2017: a radiologist answers his own profession
The earliest version we found comes from radiology. Its author is Curtis Langlotz, a radiologist at Stanford.
In 2017 radiologists argued a lot about whether AI would replace them. Langlotz answered with a tweet, and in 2019 he expanded the idea in an editorial, “Will Artificial Intelligence Replace Radiologists?”, in the journal Radiology: Artificial Intelligence. RSNA News gives its closing line as: “Radiologists who use AI will replace radiologists who don’t.”
Langlotz himself said the line started as a 2017 tweet, in an interview with ITN in January 2024. In the same interview he called the fears about radiologists being replaced unfounded.
The scope of this version is narrow. It is about one profession, and the replacement happens inside it: radiologist against radiologist. The profession itself stays where it is.
May 2023: an economist says “your job”
Six years later a similar idea was voiced outside medicine. The economist Richard Baldwin, a professor at the Geneva Graduate Institute, said it at the World Economic Forum’s Growth Summit in May 2023.
Baldwin’s words are known from a Business Insider piece republished by Yahoo: “AI won’t take your job,” and then “It’s somebody using AI that will take your job.” In the original the two sentences are split by the reporter’s note on who said it and where. This is a reporter’s record; we have no transcript of the session.
The addressee has changed. There are no radiologists in the phrase any more. There is “you”, which means any worker. “Replace” has become “take your job”. The scope is now the whole labour market. The headline of the same piece already hedges: someone who knows how to use AI “might” take it.
August 2023: the phrase becomes a headline
On 4 August 2023 Harvard Business Review published an interview with Karim Lakhani of Harvard Business School. The headline: “AI Won’t Replace Humans — But Humans With AI Will Replace Humans Without AI”.
The headline is easy to mistake for Lakhani’s own words. We could not check whether he said them verbatim: we do not have a full record of the conversation. The repost on the HBS AI Institute site carries the phrase only as the title, and quotes Lakhani separately, on other points.
This version has the widest scope. The phrase is about people in general.
Hence a rule we follow ourselves. Credit a quote to whoever the exact wording came from, and give the date. Langlotz’s line, Baldwin’s words as a reporter wrote them down, and the HBR headline are three different texts by three different authors. How to check a line like this in ten minutes is in our piece on predictions nobody actually made.
May 2025: Huang and “every job”
In May 2025, at the Milken Institute Global Conference, NVIDIA’s chief executive Jensen Huang repeated the idea. In the official transcript, at 06:40, he first says that every job will be affected: some jobs will be lost and some created. Then: “you’re not going to lose a job–your job to an AI, but you’re going to lose your job to somebody who uses AI.”
Two details show up only in the transcript. Huang corrects himself mid-sentence, from “a job” to “your job”. And he says “somebody”, while the popular retelling has “someone”.
Context changes the meaning. In Langlotz’s version the profession survives. In Huang’s, the phrase sits next to an admission that some jobs will disappear. He did not originate the idea: by 2025 it was eight years old.
What the data behind the phrase says
The data closest to the phrase comes from “Generative AI at Work” by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics in 2025.
The setup: 5,172 support agents at one Fortune 500 company that sells business-process software. An assistant built on GPT-3 watches the chat with the customer and suggests replies to the agent in real time. The agent stays responsible for the conversation and can ignore or rewrite a suggestion. The assistant was switched on for groups of agents in stages, so there was always a group to compare against.
What came out:
- on average, agents with the assistant resolve 15% more issues per hour;
- less skilled and less experienced agents gain 30%, and both their speed and their quality go up;
- the most experienced and skilled agents get a small gain in speed and a small drop in quality;
- an agent with two months of tenure and the assistant performs as well as an agent without it who has more than six months;
- customers write more politely and ask for a manager less often, and fewer newer agents quit.
If you have seen 14% and 34%, those figures are real too. They come from the NBER working paper of April 2023, revised in November, with a sample of 5,179 agents. The journal version has a slightly different sample and estimates.
The authors’ explanation: the assistant picks up the practices of the best agents and passes them on to novices. The study does not show who ends up replacing whom. The authors say plainly that their data cannot speak to employment or wages: in the longer run a firm may hire more novices, or it may set out to build AI that replaces people altogether.
What this means for a five-person sales team
Picture a team: one strong salesperson, two average ones, two novices. If the mechanism from the study carries over to sales, the novices gain first and the strong salesperson gains almost nothing. The practical question then is how quickly a novice starts answering customers the way the best person does. Who to cut is beside the point.
A caveat: the study covers support over text chat, and there is no selling in it. We carry over the mechanism. The numbers do not carry over.
The order of work that follows from this:
- Write down how your best salesperson answers the common questions: price, deadlines, objections, comparison with a competitor.
- Give the novice suggestions based on those answers, and leave the decision with the novice.
- Before you start, write down one metric: time to first reply, or the share of enquiries that reach a meeting.
- Do not move the strong salesperson onto working from suggestions. In the study, top agents followed the suggestions more and more over time, even though it slightly lowered the quality of their conversations. The authors warn that with fewer original contributions from the best workers, future versions of the model may handle new problems worse.
Our own products are built the same way. Valli, our auto-responder for Telegram Business and website widgets, is in pilot. It answers customers from a profile the owner fills in, and its rules allow prices and deadlines to come only from that profile. When the profile has no answer, it says it will check with the owner and offers to take a contact. In Telegram the owner gets a card about a hot customer and can mute Valli in that chat for 24 hours with one button, to take over the conversation. Radar collects public requests and qualifies them; the conversation with the customer stays with the manager. What effect this has on sales we do not know yet: we have no measurements.
What we don’t know
We did not open Langlotz’s tweet itself. We know about it from his own words in the 2024 interview; the date, 8 February 2017, comes from search results and the post’s ID.
We did not see the video of the WEF session. Baldwin’s words come from the Business Insider piece, which does not give the exact day of the panel. We did not check whether Lakhani said the HBR headline verbatim.
The Brynjolfsson study was done in one company, on one type of work, and describes a medium-run effect. We have no data on how AI assistants change the work of sales teams in Uzbekistan.
What an AI sales assistant does on its own and what it leaves to people is set out on Sales automation with AI.
If you want to see which step of your team’s work can go to an assistant and which is better left to people, describe the process in the quiz. We break down one process for free within 48 hours, with no intro call.
