What Happens When AI Starts Making the Decisions?

Front page
GuideCX AI33m · 17:27 GMT · 27 min read

Agentic CX in 2027: What Happens When AI Starts Making the Decisions?

The age of agentic CX is officially here; getting the technology is simple enough, making sure it’s properly implemented, adopted and governed, is what today’s leaders are struggling with most. Here’s how you can prepare to safely take part in the agentic revolution.

The trouble with agentic CX right now is that a lot of business leaders are getting so swept up in the promise of faster resolutions, fewer calls, and lower costs that they’re missing how complicated it really is to get this contact center upgrade done right.

Imagine an AI agent approving a refund. Finance’s system rejects it, but the customer still gets a cheerful confirmation email. Meanwhile, the CRM records another successfully resolved complaint, and for a while, everyone’s happy; until the complaints roll in.

There’s plenty of proof that agentic CX really can deliver some incredible results, and we’ve covered some amazing customer stories from companies like Simplyhealth and Thames Valley Police.

What I think companies need to focus more on, in those stories, however, is that success didn’t just come from a model being great. It came from strategies built to address the various things agentic AI in customer service can still get wrong.

As Jochem van der Veer, CEO of TheyDo, told us in September 2026,

“Faster and more confident doesn’t mean better.”

This is our guide to the shift itself: how much autonomy to allow, who’s accountable, what to log, and how the human role changes. If you’re shopping for the AI agent platform, or the contact center underneath it, we have separate buyer’s guides for those.

Agentic CX: Key Takeaways for Business Leaders

Agentic CX gives AI agents permission to act on customer requests across business systems; it’s up to you to decide how far that permission should go.

• Different decisions deserve different levels of autonomy. Checking a delivery date is one thing; changing someone’s payment arrangements is another matter entirely.

• Demand proof of resolution. A conversation ending successfully tells you very little about whether the customer’s problem was actually fixed.

• Put someone in charge. A proper agentic CX operating model needs a named owner who sets boundaries and has the authority to suspend an agent.

• Protect human intervention. Escalating a vulnerable customer or complicated dispute is often the smartest decision an AI agent can make.

• Watch what happens after launch. Agentic AI in customer service needs ongoing scrutiny, reliable records of decisions, and evidence that customers are getting better results.

What Is Agentic CX?

Agentic CX, or agentic customer experience, is pretty much the term we’re using to describe a customer-facing setup where autonomous agents can work through customer requests and take permitted actions using connected business systems. Our guide to AI agents in the contact center covers more detail, but for this guide, the main thing to focus on is that agentic equals action, and action equals consequences.

Defining agentic CX can sometimes get complicated, because it’s become something of a “hot term” in the industry, the same way generative or conversational AI were a while back.

So if you’re getting lost in the weeds, here’s a quick look at the current CX automation ladder:

	Points the customer toward the cancellation policy
	Explains possible reasons for the charge
	Shows an employee the account history and recommends a correction
	Checks eligibility and submits an authorized refund
	Coordinates the correction across billing, CRM, and customer communications

How Does Agentic CX Actually Work?

Our guide to AI agent architecture walks through the wiring, and you probably have a good idea of how LLMs, natural language tools, and data analysis come together to help an AI agent understand a request. Trouble is, that’s the easiest part.

Imagine we built a subscription agent for CX Today, and someone wanted to stop a paid subscription. That agent would need to establish when cancellation occurred, check the billing rules, and determine whether a refund is owed. It then needs permission to make the adjustment. Suppose the payment provider rejects it, then the agent has to recognize that failure; otherwise, the customer gets a lovely confirmation email and another reason to complain.

As Kathy Ross, VP analyst at Gartner, put it very astutely in a recent interview with us:

“AI agents are tools. They’re very powerful tools, but they’re not employees, they’re not teammates, and they have to be managed like technology.”

That’s particularly relevant when agents work across sales, service, and finance. Different systems frequently hold different pieces of the customer’s history, and someone has to make sure the AI is acting on the right information.

What Are the Five Levels of AI Autonomy?

There’s quite a difference between asking AI for advice and letting it change customer records, which is why the levels of autonomy in a contact center need to be considered very carefully.

	Retrieve the customer’s billing history
	Suggest a refund based on company policy
	Prepare the refund for an employee to authorize
	Issue eligible refunds under agreed restrictions
	Complete permitted tasks across several systems, checking the result

The important thing about this ladder is it describes the increasing authority companies give to AI agents, which is very different from describing the maturity of a system.

