The call starts well. A friendly voice, recorded sometime in the last decade, thanks me for calling and warns that options “have recently changed,” which is corporate for we moved the exit again. I press four. I am told my call matters. I press one. I am told, politely, that this query is handled online now, and the line goes dead. So I open the website, where a chatbot named after a woman asks how it can help today, and when I type “speak to a person,” it replies with a link to a help article about updating my password, which is not what I asked, and we both know it.
Over several weeks, we ran this experiment thirty times, calling and chatting our way through the customer-service apparatus of thirty major companies — airlines, banks, retailers, telecoms, the platforms that run daily life. We timed everything. We are not publishing the league table, partly because hold times vary by hour and by luck, and partly because the individual numbers turned out to be the least interesting thing we measured. What the stopwatch actually captured was not congestion. It was intent. Some companies were slow the way a busy restaurant is slow. Most were slow the way a hedge maze is slow: on purpose, by design, with the walls budgeted and the dead ends A/B tested.
That distinction matters, because we tend to experience customer service as weather — an unfortunate condition, nobody’s fault, surely temporary. It is not weather. It is architecture.
The department that wins by losing you
Every other part of a company is measured on engagement. Marketing wants more opens, product wants more sessions, sales wants more conversations. Support is the lone exception: the one department whose defining success metric is fewer conversations. The industry has a vocabulary for this — “deflection,” “containment,” “call avoidance” — and dashboards to match. A chatbot that fails to escalate you is not a broken chatbot. On someone’s weekly report, it is a containment win. Every extra menu layer, every help centre that loops back to the article you came from, every “have you tried our community forums?” is a small, countable victory that no individual employee chose but the entire organization is paid to produce.
The economics underneath are blunt. “There’s a straight-up clear hierarchy,” Harvard Business School professor Ryan Buell told Vox in 2023: face-to-face costs more than phone, phone more than chat, chat more than email, email more than no human at all. A call to a contact center runs to several dollars; a deflected one costs fractions of a cent. Multiply that by millions of customers and the phone tree stops looking like a nuisance and starts looking like a savings account. When Frontier Airlines announced in late 2022 that it was simply ending phone support, it wasn’t confessing a failure. It was publishing the logic everyone else runs quietly.
Our own timings bore this out in texture if not in publishable precision. The companies that were easiest to reach shared a tell: they still had something to sell us at the moment of contact. New-customer lines answered briskly. Cancellation lines, billing disputes, anything where we represented pure cost — those journeys went long, looping, and strange. The delay tracked the incentive almost perfectly.
What the machines found
We are not the only ones with a stopwatch. In 2022, Consumer NZ timed the path to a human at 21 major New Zealand companies and found holds ranging from 15 seconds to nearly an hour — and, more tellingly, a bank whose automated voice estimated a 10-to-20-minute wait while the actual wait ran roughly double. Even the estimate was a deflection tool. A Which? survey from around the same period found British Gas averaging more than 23 minutes to answer, up from just over four before the pandemic — the moment many companies discovered that customers would tolerate, or at least survive, far less service than previously assumed.
The largest recent audit comes with a warning label. Parloa, a company that sells AI voice agents, tested more than 10,000 enterprise contact offerings in a 2026 study and found results almost too bleak to be neutral: of 771 attempts to escalate from a chatbot to a human, 78 succeeded. Chatbots resolved the testers’ goal in under 9 percent of interactions. More than 40 percent of the enterprise websites examined showed no visible phone or chat contact at all, and 96 percent of phone systems were legacy menus, typically three or four levels deep. Parloa profits from exactly this diagnosis, so discount accordingly — but our weeks of pressing four and typing “agent” suggest the direction, if not the decimals, is right. The chatbot era did not fix the phone tree. It gave the phone tree a friendlier face and a lower salary.
Easy to buy, impossible to reach
Here is the combination that took us weeks to properly see, because it feels like a contradiction and isn’t: a company can be genuinely, impressively easy to buy from and genuinely, structurally impossible to reach. One-click checkout and a support maze are not two competing impulses. They are the same impulse — friction engineering — applied on opposite sides of the transaction. Before the sale, every unnecessary step is hunted down and eliminated, because steps cost conversions. After the sale, every necessary step is lovingly maintained, because steps cost conversations. The same design teams, the same funnel analytics, the same A/B tests, pointed in opposite directions. You are not experiencing bad service and good commerce in one company. You are experiencing one perfectly coherent system that has calculated, to the cent, the moment you stop being worth talking to.
The stopwatch, in the end, measured something simpler than hold time. It measured how much of you is still revenue. Where the answer was “plenty,” a human appeared in under a minute, warm and empowered. Where it wasn’t, we got the maze — and the maze, remember, is not broken. It is finished. It is exactly what was ordered. The next time a recorded voice tells you your call is important, the most accurate response is not frustration. It is a small, clarifying translation: your call is important enough to be measured, and the measurement said no.