Solidroad co-founder and CEO Mark Hughes discusses the changing role of AI in customer support and the need to connect customer intelligence with agent training.
The company’s partnership with Unwrap aims to shorten the gap between problems identified in customer conversations and the training required to address them.
Executive Introduction
Solidroad AI Customer Support Training and the Changing CX Model
Customer support operations are increasingly dealing with a more complex question than simply how agents perform during individual interactions: how quickly can an organization identify recurring customer problems and translate those findings into better support performance?
Traditional quality assurance processes often examine only a portion of customer interactions. That can make it difficult for support leaders to identify recurring issues across a large conversation base and then incorporate those findings into training at operational speed.
This is the context in which Solidroad AI customer support training operates. The San Francisco-based company develops AI-powered simulations designed to recreate customer scenarios, systems and workflows, allowing human and AI support agents to practice before handling live interactions.
Its September 2026 partnership with customer intelligence platform Unwrap adds another layer to that model. Unwrap analyzes customer feedback and conversations to identify recurring themes, customer issues, resolution paths and outcomes. Solidroad then uses those findings to create customized training scenarios.
For CX leaders, the significance lies in the connection between two traditionally separate activities: understanding what customers are experiencing and preparing support agents to respond effectively.
Mark Hughes, co-founder and CEO of Solidroad, provides a perspective on this intersection. The company says it works with brands including ŌURA, Ryanair and Crypto.com and has scored millions of customer conversations through its platform.
This interview explores the implications of connecting customer intelligence, AI-powered simulation and support-agent training—and what that could mean for organizations managing increasingly hybrid teams of human and AI agents.
About the Company
Solidroad AI Customer Support Training
Solidroad is an AI-native training platform focused on customer support teams. According to the supplied company information, its platform recreates real-world customer scenarios, systems and workflows through AI-powered simulations.
The objective is to give support agents an environment in which they can practice, build confidence and improve performance before handling live customer interactions.
Solidroad’s approach covers both human and AI customer-support agents. Its simulations can be configured across different personas, support channels, difficulty levels and languages.
The platform also uses customizable, self-improving rubrics informed by an organization’s guidelines, standard operating procedures and knowledge base. Agents receive feedback on their performance through the simulations.
The company is headquartered in San Francisco and says it works with global brands including ŌURA, Ryanair and Crypto.com. Its stated industry coverage includes fintech, travel, ecommerce and consumer technology.
The company’s partnership with Unwrap expands this model by connecting customer intelligence with training. Unwrap’s SupportIQ product analyzes customer conversations to identify the factors driving customer issues, the resolution paths being used and the paths associated with stronger resolution outcomes and customer satisfaction.
The combined workflow is designed to move from customer insight → identified support issue → targeted simulation → agent training.
For enterprise CX teams, this creates a potential operational link between quality assurance and learning rather than treating them as separate processes.
About the Executive
Mark Hughes and Solidroad AI Customer Support Training
Mark Hughes is the co-founder and CEO of Solidroad.
His current leadership role places him at the intersection of AI, customer support operations and workforce training. Under his leadership, Solidroad has developed an AI-powered simulation platform for training and improving both human and AI customer-support agents.
The supplied material does not provide a detailed chronology of Hughes’s earlier career, educational qualifications, certifications or previous executive positions. Those details should therefore be added only if they are supplied in the completed interview materials or independently verified before publication.
As CEO and co-founder, Hughes is particularly relevant to the discussion of how AI can be incorporated into customer-support training, how customer conversation data can inform operational learning, and how support organizations can establish consistent standards across human and AI agents.
Mark Hughes on AI Customer Support Training, Customer Intelligence and the Future of CX
Q1: Solidroad describes itself as an AI-native training platform for modern customer-support teams. What problem in conventional support training convinced you that a fundamentally different approach was needed?
Mark Hughes: My co-founder Patrick and I both spent years around support teams. We met at Intercom, where I was on the sales side and he was in engineering, and the pattern we kept seeing was that training ended right where the hard part began.
New agents get a week of documentation, they shadow someone experienced, they pass a knowledge check, and then they’re put in front of real customers who are unpredictable. Our own survey of 500 support agents found that nearly 70% are trained by shadowing, and more than half say the hardest part of onboarding is applying what they learned to an actual customer situation.