A bank might have seriously sophisticated AI and still insist that people approve certain account changes, and it probably should.

As Mick Costigan, VP of Salesforce Futures told us:

“It’s way beyond just trust. It’s access to data, everything around context engineering…”

Run any workflow up the ladder, and the complexity and stakes climb with it: retrieving tracking information is level one, while automatically arranging a replacement shipment (and verifying the warehouse actually accepted the order) is level five territory.

Where Is Agentic CX Delivering Real Value in 2026, and Where Is It Still Hype?

There are plenty of success stories, no doubt about that, but it’d be dangerous to suggest we’ve reached a state of agentic CX nirvana at this point. Even experts from leading companies like Salesforce are still asking questions:

“How good are agent capabilities going to get, number one. Number two, how do I bring them into my organization, and what does it mean in terms of changes for how the organization works? And then point three: what do humans do?”

Where Is Agentic CX Actually Working?

Scroll through our news stories, and you’ll find plenty of examples. In August 2026, we talked about how Cisco used agentic AI to successfully handle 145,000 support cases, and CEO Chuck Robbins even told us the technology was driving higher renewal rates.

Dan Schulman, the CEO of Verizon, said that the three months the company spent experimenting with AI agents to replace certain customer service reps revealed that customer satisfaction rates actually went up thanks to AI involvement.

Then there’s the Thames Valley Police’s Bobbi agent (we followed its deployment throughout 2026).

Bobbi handles non-emergency contact for Thames Valley Police and Hampshire and Isle of Wight Constabulary. Superintendent Rob Brind described it as the first agentic AI deployment in UK policing, possibly globally.

In June 2026, Bobbi was managing around 14,000 citizen contacts a year, handling 200 conversations a day with roughly 45% fully automated, and freeing up more than 3,200 operational hours. By October 2026, the forces told CX Today that automation covered 75 to 76% of the contacts coming into Bobbi. That jump didn’t happen by flicking a switch, though.

“We spent three to four weeks implementing it, and then the rest of it was all testing and checking and putting guardrails in”

That tells you a lot. Success is reachable, but it’s not automatic.

Openreach offers another helpful point. Its proactive AI program expanded to 15 million customer journeys, reducing missed appointments and inbound contact volumes by roughly a third, and a separate A/B-tested generative AI chatbot trial reduced installation cancellations by 22%.

Chris Herbert, Openreach’s Director of Customer Service, admitted in a May 2026 interview:

“We could have stopped when it was a bit hard.”

The team kept going because customer feedback justified the effort.

What Does the Latest Research Actually Show?

From what we can tell, adoption is increasing, but that doesn’t mean we’re anywhere close to full maturity.

	41% of customer-facing AI deployments had scaled fully, covering broader AI applications
	Only 24% of surveyed consumers reported their issue was completely resolved by AI without human assistance
	24% of surveyed UK contact centers claimed to use agentic AI, but only around 2% of all centers had bots capable of independently determining steps across multiple systems

More and more businesses are adopting the agentic CX label before they’ve built the demanding capabilities behind it. The technology’s ready. The preparation is still catching up.

Why Do Agentic CX Projects Still Struggle?

For the most part, in places where companies still haven’t done the groundwork.

ServiceNow covered a few of the major issues in October 2026, saying that only 5% of surveyed executives had redesigned controls and workflows for autonomous execution, and 74% were still struggling with poor data accuracy, access, or management.

Kris Lovejoy, the Global Head of Strategy at Kyndryl, also pointed out:

“Most enterprise environments were built for people running tickets and tools, not for fleets of autonomous agents executing tasks across hybrid and multi-cloud estates, and this mismatch is limiting AI from moving out of pilots to outcomes.”

Frankly, it’s quite surprising how many businesses expect AI to compensate for years of neglected operational problems. An agent won’t magically reconcile contradictory policies or fix unreliable customer records. Give it permission to act on that information, and the mistakes become considerably more expensive.

Just look at Klarna, after aggressively pursuing AI-led service, CEO Sebastian Siemiatkowski acknowledged in May 2025 that excessive emphasis on cost had contributed to lower quality. The company subsequently renewed its investment in human support.

Which Customer Experience Tasks Should AI Handle Autonomously?

We’ve answered this question before in this guide to what contact centers should be automating first.

For now, the biggest point is this: An AI agent can be trusted with one request and still be completely unsuitable for the next. The deciding factor is what a mistake would cost both you and the customer, including the damage that can’t easily be undone.

