The Pipeline Problem Hiding Inside an AI Productivity Boom

Here is a number from a new working paper that stuck with us: the time needed to grow an expert inside a firm rises from about five years to nine. That is the paper's baseline scenario for what happens once AI absorbs the codified work that juniors used to cut their teeth on.

The paper appeared in the quantitative finance listings on arXiv this week, alongside sixteen other new submissions. Most of the batch deals with narrow technical questions, such as high-frequency trading in Shenzhen or next-day price forecasts for Chinese convertible bonds. Yet a few of the entries ask bigger questions about where AI is taking the economy, and those are the ones we think investors should read slowly.

Why does it matter now? Because the dominant story in markets is still about exposure: which jobs, which sectors, which earnings lines get hit. This paper asks something different. It asks what happens to the supply of people who know how to do the hard things, once the easy things stop being done by humans.

We should say up front that the authors treat their results as scenarios, not forecasts. Nobody has measured two key parameters. Still, the structure of the argument is worth understanding, because it points at a cost that does not show up in any quarterly report.

Juniors Fall First, Seniors Fall Further

This model extends earlier work from a single worker to a firm with two cohorts, juniors and seniors. Junior work is the most codified, so it is the easiest to automate. Demand for juniors drops first.

In the baseline calibration, junior demand bottoms out about a fifth below its 2022 level, then recovers most of that ground. That recovery sounds comforting. It is not the end of the story.

Later, but harder, the senior stock falls: down a quarter, and eventually down by half. Once you see it, the logic is plain. Seniors retire, and the firm normally replaces them by promoting people who spent years doing junior work. If software has taken over that work, the replacements never accumulate the practice.

Think of a hospital that automates all routine charting and basic imaging reads. Efficiency jumps. Ten years later, nobody on staff has seen enough ordinary cases to recognise the strange one. According to the paper, expertise forms through unassisted practice on work AI cannot do, and that practice used to arrive bundled with the work AI can do.

A common misconception is that a junior recovery means the problem solved itself. Juniors may come back, but if their days are filled with supervising tools rather than struggling with hard problems, the time to expertise still stretches. Headcount and capability can diverge.

Who Pays, and Who Decides

The most interesting piece of the model is about incentives. A manager with a three-year horizon picks AI use to maximise near-term output. In the paper's scenario, that choice cuts the wage bill by 21 percent.

Meanwhile the firm's capacity on work AI cannot do falls by two-fifths. Output on that work goes unserved once the supply of seniors runs thin. The manager who made the call has usually moved on by then.

This is a classic mismatch between the party that decides and the party that bears the cost. Corporate finance has seen versions of it before: underinvestment in maintenance, in training, in research. Each looks like a gain on the income statement until the bill arrives years later.

For an investor, the point is that reported margins in an AI-adoption year may overstate durable profitability. A 21 percent drop in labour cost is real money. If it comes with a hollowed-out bench of senior talent, part of that saving is borrowed from the future.

The paper proposes three indicators organisations can track: the unaided practice juniors get on work AI cannot do, senior time spent on such work, and the years it takes to develop an expert. None of these appears in a 10-K today. Some firms may begin to disclose versions of them, and analysts who ask about them early will have an advantage in sorting the careful adopters from the aggressive ones.

How Fast Could the Machines Improve Themselves?

A second paper in the batch tackles the other end of the AI question: whether progress could feed on itself. The author models two margins. Capability improvements let AI perform a larger share of AI research tasks, while efficiency improvements raise output per unit of compute on tasks already automated.

The finding is that a software-driven intelligence explosion needs both margins to improve fast enough, and neither can make up for the other. Calibrating to recent data, capability progress looks fast enough. Efficiency progress sits on a knife-edge.

In Monte Carlo simulations, an explosion occurs in 49 percent of draws. That is a coin flip, and the author is candid that this reflects uncertainty rather than a forecast. The rate of efficiency progress alone accounts for 45 percent of the uncertainty, more than all other parameters combined.

What do we make of this as investors? A 49 percent probability of a self-sustaining acceleration is not something to build a portfolio around in either direction. It does tell us that the range of outcomes for AI-linked earnings is unusually wide, and that better data on efficiency gains in frontier models would narrow it more than anything else.

Put the two papers side by side and a tension appears. One says the human expert pipeline thins out over a decade. The other says machine progress could either stall or take off. Both are scenario work, and both warn against confident point estimates.

Market Structure, Machine Learning, and Honest Benchmarks

Three other entries in the listings carry lessons for anyone who trades or evaluates quant research. The Shenzhen study builds an order-level proxy for high-frequency activity and finds that lower latency raises short-horizon price variation and adverse-selection exposure, while realised spreads decline and pricing error falls modestly. Speed has costs and benefits, and market design decides how they split.

China's T+1 rule, the absence of maker-taker rebates, and limits on rapid quote revision tilt speed advantages toward directional liquidity taking. That is different from the developed-market pattern, where fast traders more often supply liquidity. Investors holding emerging-market funds should not assume US-style microstructure abroad.

