Part three: agent of choice, or choice of agents?

Part one argued that putting AI into customer support is the safest and least valuable thing you can do with it. Part two looked at Oura, who take measuring the AI user experience seriously, and what that takes.

Welcome to part 3, where we’ll focus on whether to build a customer-facing agent of your own, and what’s happening in your business environment while you make your mind up.

What we’re talking about, and what we’re not

AI can be a daunting topic, and even such generic terms are not helpful. Most people use it daily in some form and many have even heard of an “agent harness” but would rather not be asked to explain what it means. That’s fine, this blog will not go into the weeds.

What’s worth understanding is simpler. Let’s talk about conversational AI Agents, colloquially known as chatbots, and the strategy questions surrounding building them for your business.

An LLM-based AI agent doesn’t follow a fixed script. Even when it’s drawing on approved information and designed to give consistent answers, the way a conversation unfolds depends on what someone asks, how they ask it, and what happens next. Conversations can wander to places nobody designed them to go.

They also talk like people, so we treat them as we would our fellow humans. We explain ourselves, lose patience, give them the benefit of the doubt, and we judge them more harshly than we would a cocky intern.

Understanding that back-and-forth is our focus at Nebuly. We help companies see whether their AI agents give people a good experience, and our clients split roughly evenly between those measuring agents built for employees and agents built for customers. An agent is only valuable if it’s valuable to someone, and since you can’t script every conversation in advance, the only way to know is to read the conversations.

Our work on employee-facing AI gives us a unique vantage point as it’s where the technology has been tested most aggressively.

Before we use AI as customers, we use it in our work

Since ChatGPT Enterprise launched in August 2023, we’ve all become guinea pigs in the race to find out how this technology creates value. Microsoft actually beat OpenAI to it with Bing Chat Enterprise (now Copilot), but somehow, it’s just not as catchy.

Businesses let this new technology follow the paths of least resistance to find value. That meant pushing adoption of multiple agents with overlapping capabilities. Attending the Forrester and Gartner conferences these past two weeks I heard it referred to as Ready, Fire, Aim. But it’s starting to change.

McKinsey’s latest survey found that the small group of companies seeing significant financial returns from AI are far more likely to use it for growth as well as efficiency (74% against 47%), and twice as likely to have defined ways of measuring its impact. Similarly, Gartner found that the organisations doing well track returns continuously and stop the projects that aren’t working.

In contrast customer-facing agentic AI has mostly been retrofitted onto an existing legacy product, the FAQ or menu-based chatbot. As outlined in Part 1, this has not exactly been a success from a customer experience point of view. The problem isn’t even a lack of measurement, on the contrary, but it’s the wrong kind of measurement. AI agents deployed in service of optimising a cost centre will do just that, Terminator style, No humans!

The key difference in a customer-facing scenario is failure visibility. An employee-facing agent that isn’t delivering on users’ needs stands out because there’s more qualitative experience measurement; employees complain. A weak customer facing agent can quietly degrade your brand experience while KPIs remain in the green. Customers pay twice for doing business with you and this eventually undermines your brand.

That discipline of measuring the experience and adjusting accordingly, learned in employee-facing AI needs to transfer to teams building customer-facing agents, and it has to happen more broadly than the few I’ve highlighted hall of famers from Part 1.

But now…

Why build a customer-facing agent at all?

The question on the table, as Accenture Song put it in their Unmissable report, is whether you’ll be the choice of agents or an agent of choice. Most companies, they argue, will need to become the choice of agents, making their offerings easy for AI agents to find, evaluate and buy. A smaller number, those with real category authority and proprietary data, may also become an agent of choice: the one customers and other agents trust and turn to, again and again.

The first of those is already happening. Whether you planned for it or not, your customers are asking general assistants about your products and services, and you don’t control the answers.

Brands are already choosing how to navigate this. In the US, Delta is stopping AI agents from booking flights on customers’ behalf, and some banks including JPMorgan Chase are blocking them from customer accounts over security concerns. In the UK, John Lewis has gone the other way, making its products discoverable and purchasable through ChatGPT and Gemini, while Tesco is building its own.

Even where general agents are welcome, it isn’t seamless. The anti-bot systems that protect websites from scams and hacking catch AI agents too, and those defences exist for good reasons: nobody wants to compete with bots for concert tickets or restaurant bookings.

Your own agent doesn’t face those barriers, so that’s the first reason.

What’s already happening in the absence of strong brand agents

In September the Financial Times covered a study by the UK fintech Saturn, which put real financial questions to the most popular models, ChatGPT, Claude, Copilot, Gemini and Grok among them. The questions covered debt, mortgages, pensions, tax and retirement and the models gave wrong or incomplete advice 57% of the time.

On more complex topics the failure rate rose to 88%. Even the best performer fell short on 39% of its answers. The models are answering without the rules, the product knowledge or the customer’s circumstances in front of them, and it shows.

People act on those answers anyway. The FCA’s Mills Review found that one in five UK adults are open to letting AI make financial decisions for them, with demand strongest in debt, pensions and investments, the kind of complex question the models handled worst. A PensionBee survey found most users would accept a chatbot’s answer on financial decisions without checking.

A regulated firm giving that advice has a duty of care, a paper trail, a complaints process and a compensation scheme behind it. A human adviser getting 57% of their answers wrong would be seeking alternative employment. The general assistants have none of that: as Saturn’s chief executive put it, AI financial advice is currently unregulated.

That’s also part of why regulated firms hesitate to build their own AI agent. Put your own agent in front of customers and your duty of care comes with it. That’s a real cost.

However, when the general assistants get your sector wrong, trust suffers, whether the answer came from you or not.

