AI for customer support that keeps the human touch: 5 fixes from our webinar with Tidio

There’s a specific kind of AI support problem that’s hard to catch from the inside. Resolution rates are up, ticket volume is down, and everything looks like it’s working. Meanwhile, a portion of customers who had a frustrating experience with the bot never said anything about it. They cancelled, churned, or switched providers without a complaint. Research puts that share at 43%. The standard support dashboards don’t catch this kind of failure. By the time churn reflects it, the problem has been running for weeks.

‍

That number is what pushed EverHelp and Tidio to run this session together. Tidio brings 14 years of deployment data and direct visibility into how AI performs at scale. EverHelp brings the operations side, running hybrid AI and human support for SaaS and eCommerce clients day to day. EverHelp was founded in 2021 and has grown quickly into a recognized industry leader, most recently earning the 2026 Outsource Partner of the Year award and a place on the Clutch 100 list of Fastest-Growing Companies. Recognition that reflects what our team builds every day in client operations.

‍

The gap between those two perspectives is exactly where most implementations break down.

‍

Marius Laza, Tidio’s Chief Customer Officer, and Valentyna Dimova, EverHelp’s VP of Customer Support, spent an hour discussing where this happens and what it takes to fix it. Five steps, built on real deployment data, covering knowledge base setup, escalation design, and the feedback loop most teams quietly drop once the AI goes live.

‍

Suggested read: How can AI improve customer service? A case-based outlook

‍

The 5 fixes for better AI customer support

‍

Each fix builds on the one before it, and together they form a complete picture of what a well-functioning AI support setup actually looks like in practice.

Fix 1: Clean and segment your data first

"You have to actually treat your AI agent as a part of your team. The way to go is just to treat your AI as a human, as you would train your customer support representative."

Valentyna Dimova, VP of Customer Support, EverHelp

‍

No configuration matters if the foundation is wrong. Both Valentyna and Marius were aligned on this point from the start: before anything else, the data needs to be audited, cleaned, and segmented by audience type. A free trial user and a VIP client asking about the same billing issue may need entirely different responses, and the AI needs that structure built in from day one.

‍

One of the more surprising data points Marius shared: once teams begin training their AI, 70% of the answers it surfaces come from historical conversation data, not FAQs. Only 30% comes from the FAQ section itself. That changes how teams should approach knowledge base preparation. FAQs capture what you think customers ask. Old tickets capture what they actually say, including how they phrase frustration and which edge cases come up over and over again.

‍

As our hybrid AI support model framework makes clear, the quality ceiling of any AI setup is ultimately determined by what you feed it from the start. Clean, well-structured data is the prerequisite for everything that follows.

Fix 2: Integrate deeply with the right systems

"First it's just basically the AI is a glorified text processor. You just give it the knowledge base and other text and it makes it shorter and faster for people to see. But then integrations, you would quickly learn that people ask questions about specific things."

Marius, Chief Customer Officer, Tidio

‍

A knowledge base alone isn't enough. The AI can read documents and surface faster answers, which is useful but limited. The moment a customer asks about their specific order, their last payment, or a promotion that ended last week, a static text layer breaks down.

‍

Integrations are what close that gap. Connecting the AI to live systems (shipping dates, invoice details, refund status, product comparisons) immediately expands what it can confidently resolve. Marius noted that in most cases, the integration doesn't need to be invasive. Read-only access for surfacing shipping times or payment history is typically low-risk and quick to set up. This matters particularly for ecommerce support, where customers routinely ask about order-specific details that a knowledge base alone can never answer.

‍

From EverHelp's operational experience, inconsistency is one of the most trust-eroding things a customer can encounter. When an AI can pull accurate, current data rather than relying on static knowledge, responses become more reliable and customers start to trust them. Our guide to customer support automation goes deeper on exactly where that line gets drawn in practice.

Fix 3: Design clear escalation flows  and don't push back twice

"I don't think you should do a pushback. We have a whole section about pushbacks against this. We do recommend escalating to a human."

Marius, Chief Customer Officer, Tidio

‍

This fix generated the most discussion in the session, and for good reason. The scenario is familiar: a customer is frustrated, asks for a human, and the AI pushes back. It offers another FAQ link, asks them to rephrase, and redirects to a help article. Now the customer has two problems instead of one.

‍

Here's the data point that reframes everything: Marius shared that in Tidio's user base, 60% of customers who ask to be transferred to a human then follow up with a question the AI could have easily handled. A shipping query. A pricing question. Something routine. The escalation request wasn't about complexity. It was about comfort, frustration, or habit.

‍

Fighting that request is almost always the wrong call. The better approach is to make the handoff feel natural and immediate, and to configure the AI to detect frustration signals before the request comes, offering escalation proactively rather than waiting for the customer to demand it. FFor a closer look at what that architecture looks like in a working setup, our Human+AI customer support guide covers the structural detail.

‍

{{cta}}

Fix 4: Let support agents coach the AI

"Your support agents already have the knowledge. They know exactly how to do the job. The way we pitch Lyra is that it's another support agent. It's just another hire that you have, maybe even a junior hire at the beginning."

