SportTech
Tottenham Hotspur recently announced that it has become an ‘agentic enterprise’ via tech platform Salesforce, with more personalised, connected experiences for fans across every touchpoint.
The club’s digital transformation began with a fundamental challenge: fan data was fragmented across dozens of ticketing, retail, hospitality, app and web systems.
Among the developments in response was a new AI agent built on Agentforce, Ask Spurs, which the club says has an 86% engagement rate.
However the more pertinent figure is the 20% of interactions which are actually resolved by the agent: when Ask Spurs can’t resolve a question, it hands the conversation to a human agent on the Tottenham support team.
“Picture a scaleup in Leeds that puts one of the new AI customer service agents on its online shop and, three months later, reports that most visitors now talk to it,” Rafael Patron, president of the scientific-technical committee at the Italian Association of Artificial Intelligence Professionals (AIPIA), tells BusinessCloud.
“The board is pleased. Nobody, however, can say how many of those conversations ended with a customer who got what they came for.
“That gap is where many of these projects quietly stall, and Tottenham Hotspur’s recent move shows why.
“The club’s Ask Spurs agent answers fans in 11 languages about tickets, membership, retail and events, and Salesforce, the vendor behind it, puts its engagement rate at 86%. Then comes a second figure: the agent resolves 20% of interactions on its own. The rest go to a person on the club’s support team.
“Both numbers are the vendor’s, and neither is a scandal. Read together, though, they tell a business owner more than either does alone.”
Engagement is not resolution
Engagement measures whether people use the thing. It says little about whether the thing worked, says Patron.
“A customer who types three questions, gets three unhelpful replies and then phones the call centre has engaged enthusiastically,” he explains.
“Resolution is harder to measure and more honest. It asks whether the query was closed without anyone stepping in, and whether the customer stayed satisfied rather than returning a day later with the same question.
“One in five can be a sound result for a first year, particularly when queries involve refunds or account changes that a firm would rather keep in human hands. What matters is that the figure is known, tracked and expected to move.
“The trap for smaller firms is that the dashboards supplied with most tools put the flattering metric at the top. Conversations started and messages exchanged are easy to count, and both can rise while service gets worse.”
The unglamorous part: getting customer data into one place
The detail in the Tottenham story that deserves more attention than the percentages is what came before the agent.
Spurs spent two years replacing around 30 separate systems and connecting more than 10 others, bringing over 4.6 million fan accounts together into single profiles. Ticketing, retail, hospitality, the app and the website had each held their own slice of the same supporter.
“An agent can only answer from what it can see,” continued Patron. “If a customer’s order history sits in one system and last month’s complaint in a shared inbox, the agent will answer confidently from a partial picture. That is worse than no answer, because it sounds right.
“Most UK firms will never face a job on that scale, but the principle shrinks well. Before choosing a tool, map the five or six questions customers ask most often. For each one, write down which system holds the answer.
“Where the answer is ‘three systems and a spreadsheet’, the data work comes first. It is slow, it rarely makes a press release, and it decides whether the agent is any good.”
Designing the handoff to a human
If four out of five conversations end with a person, the handoff is the main event rather than a fallback. Yet it is often the part designed last, says Patron.
“A good handoff passes the full conversation to the colleague who picks it up, so the customer does not have to repeat themselves. It happens early when the signals are clear: a complaint, a vulnerable customer, a payment dispute, a second ‘that’s not what I asked’. And it is honest about waiting times: an agent that promises a callback within the hour and delivers it the next day is worse than a plain contact form.
“Firms should watch the quality of the handoff as closely as the volume. How many escalated customers had to explain their problem again? How long did they wait? How many came back unhappy? These questions are uncomfortable because they measure the join between machine and staff, which is exactly where nobody feels fully in charge.”
Who owns what the agent says
That sense of nobody being in charge is not peculiar to customer service. Boston Consulting Group’s ‘AI at Work’ survey, published in June 2026 with 11,749 respondents in markets including the UK, found that, among regular AI users, the integration of AI agents into workflows had more than doubled since 2025.
Governance, which BCG describes as oversight and accountability, still lags far behind the technology. Among the 8,849 regular AI users who had heard of AI agents, half said they lack clear guidance on how to manage teams where people and AI work together.
“For a customer-facing agent, ownership needs a name. One person should answer for what the agent says, keep the list of topics it must never handle alone and decide when to pause it,” says Patron. “That person does not need to be technical. They need the authority to say ‘switch it off on refunds until Monday’ and have it happen.”
Where business readers and AI professionals meet
We at BusinessCloud report on UK firms adopting new technology, from our Tech 50 rankings to deployments like Tottenham’s. The people running those firms are now asking a second-stage question: how to judge an agent and who should look after it.
That is the territory of the Italian Association of Artificial Intelligence Professionals (AIPIA), a non-profit body that brings together the people who design, integrate and oversee AI systems inside organisations.
“For UK firms with European customers, AIPIA’s guide to the AI Act sets out the transparency obligations alongside the rest of the regulation, including the rule that a chatbot must make clear it is not human,” says Patron.
“Founders and managers decide what the agent is for and what stays with their staff; AI professionals turn those decisions into permissions, escalation rules and logs that actually work. The two groups tend to talk past each other, and the most useful conversations happen when they sit down with the same set of numbers.”
Four numbers for AI customer service agents in the first six months
He continues: “The first is the resolution rate, defined strictly: queries closed by the agent with no human involvement and no repeat contact on the same issue within a week.
“The second is handoff quality. Track how long escalated customers wait and how often they have to repeat themselves, and read a sample of those conversations every month.
“The third is the error rate. Take a weekly sample of answers, check them against the facts, and record how many were wrong, misleading or out of date. An agent that is regularly wrong is a liability, however polite it is.
“The fourth is the one most firms forget: time saved, and where it went. BCG found that, in its sample of 8,989 regular AI users, about two thirds of frontline employees receive limited or no guidance on what to do with the time AI saves them.
“If an agent frees up hours on the support desk, someone should decide in advance whether they go to complex cases or to calling unhappy customers back. Otherwise the saving disappears into the working week, and the business case with it.
“None of these numbers will look as good on a slide as 86% engagement. They are the ones a finance director will ask about at year end, and the ones that show whether an agent is ready to be trusted with harder questions.”