The bot that will not pass you to a human is not broken
The bot that keeps you trapped in a loop, refusing to pass you to a person, is not broken. It is doing its job. For twenty years, customer service has been run on a metric with an ugly name, deflection, and its single purpose is to stop you reaching a human. Adding AI to that goal has given us the experience many of us now dread. The failure is not technical. It is an incentive, and until leaders name it, more capable AI will simply deliver a more sophisticated way of pushing customers away.
On The AI Adoption Podcast I spoke with Matt Price, Founder and Chief Executive of Crescendo, a customer experience company built natively on AI. He built the business in an unusual way, running 200 contact centres directly rather than selling software and leaving clients to operate it. That operational vantage point gives his argument weight, because he has watched these programmes succeed and fail from the inside.
Deflection Is A Metric That Punishes The Customer
Deflection asks a single question: can I stop this customer talking to me, or to a human? Matt was clear that overlaying AI on that mindset, as a better way to keep people out, is precisely what has not worked. I recognised the pattern from my own week. The day before we recorded, a website had me in a loop with a bot that proudly announced it was driven by AI, and no amount of asking would get me to a person. I ended up emailing the company, angry, and I work in this field. Anyone arriving less charitably will simply learn to hate the technology.
The deeper problem is that deflection flatters the people who report on it. A team can celebrate thirty, forty, fifty per cent deflection and never look at the customers the bot could not help, who now have a worse experience than before. Matt's point is that a headline resolution rate hides its own failures. Judge the interaction end to end, until the problem is actually solved, and a different picture appears. The metric that looks like efficiency is often just suppression, and suppression carries costs that never show up in the deflection number.
The Revenue Hiding Inside The Inquiries You Suppress
Here the conversation turned from cost to revenue generation. Move customer care from the back door of a business, where it is hard to find, to the front of the website, and invite people to ask anything, and the volume of engagement roughly triples. Under the old model that would have been a disaster, three times the inquiries with the same staff and a collapse in service. Matt's claim is that over half of those additional inquiries are revenue related, questions from people who intend to buy. An interaction that leads with service, he argues, converts about four times better than a sales bot on its own, for the plain reason he gives, that “most of us want to be served rather than sold to.”
The Lovepop example he gave makes it concrete. The greeting card company could always refund a card that failed to arrive, but standard support took around four hours to respond and a refund needed a second person to check, perhaps a day later. By then the moment had passed. With AI handling the immediate response and the options, and staff freed to work escalations within fifteen minutes, a problem that used to be unsolvable in time became solvable. Its Trustpilot rating, Matt said, moved from 3.6 to 4.6 within a couple of weeks.
A second example pointed the same way, toward growth rather than savings. Sweet Bee Organics, the chemical free suncream brand founded by Hollie King and sold direct to consumers, had scaled fast while its customer service lagged behind. Matt's team took on the whole service, AI with humans in the loop. The barrier to opening new European markets, in his account, was not the product but customer care in each language, and he cited Spain and Italy as examples. Because the AI handles fifty six languages and can be reinforced with human agents quickly, that barrier fell. Treated as a growth function rather than a cost to be minimised, customer care stops being the thing that holds an expansion back.
There is a genuine tension in his account worth naming. He also says that for every ten percentage points of automation a company can retire three to five points of labour, so this is not a story with no headcount effect. Yet his conclusion is the opposite of the industry's sales pitch. As volume rises and escalations rise with it, done properly this needs more people, not fewer. The saving is real, but it is dwarfed by the demand and the revenue the old deflection habit was quietly throwing away.
Transformation Dies In The Middle, Unless Leaders Force The Issue
If the prize is this clear, the puzzle is that so few organisations reach it. Matt's answer was the sharpest line of the conversation. Transformation and innovation, he said, go to die in middle management. Ask the manager who must restructure a team, or the person who owns the old technology to evaluate its replacement, and they will find a thousand and one reasons the project should wait. This is not cynicism about individuals. It is a predictable response to being asked to dismantle your own role, and it is the point at which many of these initiatives quietly stall.
Link to Neha Kabra https://open.spotify.com/episode/2dONk0xtBS40x6uIAFqgD9?si=zSyY-34GT420m-NR6rCzPw
His prescription pairs objectives set from the top with the patient work of bringing people along. That means being honest with agents about the change to their jobs, and about the escalation work that remains, which is more valuable, not less. It means keeping the technology manager involved in choosing what comes next, rather than defending what exists. Bolting a bot onto a broken process only accentuates the flaws already there, so the harder task is to redesign the slice of the experience first, quickly, with clear measures of success. None of this is cheap; Matt put the cost of the technology and the people to run it at a quarter of a million pounds or more. In a boardroom he framed the stakes plainly: a business spending at least thirty million a year on customer care could, with sustained investment and focus, halve that over two to three years while improving satisfaction. That is a transformation programme, not a software purchase. I have argued for years that transformation is a human discipline before it is a technical one, and little in this conversation suggested otherwise.
The question I would put to any board
The uncomfortable conclusion is that better AI will not rescue a company that is still measuring the wrong thing. If your customer service is judged on deflection, the technology will get more efficient at doing damage. The organisations that gain will be the ones that change the metric, align the incentives, including how they pay their vendors, and treat customer care as the place where a brand's promises are tested in public. The tools are ready. The question I would put to any board is simpler and harder: when a customer finally reaches you, does the company actually care, or has it just bought a more articulate way of saying no?
Listen to the full conversation
Comments