That is a practice problem more than a knowledge problem. In most jobs where performance matters, people rehearse before they perform. In support, the first real attempt often happens with a customer already on the line.
Partnership with Unwrap
Q2: Your partnership with Unwrap connects customer intelligence with agent training. What was missing from the traditional quality-assurance and training cycle that this partnership is designed to address?
Mark Hughes: Both halves of that loop have worked reasonably well on their own. Companies have gotten better at spotting patterns in customer conversations, and they have gotten better at running training programs. The two just have not been connected to each other.
What usually happens is that a recurring issue gets identified, it shows up in a monthly report, it gets discussed on a call, and the training stays exactly as it was. So the issue keeps coming up, and more customers run into it before anything changes.
That is the gap this partnership is built around. Unwrap can see whether a problem is showing up for a thousand other customers, and we can turn that into something agents actually practice. It shortens the distance between a team spotting an issue and doing something about it.
Recurring Themes Across Customer Conversations
Q3: Unwrap can identify recurring themes across customer conversations, while Solidroad can turn those findings into simulations. How does that workflow operate in practice—from the moment a customer issue is identified to the point where an agent is trained on it?
Mark Hughes: It starts on the Unwrap side. Their SupportIQ evaluates customer conversations and feedback and to identify what is driving customer issues, which resolution paths agents are taking, and which of those paths are working. It essentially surfaces customer support recurring themes with evidence behind it.
From there, we build a simulation of that scenario. Not a generic version of it, but something close to the real interaction, with the customer’s tone, the channel it came in on, the pushback they gave, and the systems the agent has to work in while the conversation is happening. It is scored against that company’s own guidelines and knowledge base, and the agent gets feedback as soon as they finish.
The advantage of building training from real conversations is that agents rehearse the situation they are actually going to face rather than a version of it somebody invented. They get feedback immediately instead of at the next review cycle, they can run it again until they are comfortable, and a manager can see who is ready before anyone goes live. New hires get to make their mistakes in a simulation rather than in front of a customer.
Fraction of Customer Interactions
Q4: Manual QA typically covers only a fraction of customer interactions. What changes when AI can analyze conversations at much greater scale, and how should CX leaders interpret the additional signals without confusing volume with insight?
Mark Hughes: Most teams have been working from one or two percent of their conversations and treating that as the whole picture. In our survey, 81% of agents said most of their conversations are never reviewed, so there is a lot going on that nobody has visibility into.
When a team can see across every conversation instead of a sampled few, they stop debating whether the ones that got pulled were representative and can deal with what is actually coming up. That is a real step forward.
The caution in your question is a fair one, though. I have seen teams go from having too little information to having far more than they can use, and end up with a dashboard nobody opens. Volume on its own does not create understanding.
The test I would apply is whether a finding connects to a decision. If something surfaces and nobody can say what they would do differently because of it, it is not ready to act on yet. That is a lot of why this partnership made sense to us. Unwrap does the work of prioritizing what matters, and we make the next step something a team can actually run.
Directly Shape Training
Q5: One of the more interesting claims behind this model is that customer conversations can directly shape training. Can you give an example of a customer issue that was identified through conversation data and subsequently translated into a training scenario?
Mark Hughes: Faire is a good example. They handle thousands of tickets a week, with a lot of them coming from small business owners. When we started working with them, the biggest thing that stood out was not a lack of product knowledge from their agents. It was the moment an agent had to tell somebody no.
Sometimes the answer is no. The policy does not allow it, or the product does not do the thing a customer is asking for. What came through in the conversations was how much the delivery mattered. Same answer, very different reaction depending on whether the customer felt heard first.
So we built simulations around that specific skill. Their team calls it “reframing no”. Acknowledge the frustration, make sure the person knows they have been understood, then be honest about where things stand.
This insight helped their onboarding go from twelve weeks to eight, and some of that first cohort ended up performing as well as agents who’d been on the job a lot longer.
Simulation Can Span Personas
Q6: Solidroad says its simulations can span personas, channels, difficulty levels and languages. How do you make those simulations sufficiently realistic to prepare an agent for the unpredictability of a live customer interaction?
