ISSUE 010
AI at the Front Desk
I keep getting asked whether a chatbot can cover the front desk. The research is better than I expected and narrower than anyone assumes. Almost all of it measures one firm or one country, and the two strongest studies land an order of magnitude apart. Here is what each actually measured, and why the gap between them is not a contradiction.
An idea worth thinking about
Across a whole country's workforce, chatbots saved about three percent of working hours and moved nothing else
Anders Humlum and Emilie Vestergaard surveyed roughly 25,000 workers at about 7,000 Danish workplaces in late 2023 and again in 2024, across eleven occupations including customer support specialists, and linked every answer to Danish administrative earnings and hours records. That linkage is the whole design: self-reported time savings on one side, payroll reality on the other.
Users reported saving an average of 2.8 percent of work hours. The spread ran from 0.6 percent for teachers whose employers did not encourage use, up to 6.8 percent for marketing professionals whose employers did. Employer encouragement lifted take-up from 47 percent to 83 percent for a typical worker, 38 percent of firms had an in-house model, and 30 percent of employees had received training.
On earnings and recorded hours the authors found no significant effect in any of the eleven occupations, with confidence intervals ruling out pooled effects larger than one percent. They are explicit about the limits: time savings are self-reported and may be overstated, the window closes in June 2024, employer initiatives were not randomly assigned, and Denmark is not the United States. What the paper establishes is that reported time savings and measured pay are two different quantities.
A tool I disappeared into
The federal projection for the occupation, with the number that never travels with the headline
The Bureau of Labor Statistics keeps an Occupational Outlook Handbook entry for customer service representatives, last modified 29 October 2026. The occupation holds about 2,666,000 jobs. BLS projects a 5 percent decline from 2025 to 2035, roughly 141,800 fewer positions, and attributes the fall to self-service systems, social media and mobile apps absorbing simple tasks, with retail trade affected most.
The figure that rarely travels with that headline is openings: about 289,500 a year over the decade, essentially all of them from workers transferring to other occupations or leaving the labour force. A shrinking occupation can still be among the most-hired-for jobs in the country, because turnover is an order of magnitude larger than the trend. Median pay was $21.53 an hour, or $44,770 a year, in May 2025.
One framing of my own, clearly marked as mine and not the BLS's: when owners compare a chatbot subscription to a person, the comparison in their head is usually against the whole salary. The honest comparison is against the share of that salary spent on the specific tasks a bot can finish without a human reading the result. The handbook will not do that arithmetic for you, but it gives you the wage to do it with.
Learn something
What happens to the result when you tell the customer it is a bot
A 2019 field experiment in Marketing Science is still the cleanest test of a question every owner reaches eventually. Xueming Luo and Siliang Tong of Temple University, Zheng Fang of Sichuan University and Zhe Qu of Fudan University randomly assigned about 6,000 customers of a financial services company to a human agent or a chatbot, and varied when — or whether — customers were told which one they had got.
The chatbot was roughly four times as effective at selling as inexperienced human workers. But when customers were told they were talking to a bot before they decided, purchase rates fell by more than 79.7 percent. The effect sat in the disclosure, not in the conversation. Telling people afterwards did not carry the same cost.
The caveats are heavy and I am not going to sand them down. This is 2019, several model generations before anything you would deploy today. It is one company in one sector. The INFORMS release reports the disclosure effect without baseline purchase rates, which makes the percentage harder to interpret than it looks. And the release describes no tested method for reducing the penalty. What the study establishes is that disclosure is a variable with a large coefficient, not that it carries that exact cost now.
What we’re seeing
The same technology measures 14 percent inside one firm and 2.8 percent across a whole economy, and the difference is integration rather than luck
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents handling about three million chats at a Fortune 500 enterprise software company whose own customers are small and medium U.S. businesses. Roughly 89 percent of the agents were outside the United States, mostly in the Philippines. Most adoption happened between November 2020 and February 2021, after a seven-week randomised pilot with 50 agents. The assistant suggested replies and surfaced internal documentation, and agents were free to ignore it.
Resolutions per hour rose about 14 percent on average. Novice and low-skilled agents gained about 34 percent. Agents with under a month of tenure gained about 46 percent. The top skill quintile gained nothing and showed slight declines in resolution rate and customer satisfaction. Agents two months in with the tool matched untreated agents with more than six months behind them.
Quality and staffing moved as well. Customer sentiment improved about 0.18 points on a scale running from minus one to one, roughly half a standard deviation, and requests to speak to a manager fell about 25 percent. Surveyed satisfaction showed no detectable change. Attrition fell across skill groups; for agents under six months of tenure it dropped about 10 percentage points against a 25 percent baseline, a result the authors treat cautiously because they could not include agent fixed effects and selection into AI access could bias it upward.
Now set that beside the Danish result at the top of this entry. Fourteen percent there, 2.8 percent here. Both are careful and neither is wrong, because they are not measuring the same quantity. Brynjolfsson and colleagues measure resolutions per hour inside one firm that embedded a tuned assistant into a stable product with recurring technical questions. Humlum and Vestergaard measure self-reported share of working hours saved across eleven occupations and thousands of employers, most of whom had deployed nothing more specialised than a general chatbot. Narrow task, heavy integration, high number. Broad economy, light integration, low number.
The authors say plainly that their paper is not designed to shed light on the aggregate employment or wage effects of generative AI tools, because wage and labour demand data were not available to them. They also note that their setting favours the tool: a stable product and repeating technical questions. Where a product changes quickly, they suggest, training material goes stale and the assistant could hinder rather than help.
For a small business the usable reading is narrow, and I would rather state it narrowly than inflate it. Where your questions repeat and the answers live in documents you control, a support assistant has produced large gains for the least experienced staff and roughly none for the most experienced, alongside fewer escalations and lower turnover. Everything past that line — whether it works on your catalogue, your customers, your margins — is a question your own records can answer and no published study can.
One question
Which three questions do you answer most often?
Both studies agree that repetition is where this works at all. Reply and tell me your three, and I will tell you honestly whether the evidence supports automating any of them.
More in this pillar as it's written.
— Shana
Sources link to the original publisher, never to a summary.