By Antanas Bakšys, CEO & Co-Founder of Ace Waves

For more than 2 years, the internet has shouted that AI is replacing humans in customer support. We’ve been promised that AI would reinvent customer support with instant resolutions, lower costs, and happier customers.

And yet, for years, the reality didn’t match the headlines. Have you ever interacted with a real AI agent as a consumer? Probably not yet.

We’re still forced to talk to dumb chatbots that, at best, pull up a knowledge base article. But I’m not looking for a search engine, I’m looking for someone who can actually solve my problem.

Over email, we wait days for a reply that usually doesn’t resolve anything at first, just asks more questions. By the time the issue is finally fixed, it’s no longer relevant. And on the phone, we sit on hold forever or even pay extra by the minute.

Customer support is still painful - for both consumers and businesses.

The experience we all expect now

Imagine you’ve ordered a case of wine for the weekend. It’s Thursday, you’re having lunch, and it still hasn’t arrived. Friday night is coming, friends are already invited, and you’re starting to worry.

While finishing lunch, you scroll through your inbox trying to find an update - nothing. You reply to the store’s last email asking for an ETA, grab a coffee, and hope someone responds by tomorrow. If it’s not delivered on time, you’ll need to leave work early to buy wine from a store in the city.

Instead, your phone buzzes almost instantly:

“Hi, this is Jamie from WineStore. I checked with DHL, and your delivery is planned for tomorrow between 3 PM and 5 PM. You can track & manage your delivery directly at DHL here. I'm more than happy to help you anytime. Enjoy your Friday night with your loved ones!”

No wait.

No “we’ll get back to you.”

That’s the experience we all expect now: instant, correct, any hour, any language, in any channel. And increasingly, that “Jamie” isn’t human at all.

The 85% reality

In many consumer companies, around 85% of customer support is repetitive. Issues like:  “Where’s my order?”, “I need a refund”, “My login doesn’t work”. Simple tickets. Simple logic. Yet they consume huge human effort.

For support teams, this work is monotonous and dull. For businesses, it’s costly. So costly that some brands choose to hide contact options entirely.

If AI can handle that repetitive 85% accurately, 2 things happen:

• Customers get their issues solved instantly and correctly.
• Businesses cut costs significantly without cutting quality.

The promise is huge. The headlines have been loud already for a while. But when is it actually coming?

Let’s look at what’s happened since ChatGPT’s launch and where things are headed next.

2023: The “should work” year that didn’t

After ChatGPT launched in late 2022, the world sprinted toward “AI in support.”

In 2023, rumors spread fast. Some claimed AI had already replaced customer support. Others insisted it failed miserably. Companies with high support costs rushed to build internal solutions, hiring engineers and data scientists, convinced they could make it work.

Most discovered the same wall: LLMs are great at language and writing replies, but not at following processes. We saw hallucinations delivered with confidence, weak tool-use, brittle intent trees, and guardrails that needed constant babysitting.

Everyone expected that the next model would finally fix it. But when March brought GPT-4, there was no real difference.

Later that year, enthusiastic teams kept pushing: more configs, more fine-tuning, more money. The result? Still not there. AI couldn’t yet handle real support.

In our own AI lab, V1 was not really usable - it was constantly hallucinating, which meant giving wrong answers and performing incorrect actions. None of us trusted it to work with real users.

V2 was better, but still very limited. It looked more like a decent copilot for human agents, not the autonomous worker we were after. Our goal was to make AI do the repetitive work that humans do now.

2024: Is it really working?

By early 2024, market skepticism and disappointment had already set in. Companies quietly shelved their AI experiments. Incumbent helpdesks like Zendesk and Intercom rebranded their old bots as “AI agents”, but behind the curtain, it was still intent-based decision trees, just with better marketing.

Then came the Klarna case.

In April 2024, Klarna and OpenAI announced that AI had replaced 700 full-time support agents - handling 2/3rds of their tickets without lowering customer satisfaction, saving $40 million a year (!!!).

You see, soon after ChatGPT's release, Klarna’s CEO and co-founder, Sebastian Siemiatkowski, jumped on a plane to San Francisco to convince Sam Altman and OpenAI to work with Klarna as a “guinea pig” to test and use all their AI products. And they actually built something.