Marius, Chief Customer Officer, Tidio

‍

Agents can supply natural language examples that teach the AI to match the brand's tone, not through rigid rules, but through the kind of real conversational patterns the team already uses every day. One of the fastest ways to erode customer confidence in an AI customer service agent is responses that sound robotic or generic. Brand voice training, guided by agents, is the fix.

‍

Marius confirmed this mirrors how Tidio approached the rollout of Lyra. Support agents moved into something closer to a team lead role for the AI: reviewing flagged conversations, confirming correct answers, catching gaps, and feeding corrections back into the system. When agents have a genuine stake in the AI's quality, the dynamic shifts from "AI is replacing us" to "we're training our new team member."

‍

This ongoing coaching relationship is also what drives long-term improvement in the metrics that actually matter. Our breakdown of AI agent KPIs explores why some of the most commonly tracked numbers give a misleading picture of real performance, and why the feedback loop is what separates a setup that plateaus from one that keeps improving.

Fix 5: Build a feedback loop  and keep humans in it

"Everything changes. Your business can change as well. The processes, the policies, the value you want to address to your customers, that changes. So that has to change your AI agent as well."

Valentyna Dimova, VP of Customer Support, EverHelp

‍

The fifth fix ties all the others together, and it's the one most teams quietly skip once the initial setup goes live.

‍

An attendee named Ferdinand raised the point directly during the session: AI-powered knowledge bases need regular review. Trusting answers pulled from old ticket data without ongoing verification is a real risk. Marius agreed and expanded on it: an answer that was accurate for one client six months ago might be wrong for a different client today. Policies change. Refund windows change. Shipping terms change. And if nobody catches the discrepancy, the AI keeps giving customers the wrong information with full confidence.

‍

Tidio's approach is to surface these gaps in a review dashboard, showing support agents both the answers the AI is giving and the questions it can't answer, then prompting them to confirm, correct, or add information. The review process doesn't need to be heavy. It does need to be consistent.

‍

What realistic resolution rates actually look like

‍

Marius walked through the numbers honestly, and they're more nuanced than most vendor messaging suggests:

‍

• A strong, well-configured setup typically reaches around 80% resolution
• Simple, FAQ-heavy use cases can reach 90 to 94%
• 100% automation is neither realistic nor advisable. Tidio's highest on record was 94%, for a business with highly predictable questions
• One of Tidio's largest clients started with a target of 3% resolution. Not a failure, but a deliberate choice to automate one specific, time-consuming ticket type at high volume

The 3% case is worth sitting with

That last example deserves more than a passing mention. A company handling hundreds of thousands of interactions decided that automating 3% of them, a specific, repeatable case that consistently drained agent time, was exactly the right place to start. That 3% still represented tens of thousands of resolved tickets. Starting narrow and high-impact is a legitimate strategy, and often a smarter one than chasing a headline automation rate.

How AI audits map the real baseline

Tidio uses AI audits to help teams understand their realistic starting point before committing to a deployment. The process involves exporting existing conversations, analyzing them against AI capability, and breaking down what's actually solvable. A recent example Marius shared:

‍

• 38% of a client's questions could be answered using native knowledge alone
• +29% became solvable once key systems were connected
• Deeper configuration covered most of the remaining volume
• Around 12% remained as genuine escalations

‍

That kind of breakdown gives teams something concrete to plan against, rather than guessing at what automation can realistically achieve.

‍

Where to start: the safest entry points

‍

Both speakers were consistent on this. The right place to begin is with the simplest, most repetitive questions your team handles:

‍

• FAQs and basic account queries. Low-risk, high-volume, and immediately demonstrates value to both the team and the business.
• Order status and shipping questions. Dynamic but predictable; easy wins once integrations are in place.
• Single-reply email support. Marius shared a constraint that works particularly well: only let the AI handle emails it can resolve in one reply. If a question needs back-and-forth or clarification, the AI leaves it untouched entirely. That boundary prevents the AI from starting conversations it can't finish cleanly.
• Repetitive queries that don't require human judgment. Any question your agents answer the same way, every time, is a candidate for automation.

‍

Even a tightly scoped setup like this delivers the core value of AI for customer support: instant responses, 24/7 availability, and meaningful coverage for teams limited by time zones or working hours. Getting those wins consistently is a better foundation than chasing a high resolution rate too early. Our overview of AI chatbots for customer service covers what different setups actually look like at the implementation level.

‍

Suggested read: Pros and cons of AI in customer service

‍

The takeaway worth keeping

‍

The best framing from the session came from Marius at the very start: we're fighting 20 years of bad automation experiences. Every customer who typed "can I talk to a human?" and got stonewalled is carrying that memory into their next interaction with your AI.

‍

The way out isn't better technology on its own. It's a proper setup: clean data, deep integrations, natural escalation flows, agents who coach the AI, and a feedback loop that keeps it honest over time. If you want to stay current on how the space is evolving, our contact center AI news roundup covers the latest developments worth watching.

‍

Valentyna put it plainly: treat the AI like part of the team, not a button you switch on. The teams doing that are the ones building support operations that customers actually trust.

‍

If you missed the live session, watch the full webinar here. And if you're thinking through what this looks like for your own operation, we're happy to dig in. 

‍

{{cta}}