AI makes juniors faster at answers a senior already knew

2026-09-28

If the tools are so good, can a junior employee using AI carry out what a senior employee did at half the cost?

• Ecommerce
• AI
• Leadership

Last week, four announcements occurred within three days of one another, and none of them concerned a chatbot. Kingfisher stated that the sales generated by its AI-powered personalisation amounted to about £100 million in the period in question, a rise of 16%. ASOS said that the product recommendations displayed within its search function now convert substantially better than its normal browsing paths, attributing this improvement to better digital product content which has also made the website faster and more consistent across different markets. Amazon has turned Seller Assistant into an always-on agent that monitors stock levels, prices and ratings and can take action on them. Morrisons has appointed a Chief AI Officer who reports directly to the chief executive.

This is the side of AI in ecommerce I find genuinely interesting. Catalogue enrichment, demand forecasting, repricing, monitoring thousands of SKUs at three in the morning. Work that used to be done slowly, or badly, or by nobody at all because there weren't enough hours for a person to look at every listing.

It also raises a question which I am being asked more and more by owners and CFOs, typically put in terms of cost savings. Since the tools are so good, do we still really need a senior employee to be in charge of trading, or of product data, or of pricing? Might a capable junior worker, using the appropriate AI system, be able to carry out the work for half the salary?

I don't think so. I want to be careful about why, because the research is less flattering to my position than most people who share it admit.

What the studies measure, and who comes out ahead

The strongest field evidence we have points the other way at first glance. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,172 customer support agents as a generative AI assistant was rolled out in stages. In the version published in the Quarterly Journal of Economics in 2025, issues resolved per hour rose 15% on average. Less experienced agents got faster and better. The most experienced got slightly faster and, on quality, slightly worse.

Software looks similar. Kevin Cui, Mert Demirer and colleagues pooled three randomised trials at Microsoft, Accenture and an anonymous Fortune 100 company, 4,867 developers in all, and found a 26.08% increase in completed tasks among those given an AI coding assistant. The standard error was 10.3%, so the true effect could sit some way either side of that figure. The paper, now published in Management Science, is clear that less experienced developers adopted the tool more and gained more from it.

So if you measure throughput, a junior with AI closes a lot of the gap. I'd rather say that plainly than pretend otherwise, because a finance director who reads these papers will say it for me.

Read past the headline, though, and both studies also show why it happens. Brynjolfsson and his co-authors suggest the assistant works by spreading the habits of the best agents to everyone else. The model learned from conversations handled by people who already knew how to fix the problem. A junior using it is being handed a senior's answer, compressed and delivered at speed. That is enormously useful, and it only works where a good answer already exists in the data.

In 1966 Michael Polanyi made the observation that we know more than we can express. The portion of a senior's knowledge that has been put into written form, such as tickets, lists, reports and previous decisions, is precisely the kind of information that a model can take in and redistribute. The implicit part of this knowledge is not included in the training data; for example, knowing that one supplier always ships short in November, or that the last time you matched a particular competitor on price you caused your customers to wait for it. A considerable amount of money in ecommerce trading is based on this second kind of knowledge. Performance figures such as the number of listings per hour or the number of tickets closed will go up regardless of the situation, which is why these metrics are deceptive.

The frontier you can't see from inside

The paper I'd put in front of any board considering the swap is by Fabrizio Dell'Acqua and colleagues, run with Boston Consulting Group. The experiment took place in 2023 on GPT-4 with 758 consultants and was published in Organization Science in March 2026. On 18 realistic tasks that sat within what the model could do, consultants using it completed 12.2% more tasks, finished 25.1% faster and produced work rated more than 40% higher in quality. On one task deliberately chosen to sit outside the model's capability, consultants with AI were 19 percentage points less likely to reach the correct answer than those working without it.

The detail that matters most for a trading team is that nobody was told which task was which. The authors call it a jagged frontier because tasks that look equally difficult to a person can fall on opposite sides of it, and the model gives no signal when it has crossed over. The wrong answers also read well. The published paper notes that AI made responses more persuasive and coherent on the very task where it made them less correct. That is processing fluency at work: text that is clear, confident and easy to read feels truer, whether or not it is.

Models have moved a long way since GPT-4, so the specific tasks on each side of the line will have changed. Whether the frontier smooths out as models improve, or simply moves while staying jagged, hasn't been properly studied yet, and I'd be wary of anyone who tells you confidently either way.

Here is how it plays out in replenishment. A forecasting model gives you a demand distribution for each SKU. Someone still has to decide how much stock to buy against it, and that decision comes from the costs, not the forecast. The newsvendor logic every planner learns sets the stocking level at a critical ratio: the cost of being short, divided by the cost of being short plus the cost of being over.

Take an illustrative seasonal knitwear line. Say each lost sale costs £30 of margin, and each unit left at the end of the season loses £12 once it has been marked down and cleared. The critical ratio is 30 divided by 42, about 0.71, so you'd buy to roughly the 71st percentile of forecast demand. A basic line that carries into next season has a much smaller overage cost, so its right answer sits far higher. Nothing in the forecast tells you which situation you're in. Those numbers come from knowing the category, the supplier terms, the markdown curve and what happens to the brand when a hero product sells out in week two.

A junior user of the tool will typically go with the default service level, usually around 95%, since this seems sensible. On the knitwear line, this means purchasing well beyond the point at which another unit should be included. The cost is then felt months later in the form of deeper discounts and a poorer full-price sell-through, long after the item has been linked to a setting that no one questioned in June.