Mark Hughes: Two things, really. The first is that we build from real conversations rather than invented scenarios. The personas, the phrasing, the way somebody escalates, all of it comes out of what that company’s customers actually do.
The second is that the simulated customer behaves like a person. They ask follow-up questions, they get frustrated, they change direction, so an agent cannot get through it by following a script.
We also recreate the systems and the workflow around the conversation, because a lot of what makes a live interaction difficult is handling several things at once while staying present with the customer.
Rubrics Informed by Company Guidelines
Q7: Your platform uses customizable, self-improving rubrics informed by company guidelines, SOPs and knowledge bases. Who decides whether the rubric itself is evaluating the right things, and how do you prevent an AI system from reinforcing an organization’s existing blind spots?
Mark Hughes: That is exactly what this partnership is designed to help with. Any system built solely on a company’s own material has the same weakness, which is that it can only ever check you against what you already believed. If something is missing from that picture, nothing in the system is going to tell you.
Unwrap is reading what customers are actually saying, acting like a second set of eyes coming in from outside the company. Most of the time the two will agree, which is reassuring. The useful moments are when they don’t. If the internal numbers look strong and customers keep raising the same issue, that gap is the most valuable thing on the table, because it usually means the standard has drifted from what customers actually care about.”. We would rather treat that as a finding than explain it away.
The other part of the answer is that a real person still makes the final decision. The company owns what good looks like for their customers, their team signs off on it, and they change it when the evidence says they should. Our platform gives them better evidence to make that call with, but it does not make the call for them.
Human Agents and AI Agents
Q8: As AI agents increasingly participate directly in customer service, should human agents and AI agents ultimately be measured against the same customer-experience standards? Where should their evaluation frameworks differ, if at all?
Mark Hughes: The bar should be the same. A customer does not really care who or what handled their conversation, they care whether they got help and felt looked after. If a company holds the two to different standards, the experience gets inconsistent, and customers tend to notice that before the company does.
Where they differ is in how things go wrong. A person might have one conversation go sideways on a difficult day. An AI agent working from the wrong instruction can repeat the same mistake across a lot of conversations before anyone notices, so it needs checking more regularly rather than sampling now and then.
The fix is different too. With a person, you coach them. With an AI agent, you are usually changing an instruction, a policy, or the tooling around it. Same bar, different response when it gets missed.
A Prescribed Process
Q9: There is an important distinction between an agent following a prescribed process and actually resolving the customer’s underlying problem. How does Solidroad evaluate resolution quality rather than simply compliance with a script or SOP?
Mark Hughes: Process compliance is straightforward to measure, which is part of why so many teams lean on it. The limitation is that an agent can follow every step and the customer can still come away holding the same problem they started with.
So the more useful question is whether the issue actually got resolved. Did the customer understand the answer? Did they have to come back about it? And, did the agent pick up on what the person actually needed, which is not always the same as what they typed in the first message?
That is what we try to build into the training itself. If a scenario only rewards hitting the steps, you get agents who hit the steps. If it rewards getting the customer sorted, you get something closer to what the business actually wants.
Recurring Customer Problems
Q10: If customer intelligence identifies that a recurring customer problem is actually caused by a product, policy or process rather than agent behavior, how does your model prevent organizations from trying to “train away” a problem that needs to be fixed elsewhere?
Mark Hughes: Training is not the right tool for every problem, and we are fairly direct with customers about that. If the cause is a confusing refund policy or a checkout flow that is not working, coaching an agent to explain it more gracefully does not fix anything for the customer, and it puts the agent in a difficult position.
The tell is easy to spot once you’re looking for it. Agents are scoring well on the interaction, the customer is still unhappy, and the volume isn’t going down. That combination means the problem lives upstream of the conversation.
The pattern is usually easy to spot once you are looking for it. Agents are scoring well on the interaction, the customer is still unhappy, and the volume is not coming down. That combination points to something upstream of the conversation.
AI-powered Support Training
Q11: ŌURA is cited as an organization using both platforms across its member-support operation. What have you learned from real-world deployment that changed or challenged your assumptions about AI-powered support training?
