Why Complex CX Is the Real Proving Ground for Agentic AI

What if the impressive agentic AI demos fail in real use? Gartner says over 40% of agentic AI projects may be canceled by 2027. This is largely due to poor returns and integration issues.

This gap often does not come up in demos.

Basic requests like checking order status, resetting passwords, and asking about returns show where agentic AI in customer service seems simple.

However, it struggles with complex tasks.

A 2022 Klarna study found that the company replaced 700 customer-service jobs with AI. However, within two years, it had to rehire staff because quality dropped on complex questions.

This change is an important story to consider.

Ambiguous requests and switching channels during a problem are common.

Handoffs that lose important context are where customer trust breaks down.

Complex CX automation must show its value in those key moments, not just in simple situations.

Deflection rates do not reflect this.

What matters is resolution.

Quick Answer: Complex customer experience (CX) is where agentic AI shows its strengths. Gartner predicts that over 40% of these projects could fail by 2027 because of poor ROI and integration challenges. The challenge lies in handling nuanced, context-dependent queries effectively, as shown by Klarna's reversal after initially replacing 700 customer service roles. Success in complex CX automation is paramount for building and maintaining customer trust.

The development: agentic AI is moving closer to complex customer work

What if your best deflection rates are hiding important conversations that decide if customers stay?

This question affects how agentic AI in customer experience develops as we move into 2026.

The technology has moved beyond answering simple FAQs and is now handling multi-step tasks.

In January 2026, Microsoft launched Copilot Checkout, allowing shoppers to complete purchases without leaving the platform. This feature works in partnership with PayPal, Shopify, and Stripe.

Industry data suggests that the numbers being reported are too optimistic. Gartner projects that agentic AI will resolve 80% of common customer service issues by 2029.

The discussion becomes more focused here.

The word "common" is carrying enormous weight in that forecast.

Password resets and shipping updates have never been the tough problems.

Complex CX automation is different. It involves reading unclear emails, getting order history from various systems, deciding if a policy exception is needed, and knowing when to pass an issue to someone else.

Klarna shows the gap in practice.

The fintech replaced 700 customer-service roles with AI in 2022, then rehired staff within two years after agents fell short on complex questions and service quality.

This timing is uncomfortable for support leaders.

Budgets for the next round of agentic AI projects are being approved even though the first round is still testing its efficiency on simple tasks.

The real test isn't how many tickets an agent resolves.

It's about whether it can handle unclear situations, transitions, and rules.

Why complex CX is the real proving ground for agentic AI

Basic automation can provide answers.

Complex CX must determine what to do next.

This is where agentic AI in customer service builds or loses trust.

A bot that reads a return policy may seem competent until the customer's package is late, the refund has been processed, and they are still waiting for a replacement.

Then the system must check multiple records, figure out what caused the problem, and decide what action to take next.

Ambiguity is the first challenge.

Real customers write “this is the second time” without stating what broke.

Then there’s continuity.

Generating a plausible answer is cheap in 2026.

Holding context across a chat that started Tuesday, moved to email overnight, and resumed on WhatsApp is not.

Gladly’s analysis of agentic AI customer service makes the same point: the system has to know who the customer is and what happened last time before it acts.

Klarna showed what happens without that.

The company replaced 700 customer-service roles with AI in 2022, then rehired staff within two years once complex questions exposed gaps in judgment and service quality.

The four tests that separate useful agents from impressive demos

CX challenge	What a basic bot does	What an agentic system must do	Operational evidence to watch
Ambiguous intent	Matches keywords to a canned article	Asks one clarifying question, then reasons from account state	Share of vague chats resolved without escalation
Multiple knowledge sources	Searches a single help-center index	Combines help docs, order data, and past tickets	Citation accuracy in agent replies
Routing and handoffs	Drops the ticket into a general queue	Passes full context to the right team or person	Repeat-contact rate within seven days
Policy or permission limits	Refunds anything, or nothing at all	Applies the real rule, explains it, stops when required	Exceptions logged for human review

What the shift means for CX and support operations

Please provide the section text to revise.

The source content did not come through.

Agentic AI should resolve issues, route them correctly, and learn from every interaction.

Agentic AI should resolve issues, route them correctly, and learn from every interaction.

A customer asks about a delayed order in chat, then follows up by email two days later. It is the same order and the same frustration.

Yet, it shows up as two separate tickets with no shared memory.