We call the convertible bond paper a reality check, and the label fits in spirit. It asks whether deep networks that look good in-sample beat a simple persistence forecast for next-day prices. Many published results never faced that benchmark.

The survey on uncertainty in financial foundation models makes a related point. Uncertainty arises at different stages: whether evidence is valid at the relevant time, whether an interpretation is justified, which outcomes remain plausible. Most studies address one stage and not the chain from evidence to action. If a vendor sells you an AI research tool, ask which of those uncertainties it measures.

What This Means for VEA, UNG, C, IEF, and QUAL

None of these papers names a ticker, so any portfolio link is our inference. Treat it as a way to organise thinking, not as a call.

Start with QUAL, the iShares MSCI USA Quality Factor ETF. Quality screens favour high return on equity, stable earnings and low leverage. If AI adoption inflates near-term margins through labour cuts, quality screens may reward firms whose profitability is partly borrowed from their talent bench. That is a risk worth monitoring, since backward-looking screens cannot see it.

Now C, Citigroup. Large banks employ big cohorts of junior analysts doing codified work: reconciliations, reports, first-pass reviews. They are natural early adopters and could show wage-bill savings quickly. The question of how seniors get trained applies directly, since senior risk and credit judgment is built from years of junior grind. A bank that cuts too deep may enjoy two good years and a weaker decade.

VEA, the developed-markets ex-US fund, offers diversification across labour regimes, with different hiring norms, apprenticeship traditions and regulation. Firms in some of those countries face stronger constraints on rapid workforce cuts, which could slow the near-term gains and also protect the training of future seniors. We would not bet on that, but it is a reason not to treat all developed-market equities as one AI trade.

IEF, the intermediate Treasury ETF, matters for a different reason. If a productivity boom lifts growth and rates, IEF suffers. If the transition produces weak junior hiring and softer consumer demand, IEF benefits. Its role as ballast is useful because the AI outcome range is wide.

UNG, the natural gas fund, connects through the data-centre power story. Compute demand lifts electricity use, and gas fills a lot of that demand. UNG is a futures-based product with roll costs that can erode returns even when the thesis is right, so we would treat it as a small tactical position at most.

Consider a simple scenario. Take a $100,000 portfolio split 40 percent QUAL, 15 percent VEA, 10 percent C, 30 percent IEF and 5 percent UNG. If the adoption-boom case plays out and long yields rise a point, IEF's roughly six-year duration implies a loss near 6 percent on that sleeve, about $1,800. A 10 percent rally in QUAL would add $4,000. The mix survives, but the bond sleeve does less hedging than people expect in that particular world.

Putting This to Work Without Overreacting

The practical question is what to do with papers that explicitly refuse to forecast. Our answer is to change the questions you ask, not necessarily the positions you hold.

When reading earnings calls from firms announcing AI-driven headcount reductions, listen for what they say about entry-level hiring and training. A company that keeps a deliberate apprenticeship track while automating is making a different trade-off from one that just cuts. Neither is automatically right, but they deserve different multiples.

On timing, the paper's own numbers argue for patience. Junior demand takes years to bottom out and seniors fall later still. Any damage from a thin pipeline would show up well beyond a typical three-year manager horizon, which means the market may not price it until it is visible in results. That favours a slow-moving watch list over a quick trade.

Implementation has real obstacles. The three indicators the paper proposes are not publicly reported, so you will rely on proxies: turnover commentary, graduate hiring announcements, training spend, and the age profile of senior staff where disclosed. Expect noisy data and avoid building a rule on a single quarter.

For funds, rebalancing discipline matters more than prediction. If QUAL has run well and now exceeds your target weight, trim it. Check that your IEF allocation still matches the duration risk you want. Keep UNG small enough that a bad roll year does not matter.

Three Habits Worth Keeping From This Week's Papers

The week's listings leave us with a view that is more about method than conclusions. AI transitions may impose costs on a delay, on a different party from the one that chose them, and in places that standard metrics do not capture.

Scenario thinking beats point forecasting here. A 49 percent chance of an intelligence explosion and a nine-year expert pipeline are both reminders that the honest range is wide. Portfolios built to survive several outcomes will do better than portfolios built to win one.

Here is what you can do this week:

- We suggest you review your equity holdings for heavy exposure to firms that rely on large junior cohorts doing codified work, and read their latest commentary on hiring and training. - Check whether your quality-factor exposure through QUAL is concentrated in sectors where margin gains from automation may be flattering the screen. - Confirm that your bond sleeve, whether IEF or something similar, matches the rate scenario you are worried about. - Keep any UNG position small and treat it as a tactical bet, not a core holding. - When any research tool or fund claims AI-driven alpha, ask what it was benchmarked against and which stages of uncertainty it measures.

None of this requires predicting whether the explosion happens. It requires noticing that the costs and benefits of automation land at different times and on different people, and that markets are usually slow to price the later ones.