Beyond the obvious sectors like banking, insurance, health, pharma, legal, or anywhere the answer has consequences, I don’t believe any business wants their customers to lose faith in their market.

What you’re playing for

Everything so far is the defensive case. One company I spoke to recently described their AI agent as both defensive and offensive, and the distinction is useful.

Defensive means protecting customers from bad generic advice and protecting the brand from the fallout when they act on it. Offensive means giving customers a reason to come back: an experience good enough to build loyalty and grow the business.

Most are playing defence right now but going on offense holds the bigger prize. JKR and Kantar’s Be Distinctive Everywhere report found that 71% of brand equity comes from experience: 30% from what a brand owns and delivers directly; your product, and 41% from what’s earned through friends, reviews and social; proof.

Put that next to Accenture’s two paths and becoming an agent of choice is owned experience in its most direct form: a one-to-one conversation between your customer and your brand.

Once you have an agent, the conversation is part of the product, and it’s where customers decide whether you’re any different.

The report goes further to say that frictionless isn’t enough: an experience anyone else could have offered meets the need but doesn’t set the brand apart. While general assistants still disappoint simply knowing who you’re talking to sets you apart. Going further Kantar associates stronger growth and loyalty with meaningfully differentiated brands, so the better your agent the better your return.

Who gets to be the agent of choice?

Accenture reserves the agent of choice for companies that already have deep category authority and proprietary data. That treats authority as something a brand brings with it into AI. I’d argue that’s not a given considering how many bad agentic products we see from household names. In practice it’s built or lost in the conversations themselves, and that changes who can win.

Medallia’s 2026 research found that more than 40% of consumers find human errors easier to forgive than AI ones, against 7% who say the opposite. That makes the stakes uneven. For a challenger, a customer-facing agent is a native shop window, and sometimes it’s the product itself, so every good conversation builds authority it couldn’t otherwise buy.

For a household name, the agent is an extra surface on top of a large existing base, and a bad one puts more at risk than a good one adds. An established brand starts ahead, but it can still lose.

Virgin Atlantic is doing both. It runs its own AI Concierge, built to carry what its Chief Experience Officer calls the ambition of becoming “the most loved travel company”, and in April it became the first airline to launch an app inside ChatGPT.

So who should build a customer-facing agent?

Not everyone. The vendor version of this argument is that every company needs one, and that isn’t true.

Staying with Virgin Atlantic for a moment, they’ve shown that a customer-facing agent needn’t be all singing, all dancing in order to be successful. They started with a narrow, high-value job, using AI for what it’s good at; unstructured conversation, so they deployed Concierge in a customer journey they knew intimately. It has to be good at the job it’s given, and good means the customer got what they came for.

Two questions are worth asking. Do you hold domain or customer knowledge that a general model can’t have? And does a wrong answer cost your customer something real?

Two yeses and the case is strong. Two noes and being easy for the general assistants to find and recommend is a sensible strategy, so spend the money somewhere more useful. One of each, and it depends on what your competitors do next.

The bar here is not a flagship assistant either: a machinery hire firm answering availability and delivery questions through an AI agent on WhatsApp is carrying its brand in every conversation, and a mistake can cost its customers days on a project.

Ask those questions thoughtfully with the team and more would have a case than are currently building. I think many are currently answering no by default, and that has to change.

Capability is no longer the limitation it was. A deep market of product engineering firms, systems integrators and consultancies that build these things for a living, have grown up in recent years.

What’s usually missing, and why the case never gets made, is a decision about what good would look like and who is accountable when the agent gets it wrong. Neither comes with the technology.

Building it is only half the challenge

The harder part is knowing whether what you’ve built is any good, and this is where the lessons from employee-facing AI come back in.

The companies getting real value from AI internally are the ones that measure it and adapt their businesses to better leverage the technology. That’s the lesson worth bringing across.

Klarna shows what happens with poor measurement. In March 2024 it announced that its OpenAI-built assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats and the work of 700 full-time staff, with satisfaction on par with humans and a projected $40m profit improvement.

Fourteen months later, its CEO told Bloomberg that prioritising cost above experience resulted in “lower quality” and that Klarna was making sure customers could always reach a human.

Every number in the first announcement said the product was working fine. The correction came from customers and the CEO’s own judgement, more than a year later and in public.

Klarna didn’t retreat from AI. By late 2025 its assistant was doing the work of 853 full-time staff and had saved around $60m. But the damage was real, borne by customers and Klarna’s brand became synonymous with bad AI customer service. This doesn’t do justice to the efforts they’ve made to build a now successful AI agent. Everybody loses.

It shouldn’t take that, and today many leaders would catch it sooner, but judgement on its own is still a bad strategy. Evidence that an agent is letting customers down should be caught much earlier, and ideally privately.

Get building, and measure the experience

Conversational AI is becoming the interface. As customers and employees shift from navigating digital channels to delegating tasks to AI agents, traditional measures of engagement lose relevance. The conversation itself becomes the only record of the user experience, whether or not they got what they came for, and why. This creates an unusually rich source of insights, but also a new point of risk; poor answers rapidly erode trust, push users back to higher-cost channels and compound brand damage. Yet most organisations are measuring AI adoption, not value.

Technical observability dashboards show quantitative metrics (tokens, latency, traces), but not whether an agent delivered the right outcome, if that outcome was valuable or where and why it failed. Without that qualitative evidence, businesses cannot reliably improve those agentic products, manage risk or justify the next investment.

Digital products are built on analytics and user research. AI products need the same, drawn from the conversations themselves.

Let’s talk

We’re hosting a Nebuly Sessions roundtable on 21 October at The Hoxton, Southwark, on customer-facing AI that differentiates your brand. It’s a dinner, not a panel, so expect healthy debate and disagreement rather than slides. Register below: