customer
In full transparency, we are early in this next phase. Our goal is that in the next few months this feedback loop could take hours or mere minutes (especially if we remove the human-in-the-loop approval). In this new model, feedback moves directly from the customer to the codebase. The gap between a customer-reported problem and an engineered solution collapses.
Next Steps for CEOs
If you are a leader or CEO that is serious about adopting an AI-first customer experience, here is what you need:
True buy-in from leadership, product, and engineering.
This is not just a thumbs up from your team, it’s about real partnership. The AI feedback loop only works when CX, product, and engineering are aligned and committed to closing the gap between customers and the roadmap. AI makes this faster and more visible, but the discipline of listening and acting on feedback is required for even traditional support organizations.
A CX leader who is ready to embrace this change.
At a younger company, you could build your CX organization as AI forward from day one. But at a more established company, it likely means rethinking the entire org structure, and eliminating roles that no longer make sense in an AI-driven model. In many cases, frontline agents, Tier 2 agents, and even QA specialists may be replaced or significantly reduced. You need a CX leader who is willing to own and navigate that transformation.
A company-wide culture that embraces AI and its failures.
AI isn’t perfect. There will be misses, misfires, and awkward moments along the way. When we first launched, someone asked if the agent they were chatting with was AI and it kept introducing itself over and over again. What matters is whether your company is willing to treat those as learning opportunities and move forward. The organizations that win will be the ones that iterate quickly, not the ones that wait for perfection.
As CEO, the practical next step is to sit down with your Head of CX or Support and ask four straightforward questions:
What percentage of our messages are being resolved automatically by an AI agent today? If the answer is 0%, ask them why and what their plan is.
What is our CPR (Cost per Resolution) or CPT (Cost per Ticket)?
Find out how much it costs to resolve a single customer contact. Add salaries and tools and divide by number of tickets. AI can provide a huge cost savings opportunity while improving customer happiness.Are you, as CX leader, capable and willing to adopt this model?
Head in the sand simply isn’t an option. Transformation will stall before it starts unless there is acceptance and excitement.Which vendors are we evaluating?
Schedule meetings with vendors. Vendors and models are moving so quickly, it’s hard to keep up.
Start adopting an AI-first CX mindset today
If you will indulge me, here are some predictions about the future of Customer Experience and where I am keen to go in the medium to long term:
Customer Support will become part of Product. Heads of Support will report directly to the Head of Product because at the end of the day, support is responsible for quality. Aligning under Product ensures customer signals turn into product improvements, especially in a world where human engineers are no longer needed to implement most bug fixes.
90% AI resolution rates will be common. For the humans left in the support loop, many will be technical support engineers reviewing complex edge cases as well as high EQ, cross-functional agents. Clearly written PRDs and procedures will be more important than ever.
Customers will vote without knowing it. Customer sentiment will flow directly into product development pipelines, influencing what gets built and in what order. A/B testing products may disappear or happen more autonomously.
Many roles inside CX will disappear. Front-line agents, Tier 2 agents, and even QA specialists will likely be reduced drastically. The hardest part of running a support team has always been building a bridge to engineering and staying aligned with the product team. AI is now closing that gap.
Engineering graveyards will disappear. A few small issues in your product don’t amount to much, but thousands of these papercuts add up. As customers report these small issues, small code fixes will be automated. Customers can’t see one small change, but when many are made, they can feel it.
AI-native customer experience will dominate. Companies with CX teams adopting AI the right way will offer better experiences, build products faster, and outcompete traditional teams with siloed feedback loops.
There is of course still a place for white-glove experiences, especially if you’re B2B with very high ACVs, where building rapport and long-term relationships matters. Customer Success needs to be part of the equation and work hand-in-hand with Customer Support.
But customer support teams have always been the first layer between you and your customers. Today, twelve months into our journey of adopting an AI-first CX mindset, I can tell you that the impact of AI agents goes far beyond cost savings.
AI agents are enabling our team members to focus on delivering an exceptional customer experience. This team is providing our customers with answers more quickly, more accurately, and with more empathy than ever before. They are enabling anyone to get access to accurate documentation wherever they are. And they are collapsing the time for engineering to resolve issues and implement fixes. If you are a CX leader or CEO, start adopting an AI-first CX mindset today. Otherwise, you are doing your customers and team a disservice, and your company is already behind.
I would like to thank the members of my team who have embraced this AI-first experiment with me. Thank you: Shrenik, Marian, Karl, Mehdi, Nate, and Siobhan.
About Eli
Eli Winderbaum is currently the Head of Customer Experience at Mirage (formerly Captions). Prior to Mirage, Eli led Customer Support teams at Domain Money and Clarity Money (acquired by Goldman Sachs). Eli was one of the first hires at BetterCloud, where he built and grew the Customer Support and Success teams. Eli lives in Brooklyn with his wife, Shaun.
Eli would love to hear from you. To get in touch, reach out on LinkedIn or send him an email eli@eliwinderbaum.com. Eli is intentionally not on Twitter.
If you live in New York, Eli will be speaking more about how to scale Customer Experience with AI at the following events:
September 24 at Founders Club NY: CX AI Tools & Strategies
October 14 at the App Growth Annual Conference
 | A guest post by
|
Had a great time writing this specifically for the NACEO audience. DM me if you have any questions or want to chat!
Excellent. This article is a keeper for anyone who is pondering how to strike a balance between AI and human customer support.