That disconnect is the main problem that agentic AI needs to fix.

The shift is simple: customer questions aren’t isolated tickets.

They’re operational signals.

When five customers in one week ask why a refund took twelve days, it is not five support issues. It is one process problem appearing across channels.

If you treat chat, email, and your knowledge base as separate silos, you lose that signal.

Why silos break complex CX. The customer doesn’t experience channels.

They experience one company that should already know their history.

Gladly’s 2026 perspective on knowing customer context before acting frames it directly: the question isn’t how much gets automated, but whether the AI actually knows who it’s talking to.

That memory has to span every touchpoint, or it isn’t memory at all.

What responsible automation looks like. Resolution alone isn’t enough.

In complex CX, responsible automation means three connected jobs:

• Resolve what’s genuinely answerable, and stop when policy or permission says stop
• Route what isn’t, with full context attached
• Learn from every interaction by feeding recurring questions back into the knowledge base

Skip routing and you get Klarna’s 2022 outcome: it replaced 700 customer-service roles with AI, then had to rehire within two years because quality slipped on complex cases.

Skip learning and the AI just gets busier.

Here at AnswerRidge, that resolve-route-learn loop across chat, email, and your knowledge base is the whole design.

Near-term watch points for support leaders

Containment rates make neat headlines.

They often do not show whether a system performed well when a customer changed their mind during a conversation.

As agentic AI in customer experience shifts from pilot to production, four signs will indicate which deployments are truly designed for complex work.

Published evidence beyond containment. Ask vendors for resolution accuracy, repeat-contact rates, and escalation patterns — not just the share of chats closed without a human. Gladly's 2026 guide to agentic customer service makes the point bluntly: the industry has fixed on how much it can automate rather than whether the AI knows the customer at all.

Behavior under uncertainty. Watch how an agent responds when it cannot verify a policy, a refund window, or an order status.

Does it guess, or stop and hand off with context intact? Research from Klarna shows that the company's reversal after replacing 700 service roles in 2022 illustrated the consequences when systems push past their confidence limits.

Knowledge that compounds. A knowledge base should sharpen with every real interaction — new exceptions logged, stale articles flagged, recurring gaps surfaced for someone to fix.

If your article list looks the same in six months, the learning loop is not running.

Measurement that reflects total effort. Track resolution accuracy, routing quality, and end-to-end handling time together, not in isolation.

Gartner expects more than 40% of agentic AI projects to be scrapped by 2027 for weak ROI and integration problems, and single-metric dashboards are part of how that happens.

The coming year rewards leaders who ask harder questions before signing, not after.

Containment shows what left the queue.

What are the limitations of agentic AI in customer support?

Agentic AI struggles with complex customer scenarios that require nuanced understanding and contextual continuity. It excels at handling simple requests but often fails when faced with ambiguous questions or the need for multiple-step resolutions, leading to inadequate customer support.

Why is agentic AI failing to handle complex customer issues?

Agentic AI fails with complex customer issues primarily due to its inability to maintain context over multi-turn interactions and manage nuanced, context-specific queries. This results in lost customer trust when the AI can't effectively resolve unique or ambiguous inquiries.

What is one major risk of agentic AI in CX?

A major risk of agentic AI in customer experience is its tendency to lose context during handoffs or when channel switching, which can lead to a breakdown in customer trust. If the AI cannot remember previous interactions, it diminishes the quality of support.

Will AI replace the help desk in 2026?

AI is unlikely to fully replace the help desk by 2026, although it may take on more tasks. Successful integration will depend on the ability of AI to handle complex customer issues effectively, which remains a significant challenge.

Why may 40% of agentic AI projects be canceled by 2027?

Gartner predicts that over 40% of agentic AI projects could be canceled by 2027 due to weak return on investment and integration challenges. Many projects fail to translate impressive demos into reliable, measurable outcomes in real-world scenarios.

According to Gartner, the projection that more than 40% of agentic AI projects will be scrapped is not a verdict on the technology.

It is a warning about how the work is sold.

Demos reward polish, speed, and the appearance of autonomy.

Production rewards reliability, accountability, and measurable outcomes.

The gap between those two tests is where most projects fail.

That gap is not closed by better prompts or a louder launch.

It is closed by treating agents as systems: defining failure, logging decisions, testing edge cases, and measuring value after deployment.

The demos never take that test.

The organizations that do will not need to fear Gartner's number.

They will have already passed it.