Customer service in the age of AI

There is a familiar moment in almost every delivery journey: the customer who opens a tracking page, sees that the parcel is still "in transit" just like an hour ago, and presses refresh again.

The information may not have changed, but there might be so many reasons worth the extra check: the parcel might contain a birthday present, an urgently needed component, or something the customer has arranged their day around. The shipment is operational data and the refresh is an expression of uncertainty. 

At nShift, we believe customer service in the age of AI should begin there. Technology can make service faster and more scalable, but its purpose should not simply be to reduce conversations; it should reduce uncertainty, help people solve problems well, and provide accurate support when something does not go to plan. 

AI is not a substitute for care, judgment, or experience. Quite the opposite: it should help make those qualities available more consistently. 

The people who benefit first

A study of generative AI in customer support followed 5,172 service professionals through the rollout of an AI assistant. This was a live deployment across a working support organization, not a lab test. Published in The Quarterly Journal of Economics in 2025, it found that access to the assistant raised issues resolved per hour by 15%.

The greatest gains appeared among colleagues who were earlier in their service careers. With support from the assistant, they became faster and more effective, applying knowledge that would normally take longer to acquire. For nShift, this points to a constructive role for AI: namely helping people build confidence sooner and making proven knowledge easier to find.  

The same study also reinforces the value of experience. The most seasoned professionals gained less from standardized guidance, and quality dipped slightly among the strongest performers. I read that as confirmation: expertise goes past knowing the established answer, into recognizing when the established answer no longer fits, and that recognition is the part no assistant supplies.

So people should start from a stronger position and keep the freedom to think and apply judgment. Our people should feel supported by AI, never boxed in by it. Their knowledge is what makes responsible automation possible in the first place, and the person still has to own the outcome. In other words, AI shortens the path to the right answer without handing anyone a way to pass the responsibility on.

But improving support after a question is raised is only part of the opportunity. The next step is to reduce the uncertainty that caused it.

The unnecessary question that shouldn't exist

A "Where is my order?" inquiry usually starts with an information gap: the customer doesn't know whether the parcel is moving, whether it is late, or what happens next. An automated answer may save time, but it does not necessarily improve the experience that created the question.

Our view is that service begins earlier, with a realistic delivery promise, reliable carrier events, clear communication, and the ability to notice when expected information has not arrived. 

nShift Track is built for that. It communicates across delivery milestones, identifies stalled parcels and other non-events, and gives service teams a clearer view of what is happening on the ground. Customers using Track see up to 50% fewer delivery-related support tickets.

I care less about the ticket count than about the change in the customer’s experience. When people get useful information at the right time, some questions never need to be raised, and every accurate update removes one reason to refresh the page and wonder whether anyone is paying attention.

Customer service does not begin at the support desk. The accuracy of the promise and the handling of the exception decide whether a customer feels looked after.

A convincing answer can still be wrong

Generative AI can produce polished answers from incomplete information. In customer service, a fluent answer hasn't earned our confidence until it is also correct.

So our AI is grounded in the products and business rules it is asked to explain - a principle that appears in nShift Companion, though its role differs depending on where you meet it.

• In Checkout, Companion works with a customer’s own setup and business rules to help teams understand delivery choices and configuration.

• In Track, it currently explains how the product works and how tracking pages, notifications, and reporting are configured. Those answers come from nShift’s own product content. Shipment-aware support, drawing on individual delivery data, is planned as a future development.

AI reliability depends less on the model than on the discipline behind it. Every shortcut someone takes when closing a case becomes a gap the AI inherits months later, which makes grounding a matter of habits before architecture. So we stay precise about what AI does today and what still needs a person. It does not improvise where a verified answer is required, and it hands over to human expertise when interpretation or accountability is needed.

The knowledge the machine inherits

The same standard applies behind the customer-facing experience.

We are using proprietary agentic AI across the design, testing, maintenance, and scaling of our carrier integrations. That estate runs to more than 1,000 carrier connections and tens of thousands of carrier services, each with shifting technical requirements and its own practical edge cases.

The tool our specialists use reads a carrier’s own documentation and drafts the integration definition from it. What happens next is the part I would point to: the draft that lands in the carrier editor is never published automatically. A specialist opens it, finds what the AI misread, writes down the correct value, and sends it back to be generated again; that loop runs as many times as it takes. A person decides when the integration is right, and a person publishes it.

We built it that way on purpose, because our Carrier Integration Specialists are the ones who recognize the case that does not look like the others. That recognition is what makes integrations hold up in the real world - the technology scales because these people spent years building the knowledge it runs on. AI preserves and extends that knowledge without manufacturing the judgment that comes from understanding why the circumstances changed.

The problem with invisible intelligence

Research shows that people behave differently when they know they are talking to AI. A field experiment with more than 6,200 customers found that undisclosed chatbots performed as well as proficient human agents in a structured sales setting. When the chatbot identity was disclosed before the conversation, purchase rates fell by more than 79%.

That figure deserves a caveat before anyone steers their strategy: it measures purchase behavior in a sales conversation rather than satisfaction in a service, and it reflects what customers expected of chatbots at the time of the study. Read carefully, it says that expectations govern how AI gets judged, and expectations are shifting while confidence has to be earned through clarity and accuracy, over time.

Regulation is heading the same way. Article 50 of the EU AI Act requires systems designed to interact directly with people to tell users they are interacting with AI, unless that is already obvious. Those transparency obligations apply from 2 August 2026.

Transparency belongs alongside service quality for us at nShift. People should know when AI is involved, what it helps with, and how to reach a person when they need one, in the language they prefer. Straightforward questions get resolved efficiently while no one has to fight the system to reach someone who can take responsibility.

The customer does not see the org chart

At nShift, Professional Services, Customer Support, and Carrier teams sit within one Customer organization. We combined them because customers experience them as one relationship.

Onboarding affects how quickly value arrives. Support shapes confidence at the moment help is needed. Carrier connectivity helps determine whether delivery promises can be kept at all. Different nShift specialists own each of those, and the customer should never have to work out where one team ends and the next begins.

That is also why each part of the business should not build its own separate AI toolset. A support bot here, a carrier assistant there, and an onboarding chatbot somewhere else would only recreate the fragmentation customers already might feel from the org chart. One connected approach, grounded in the same knowledge, is what makes a single relationship possible. 

So the priority is getting these teams to learn from each other:

• Repeated support questions improve guidance and onboarding.

• Carrier changes reach customer workflows sooner.

• Delivery data shows us where uncertainty is being created.

Run this way, customer service becomes a way of seeing the whole customer journey and acting on what it shows us. AI helps us find recurring issues and respond more consistently. But the purpose stays human: helping our people solve problems earlier and spend their time where their experience counts. AI should make service more informed and more dependable, and it should strengthen the people who serve customers. It should also remove avoidable uncertainty without pretending every situation can be automated.

Customers will still press refresh, like the example I gave in the opening. But our job is to make sure the information is accurate, the promise is credible, and someone is paying attention when the answer is not yet clear.

For more CX insights and meaningful conversations, let's connect on LinkedIn.