Mark Hughes: A couple of things. The first is that spotting the problem is the easy part. Early on we assumed the win was visibility, showing a manager everything happening across their conversations. Visibility matters, but what teams kept telling us was that they wanted the insight to come with something they could run that week.
The second was where the improvements actually showed up. We expected the biggest impact in the high-volume scenarios, the situations everyone handles all day. In practice, a lot of it came from onboarding and from the messier, less common conversations that come up rarely enough that nobody gets much practice at them.
ŌURA is a good illustration of the model working the way it is supposed to. Unwrap is able to understand what is coming up across member conversations, and then Solidorad can turn that into training their team can actually practice.
Millions of Customer Conversations
Q12: Solidroad says it has scored millions of customer conversations. What metrics do you believe CX leaders should examine to determine whether AI-powered training is actually improving customer outcomes rather than simply generating more training activity?
Mark Hughes: Customer satisfaction is the one I would start with, because it is the closest thing you have to customers telling you directly if their experience got better.
After that, I would look at whether customers are coming back about the same issues you trained on, whether conversations are getting resolved, how quickly new hires reach the standard, and, what gets often overlooked, how consistent the team is. If your strongest agent and your average agent handle the same situation very differently, that is a problem no average is going to show you.
Activity measures like simulations completed or training hours logged are useful for knowing a program is running. I just would not treat them as evidence that a customer had a better experience. They are inputs rather than outcomes.
Business Performance
Q13: How should enterprises distinguish between improvements in agent proficiency, improvements in customer experience and improvements in business performance? Are there situations where one improves without the others?
Mark Hughes: They do come apart, and where they come apart usually tells you something useful.
If agents are measurably better and satisfaction has not moved, the issue is usually sitting in the product or the policy rather than in the conversation, so more training will not shift it. If the experience improves and the business numbers do not, check the volume, because the improvement has probably landed somewhere that does not come up often enough to matter
The thing underneath all of this is that most of the numbers a support team has, response time, ticket volume, resolution rate, tell you something real but they do not tell you whether the interaction was actually helpful. That was the gap Patrick and I kept running into before we started the company. So I would not read any one of these moving as proof on its own. If proficiency moves and nothing follows, I would not call that a training failure. I would treat it as a signal to go back and check whether you picked the right problem.
Analysis and Training
Q14: Customer-support AI raises questions around privacy, data handling and the use of customer conversations for analysis and training. What governance principles should enterprises establish before allowing customer interaction data to feed an AI training environment?
Mark Hughes: A few basics. Know what data is actually in scope and keep it to the minimum the work requires. Be explicit about whether customer conversations can be used to train general-purpose models. Make a real decision about how long you are keeping the data rather than defaulting to keeping everything. And be deliberate about who can see raw transcripts.
The part people forget is that this applies to your own team as well as your customers. Tell agents what is being looked at and why. If they find out after the fact that their conversations were being reviewed, you have created a trust problem that is hard to undo with a policy document.
Address the Risk
Q15: How do you address the risk that training data itself may contain inaccurate information, biased customer feedback or outdated policies? What controls are necessary before such data influences an agent’s training?
Mark Hughes: The main control is anchoring to current approved policy rather than to what agents have historically done. If you build training out of past behavior alone, you risk teaching expired policies and old habits along with the good practice. So the source of truth is the company’s own guidelines and knowledge base, that’s kept up to date, with someone who owns it.
On feedback, it is worth separating a customer being unhappy from a policy being wrong. Both are worth knowing about, and they call for different responses. A theme showing up because people do not like a policy that is working as intended is not a training gap.
And a real person should review a scenario before it goes live. That is something deliberate. It is a small amount of friction, and it stops one wrong assumption from being taught to a few hundred agents at once.
Successful Adoption of Solidroad
Q16: Many enterprises already have contact-center platforms, QA systems, knowledge bases, workforce-management tools and learning-management systems. What does successful adoption of Solidroad look like within that existing technology stack?
Mark Hughes: We are built to sit alongside what a company already has rather than replace it completely. We read from the conversation platform and the knowledge base so that training reflects real interactions and current policy, we work with the workforce tools teams already use to schedule and assign practice, and results can feed back into whatever system of record they use for learning.