Compare a missed delivery with a repayment-plan request from someone who’s just been laid off. They don’t belong under the same approval rules, even if the software could technically process both.

How Do You Decide Which Actions AI Can Take On?

Start with what a mistake would cost. How easily could the business undo it? What would the customer lose? And does the AI have reliable information as well as permission to act?

		Account authentication and accurate tracking data
Changing a delivery appointment		Available slots and confirmation
		Eligibility, payment permissions, transaction limits
Offering exceptional compensation		Financial authority and customer circumstances
		Account security and potential customer harm
Resolving a vulnerable customer’s dispute		Context, fairness, and appropriate specialist support

Those are just recommendations, though, not universal rules. A $20 refund under a straightforward returns policy deserves a different assessment from a disputed $2,000 insurance payment.

When Should AI Escalate to a Human?

The goal should always be safety and control; even if an AI agent should be able to resolve a problem for a customer faster, that doesn’t mean customers should be locked into automation.

Remember the Bobbi agent? It has very specific guardrails. Bobbi only helps people with non-emergency inquiries, drawing on around 1,900 knowledge articles, and recognizes topics involving vulnerability or potential harm, directing those interactions toward human support.

In one case the organization told us about, a vulnerable person used Bobbi to understand what was happening after criminals had taken over their home. The subsequent human intervention helped safeguard the individual, and police reported four arrests.

“Crucially, the right 25% are escalated.”

Escalation was treated as the right result for sensitive contacts. A containment percentage alone would miss that completely.

How Much Control Should Customers Retain?

Control isn’t just for contact centers as the agentic CX trust gap keeps widening.

Customers should know when AI is handling their request, and they should have a practical route to someone who can fix problems beyond the agent’s authority. Those are the main rules.

Five9’s 2026 consumer research found that trust in AI rose from 26% without a human option to 55% when one was available.

Permissions are essential too, as Frances Zelazny, General Manager of New Market Innovations at identity verification company Prove, told us in August 2026:

“One [option] is to tell the agent as you cross thresholds, you have to get new permissions.”

That needs to go a step further, though. Those restrictions have to be enforced through the systems the agent accesses. Telling an AI to ask for approval isn’t very reassuring if its credentials already let it skip that step.

Who Is Accountable? Building an Agentic CX Operating Model

One of the biggest mistakes CX leaders keep making with agentic AI is believing that if something goes wrong, they can still blame the model, or the vendor. Regulators probably won’t agree with that, as CheckPoint told us in August 2026:

“Controls can be outsourced. Accountability cannot. The risk decision stays with you.”

An effective agentic CX operating model needs named decision-makers, with clear authority over the actions an AI can take.

	Chief Customer Officer or CX executive
Daily monitoring and intervention
System access and technical controls
Risk policies and regulatory requirements
Escalations and employee oversight	Contact center operations and QA
Approval of expanded AI authority	Designated executive or governance committee

We won’t pretend we can give you a complete guide to legal liability here, since the risks depend on the circumstances, jurisdiction, and contracts. Operational responsibility, though, should never be left open to interpretation.

What Should an Agent Authority Charter Include?

Every company won’t necessarily need to go out and hire new experts before they deploy agentic AI, but if you’re preparing for risk, start early.

Back in December 2025, Simplyhealth’s Chief Customer Officer Claudia Nicholls told us that operating as a regulated business means working within a risk appetite agreed with the board, and the insurer established an AI forum (co-chaired by its customer service director and CTO, with input from fraud, risk, and compliance specialists) to assess every new proposal against it.

That’s a sensible arrangement. Someone has to challenge an idea before customers become the testing ground.

Even if you don’t hire new team members, an Agent Authority Charter should record the business owner, approved purpose, allowed and prohibited actions, system permissions, approval conditions, and the person able to restrict or suspend the agent. Review it whenever the agent gains a new tool or decision right.

A delivery agent’s charter might permit one replacement shipment after identity and inventory checks, send anything outside standard policy to a supervisor, and give CX operations the power to suspend the action if duplicate shipments appear.

What If Your AI Agents Start Working Against Each Other?

It’s a harder problem than it looks. Lots of companies are busy building guardrails against hallucinations, data leaks, and dodgy answers, but not many have considered what happens when agents with different rules collide.

Picture a customer who wants to cancel a subscription. The service agent accepts the request, but a retention agent has instructions to prevent cancellations, and billing has another payment scheduled. Every system is following its instructions, yet the customer ends up arguing with a company that can’t agree with itself.