Automation bias, and who catches the wrong answer

People have a well-documented tendency to accept a suggestion because a machine made it. It's called automation bias, and the cleanest recent measurement I know comes from a field with nothing to do with retail.

Thomas Dratsch and colleagues at University Hospital Cologne asked 27 radiologists to read mammograms with help from what they were told was an AI system, and published the results in Radiology in 2023. When the system's suggestion was correct, every group scored around 80%. When it was deliberately wrong, the least experienced readers fell to 19.8% correct, the moderately experienced to 24.8% and the very experienced to 45.5%.

It's only a small sample from a completely different area, so I'd base my conclusion on the general pattern rather than the precise figures. No one was completely immune, and experience more than doubled the chance of choosing the correct answer when faced with a confident incorrect one, this difference being most significant in the case of keeping a senior informed.

The way that Amazon has announced this reveals the way the platforms view the issue. The new Seller Assistant workflows allow a seller to choose whether the agent merely makes recommendations or actually takes action, with the seller setting limits, specifying approval steps and maintaining a full audit trail. This kind of design means that someone has to establish the guardrails. For a retailer the essential question is then who.

Imagine an agent who is instructed to remain competitive on price; if a competitor lowers one of their best-selling products by, for example, 15% for a weekend, the agent suggests matching it, the individual who has to approve the move sees a reasonable proposal with a clear explanation and then says yes. When this situation is repeated over a few hundred SKUs during a quarter, two outcomes occur. The gross margin is lost without anyone noticing any major individual decisions, and customers who keep encountering the reduced price start to anchor on it so that their internal reference price drops and the full price begins to seem like a markup; as a result, full-price conversions decline, which in turn causes the following round of recommendations to be set even lower.

A senior trader establishes a different type of limit, namely, a floor based on landed cost, maintains a short list of items where you never allow the market to dictate the price since the brand cannot afford to appear cheap, and has the rule that a competitor's weekend promotion does not count as a price change. Nothing about this is unusual; it just requires someone who has seen it go wrong in the past.

Catalogue entries have the same flaw. Although generative models are excellent at producing descriptions and assigning attributes on a large scale, Google now requires merchants to mark AI-generated titles and descriptions using a separate structured attribute. The problem is not obvious. When a model says that a wool blend is 100% merino or refers to a relaxed cut as slim, the item listed looks entirely normal and leads the customer to expect that the product will not live up to this description. The gap between the promise made and the actual product received is called expectancy disconfirmation, and in the fashion industry this typically ends in a return instead of a complaint. Individuals who are familiar with the category can detect the inconsistency within a few seconds, whereas those who are new to it have no reason to think there is anything wrong.

What employers are actually doing

The labour market data is interesting here, and it doesn't support the cost-saving story either. Seyed Mahdi Hosseini Maasoum and Guy Lichtinger analysed resume data covering 62 million US workers across 285,000 firms. After firms began actively integrating generative AI, junior employment fell relative to comparable firms that hadn't, while senior employment stayed broadly flat. The fall came mainly from hiring fewer juniors, with little change in people leaving.

Two cautions. The size of the effect differs between versions of the paper as the authors have revised it, so I'm not quoting a single number. And other researchers, including Brynjolfsson and colleagues in separate work, point out that entry-level hiring in exposed jobs had begun weakening before ChatGPT existed, so how much of this is AI and how much is the post-pandemic slowdown is still argued over.

Despite those qualifications, the trend is clear: the companies that are most actively using AI are keeping their senior staff and reducing the number of new applicants they take in below them. Whatever the spreadsheet indicates, their actions show that they regard judgement as the stage which cannot be automated.

A second paper from the same research group is worth reading if you plan to let juniors lead AI adoption. Katherine Kellogg and colleagues asked junior consultants how senior colleagues should manage the risks of generative AI. The juniors mostly proposed what the authors call novice risk mitigation tactics, grounded in a limited grasp of what the technology can and can't do, and aimed at changing individual habits instead of the system around the work. They were enthusiastic and fluent with the tools. They simply hadn't yet seen enough to know where it would fail.

There is also a longer-term problem that never makes it into the business case. If every business keeps its seniors and stops hiring juniors, the route that produces the next generation of seniors narrows. Ecommerce is a young discipline, and the bench was never deep to begin with.

How I'd actually set it up

The teams I'd build now look different from the ones I ran a few years ago, and AI is a big part of the reason. A senior who owns trading, product data or pricing can now cover far more ground, because the monitoring, drafting and first-pass analysis that used to fill a junior's week happens overnight. That's the real gain: range and speed for the person whose judgement you're paying for.

In reality, it is the senior who draws up the rules that the tools operate on, including the cost assumptions used in the replenishment model, the floor below every automatically set price, the list of attributes which an AI can fill in without human review and those it can only suggest, and the junior then works within these rules, checking the exceptions with the senior, thus passing on the unspoken aspects; the AI carries out the bulk of the work.

So long as you substitute the senior employee with a junior one using the same tools, the results will look all right for a time. The listings will be complete, the prices will be competitive, and the stock will be ordered on time too. It is only later that the problems show up elsewhere: in the number of markdowns, in the return rates, and in the fact that the customer base has come to expect the discount. By then, the original decision is rarely investigated.

AI is at its most effective when someone with real expertise is steering it and checking what it produces. Take that person out, and you still get answers. You just lose the ability to tell which of them are wrong.