As for what good adoption looks like, the ones that work tend to start narrow. One team, one problem the leader already cares about, and something to show inside the first month. Onboarding is usually the best place to start because the baseline is clear and you see movement quickly.
The version that struggles is the broad rollout where everybody gets access and nobody owns a specific outcome. One scenario that measurably improves will do more for adoption than a full deployment with no clear owner.
Improve Human Agents
Q17: Is the larger opportunity for Solidroad to improve human agents, evaluate AI agents, or create a common performance layer across both? How do you see that balance evolving over the next few years?
Mark Hughes: We’re aiming to create a common layer.
Most companies are running two separate systems today. Human agents get coaching and scorecards. AI agents are usually reviewed by whatever came built into the platform running them, which can mean the system handling the conversation is also the system reporting on how it went.
Where we think this ends up is a single standard that applies to both, so a support leader can ask one question, whether customers are getting the experience the company promised them, and get one answer regardless of who or what handled it.
That is the thinking behind wanting “Solidroad Certified” to mean something. Security certifications became a requirement once companies moved to the cloud, and there is a potential for something similar to happen with conversation quality as more of it moves to AI.
Biggest Adoption Barriers
Q18: There is often a gap between what an AI vendor can demonstrate in a controlled environment and what an enterprise can operationalize at scale. What are the biggest adoption barriers you encounter with CX and contact-center leaders?
Mark Hughes: The biggest one is that these leaders have been sold insight before. A lot of them have bought a platform that produced very good reports and did not change what happened on the floor, so they have learned to be careful. They are not skeptical of AI. They are skeptical of anything that arrives without a next step attached.
So the burden is on us to show the path from a finding to something an agent runs that week, on their own conversations and their own definition of good rather than in a demo environment. That is also why the partnership matters here. Unwrap surfaces the issue, we turn it into practice, and a leader can see the whole distance from problem to training rather than being handed a report and left to work out the rest.
First 90 Days
Q19: What should a CX leader measure during the first 90 days after deploying an AI-powered training program to determine whether it deserves broader enterprise adoption?
Mark Hughes: Take a baseline before you turn anything on. It is easy to skip when everyone is keen to get going, and it makes the conversation in month four much simpler when you are trying to work out what actually moved.
Then pick one group and keep it tight. How long it takes new hires to reach proficiency against whatever your historical number is. Whether customers are coming back on the scenarios you trained. How consistent that group is, meaning the distance between your strongest and weakest performer on the same situation.
The one I’d watch most closely is whether managers are actually using it. If the people who own the team aren’t opening it by week six, it won’t survive a wider rollout no matter how good the numbers look.
Ninety days is enough to see proficiency and consistency move. It’s usually not enough to see business impact, so I wouldn’t promise that in a pilot.
Human and AI Agents
Q20: Finally, if customer support moves toward a model in which human and AI agents continuously learn from customer interactions, what fundamental change do you expect in the role of the CX leader?
Mark Hughes: I think the job moves from managing capacity to owning a standard.
For a long time, the central problem for a CX leader was a math problem. Volume is going up, here is my headcount, how do I cover it. As AI takes on more of the volume, that problem gets smaller and a harder one takes its place, which is defining what good actually looks like for your brand and then showing that it is happening in every conversation, whoever handled it.
That is closer to running an operation than managing a team. You are designing a system, watching it for drift, and fixing it when it moves. The leaders already thinking that way are the ones I would back. The ones treating AI purely as a cost line risk hearing about their quality problems from customers before they see them in their own reporting.
CXQuest Analysis
Why This Interview Matters
The central issue is not simply whether AI can automate parts of customer support. It is whether organizations can create a continuous operational feedback loop in which customer interactions reveal problems, those problems inform training, and training subsequently improves how support interactions are handled.
The Solidroad-Unwrap partnership addresses that connection directly.
Customer intelligence and agent training have historically operated as distinct functions. A customer-insights platform may identify recurring complaints, while a training platform may prepare agents for known scenarios. Connecting the two potentially reduces the time between identifying an emerging customer problem and preparing agents to address it.
The significance becomes greater as organizations introduce AI agents alongside human agents. The same customer experience increasingly depends on interactions across both types of systems.
CX Perspective
From a customer-experience perspective, the relevant question is how effectively customer feedback reaches the people—or systems—responsible for resolving the underlying problem.