The industry froze. Fear of missing out spread across the market.

Was it real? Was it hype? Is this actually possible? How? Do you need a special deal with OpenAI to access the magic? Or are we simply missing something? If Klarna pulled this off, why can’t Zendesk or Intercom deliver the same? Why can’t we?

For most, it was a headline. For us, it was a signal.

That announcement proved one thing: AI could finally do the work - not just generate text.

2025: Our own “it finally works” moment

Sometime in 2024, after months of trial and error, we found ourselves hesitating. Should we go back to traditional decision-tree logic - those predictable “if-then” systems that react to keywords and follow preprogrammed rules? They would’ve worked, but they wouldn’t have been AI. Those setups are limited, rigid, and demand constant maintenance.

We stayed the course. Klarna’s case gave us optimism but also doubt - were we missing something fundamental? Were we building the wrong way?

When we began working on the idea that later became Ace Waves, our goal was simple: to build AI agents that could handle customer support as reliably as humans - but at scale, in any language, across every channel. We kept iterating, pushing our AI lab to full capacity, until one version finally broke through.

In early 2025, while working with one of our patient design partners, one of our deployed AI agents hit a milestone. It followed every process flawlessly. It was able to do real actions in systems and use real tools. It didn’t hallucinate. It handled thousands of customer issues per day with full autonomy.

That was the game-changing “it finally works” moment - the point where we realized we weren’t delusional. We’d fu**ing proved it!

Of course, it was just one company and one AI agent. The road ahead was still long, but it gave us the confidence and knowledge to keep building.

For years, we were too early. The tech wasn’t ready. Now it finally is.

The inflection point

The gap between “talks like a human” and “works like your best agent” just recently quietly closed. Not because of one big breakthrough, but because dozens of small ones finally clicked together: new LLMs, multi-agent systems, orchestration principles, reasoning frameworks, improved memory, grounding, context control, etc.

What emerged isn’t another chatbot. It’s a new category: agentic AI - systems that follow procedures, understand context, perform real actions in existing tech stack, and operate within strict guardrails.

This is what I mean when I say “inflection point.” It’s that quiet shift when experiments stop being prototypes and start being infrastructure. When the conversation moves from “Can AI do support?” to “How much support can AI handle reliably?”

Where AI agents shine

Ace Waves AI agents now outperform humans in repetitive, well-defined support tasks such as:

• Order tracking and delivery updates
• Refunds, returns, and modifications
• Buyer–seller mediation in marketplaces
• Product warranty coordination and services ordering
• Subscription changes
• Level-1 payment disputes
• Informational requests
• Product help
• etc.

They can operate inside existing systems like Zendesk, Intercom, Kustomer, or custom helpdesks, across all channels: email, chat, and voice. They speak 90+ languages fluently and work 24/7.

AI agents can do these tasks faster, cheaper, and more consistently than humans.

It’s not about replacing people. It’s about removing the repetition from their day.

Humans are still needed in high-stakes, ambiguous, or policy-gray areas. Here, AI helps by gathering context, checking documents, and preparing a clean handover so your best people can focus on decisions, not fetches.

In other words, AI takes the weight; humans take the lead.

What this means for businesses (a.k.a. why CFOs smile)

Every company that handles hundreds of thousands of support tickets sits on the edge of a new curve. For the first time, the economics of customer support are rewriting themselves in real time.

The same ticket that once cost several dollars to resolve now costs dramatically less, handled in seconds with a higher SLA and consistency. CSAT doesn’t have to fall - in many deployments, it rises, because instant, accurate, and on-brand resolutions beat long waiting queues every time.

And this isn’t a future scenario. It’s starting to happen in production right now, and our clients like Eneba are already enjoying the advantage they have over the competition.

The next wave

For years, “AI in customer support” was a pitch. Today, it's an infrastructure. The technology, the data, and the know-how have aligned. The inflection point isn’t theoretical - it’s operational.

In the coming year, every consumer brand will face a new standard: instant, multilingual, 24/7 support that is indistinguishable from humans, powered by systems that never sleep.

If you’re leading CX, Ops, or Finance, open your P&L and mark the support line. That line won’t look the same 12 months from now.

The inflection point is here.

The question isn’t if you’ll deploy AI agents, it’s when you’ll start. Don’t wait too long.