Analyzing conversations at scale can reveal recurring issues that may not become visible through manual quality assurance alone. Turning those insights into training introduces a mechanism for translating customer feedback into operational behavior.
The potential CX value therefore lies less in AI itself and more in the feedback loop it can enable: listen, identify, train, measure and refine.
The model also raises an important question for CX leaders: whether training scenarios remain aligned with what customers are actually experiencing rather than relying primarily on predetermined examples.
Business Perspective
For businesses, support training carries both operational and financial implications. Large support organizations must continually onboard personnel, maintain consistency and adapt to changes in products, policies and customer expectations.
If customer intelligence can automatically identify recurring support problems and training systems can rapidly convert those findings into practice scenarios, organizations may be able to reduce the delay between operational discovery and employee development.
The business case, however, ultimately depends on measurable outcomes. Those could include resolution performance, customer satisfaction, handling quality, training time, agent readiness and the effectiveness of AI-supported interactions.
The partnership therefore presents a testable operational proposition rather than simply an AI adoption story.
Technology Perspective
The technology architecture described in the release connects several capabilities: conversational analytics, customer intelligence, AI simulation, customizable evaluation rubrics and knowledge-base-informed training.
One important technical consideration is the quality of the information flowing through this chain. If conversation analysis identifies the wrong problem, training can reinforce the wrong behavior. Similarly, if an organization’s knowledge base or operating procedures are outdated, simulated training may reproduce those weaknesses.
The use of self-improving evaluation rubrics also raises questions around governance. Enterprise customers need visibility into how performance is assessed, what organizational policies inform the assessment and how changes to those standards are controlled.
For AI agents, the challenge is even broader because training and evaluation need to account for system behavior, escalation logic, policy compliance and consistency—not only conversational fluency.
Leadership Perspective
For CX and customer-support leaders, the emerging responsibility is increasingly one of orchestration.
Leaders must determine how customer intelligence, quality assurance, human-agent development and AI-agent governance work together. A technology platform can facilitate these processes, but the underlying standards remain organizational decisions.
The Solidroad model also highlights a broader leadership question: should human and AI agents be evaluated against fundamentally different expectations, or should they ultimately be held to the same customer-experience standards?
As hybrid support models become more common, defining those standards—and establishing credible mechanisms to measure them—will become increasingly important.
Industry Outlook
Customer support is moving toward operating models that combine human expertise with increasingly capable AI systems. This changes the role of training from a primarily onboarding function into a potentially continuous performance-management process.
At the same time, customer conversations are becoming an increasingly important source of operational intelligence. The opportunity is to shorten the distance between what customers report, what organizations learn and what support agents subsequently do.
The longer-term direction will depend on several factors: the reliability of conversation analytics, the quality of organizational knowledge bases, the transparency of AI evaluation, integration with existing contact-center systems and the ability to demonstrate measurable improvements in customer and business outcomes.
The Solidroad-Unwrap partnership provides one example of this emerging convergence between customer intelligence and AI-enabled workforce development.
Key Takeaways
Strategic Implications
1. Customer intelligence becomes more valuable when it can influence operational behavior. Identifying recurring customer problems is only one part of the CX improvement cycle.
2. Training can become a continuous process rather than a one-time onboarding activity. Emerging customer issues can potentially be converted into new scenarios and learning requirements.
3. Human and AI agents create a common quality challenge. Organizations need clear standards for evaluating both types of support interaction.
4. The quality of the data-to-training pipeline matters. Poor conversation analysis, outdated knowledge or inappropriate evaluation criteria can undermine the resulting training.
5. AI adoption requires governance as well as technology. Organizations need clarity around evaluation criteria, operating procedures, knowledge sources and accountability.
6. Customer experience and workforce development are becoming increasingly interconnected. Insights from customer interactions can inform not only product and process changes but also agent preparation.
7. Enterprise value ultimately requires measurable outcomes. Training effectiveness should be assessed against relevant customer, operational and business metrics rather than AI adoption alone.
8. The emerging CX model is increasingly a feedback loop. Customer interactions generate intelligence; intelligence informs training; training influences subsequent interactions; and those interactions generate new data for refinement.
