It’s a crisp morning at an AI conference. You’re in the front row of a session titled “The Death of the SDR.” The speaker, a VP of Sales at a hot AI company, clicks to a slide showing pipeline up 60%, SDR headcount flat. The room leans in. “We deployed an AI SDR six months ago,” she says. “It doesn’t burn out. It doesn’t have a bad quarter. It doesn’t resign to go work at a competitor. The variance is gone.”
The room nods. You nod. It makes intuitive sense: if the scaling problem is human variance, remove the humans.
But is the variance actually gone? Or has it simply moved somewhere you’re not looking yet?
Welcome back to “the Maths Behind the Business” (formerly “Maths meet B2B SaaS”), the series where I take a mathematical theorem — usually one that lives in a textbook nobody particularly enjoyed — and show how it has been quietly governing your business all along. Like the laws of physics, these principles operate whether or not you’ve named them.
In a previous article, I argued that sales team performance regresses to the mean as you scale: your founding SDRs and AEs are statistical outliers, and the act of scaling your team mathematically guarantees a drift toward average. That article got brought up a few times recently, as founders started deploying AI SDRs as a structural fix. Which is exactly why I wanted to write this follow-up.
A brief return to Galton’s garden
Regression to the mean was first described by Francis Galton in the 1880s. He was studying something entirely unrelated to software sales: the height of parents and their children. What he found was that very tall parents tended to have children who were tall — but not quite as tall. Very short parents had children who were short — but not quite as short. Over generations, heights drifted back toward the population average.
This wasn’t because something mysterious was pulling everyone toward mediocrity. It was because extreme outcomes are, in part, the product of conditions that are unlikely to repeat perfectly — lucky timing, a particular alignment of circumstances, a stretch of exceptional effort. When those conditions don’t recur in exactly the same way, the result moves back toward the mean.
The mathematical expression for this is straightforward. If we denote an individual’s observed performance as X and the true underlying mean of the population as μ, then an extreme observation X₁ that sits far above μ will, on average, be followed by an observation X₂ that is closer to μ. The more extreme X₁ was, the more pronounced the reversion.
In sales teams: your founding SDRs are your X₁. They self-selected into an unproven company with no playbook, took below-market comp, and figured it out. That is a very specific personality profile, and it is not the profile of the average person who applies to an SDR role once your company has a careers page, a structured onboarding process, and a recognisable name. X₂ is the broader hiring pool. And X₂ is, by definition, closer to the mean.
So, what happens when you replace the human SDR with an AI?
The AI SDR pitch is, at its core, a regression-to-the-mean solution. Remove the human, remove the variance, keep the X₁ performance permanently. And on the narrow question of consistency of execution, the vendors are not wrong — most AI SDR platforms today categorise reply signals, run A/B tests on messaging variants, adapt cadence timing based on engagement, and refine sequences based on what’s working. The execution layer genuinely does not have bad Mondays.
But here is what makes regression to the mean such a stubborn principle: it acts on any system that relies on a distribution of inputs, not just on the individual performing the task. 🤓
Removing human variance from execution does not remove variance from the strategic inputs the execution depends on. And it does nothing whatsoever about the environment the execution is happening inside.
So the regression doesn’t disappear. It makes two moves. One immediate, one slow-building. And together, they are arguably more expensive than the original problem.
First move: the bottleneck migrates downstream
The better AI SDR platforms do learn. They analyse reply patterns, flag recurring objections, and optimise send timing based on engagement data. The execution layer is genuinely adaptive. But there is a ceiling on what it adapts within: the ICP definition, the value proposition, the messaging framework, the competitive positioning. These are inputs the AI works from, not problems it solves. The strategic envelope is drawn by humans. The AI operates inside it.
That distinction has a direct operational consequence.
When the envelope is well-drawn, AI SDRs deliver what vendors promise: sequences at scale, meetings booked, pipeline volume up. The Q1 board slide looks good. But the envelope is rarely perfect, and it degrades over time as markets shift — and the AI has no mechanism to flag when that’s happening. It will optimise execution of a stale brief just as efficiently as a sharp one.
So the problem doesn’t announce itself at the top of the funnel. It appears one layer down. AEs start noticing that meetings are coming in but qualified pipeline isn’t growing proportionally. Discovery calls are being used to disqualify leads that should never have been booked. Deal cycles lengthen because the buying intent behind the meeting was always softer than the calendar invite suggested.
The bottleneck hasn’t been solved. It has migrated from lead generation — the cheapest part of the funnel to run at scale — into conversion, where it now consumes your most expensive GTM resource: AE time. And because the volume metrics at the top still look healthy, the migration can go undetected for quarters.
If you’re finding that:
- Meeting volume is up but pipeline-to-close conversion is softening
- AEs are spending a growing proportion of their discovery calls on fundamental qualification rather than value selling
- Your average sales cycle has lengthened since deploying the AI SDR
- The deals that are closing are disproportionately inbound or warm, while AI-sourced meetings are stalling at the same stage
Then it is worth asking whether you have solved a lead generation problem or simply relocated it to a part of the funnel where it is harder to see and more costly to address.
Second move: the whole channel regresses
The first move is a company-level problem. The second is a market-level one, and it compounds over time.
Cold email reply rates have fallen from 8.5% in 2019 to 3.43% in 2026. That is a decline of nearly 60% in under a decade, and AI-generated outreach at scale is explicitly cited alongside inbox saturation and tightening spam filters as a structural driver. Enterprise buyers currently receive between 15 and 40 cold emails per week. And adoption of AI SDRs is still broadening.
The mechanism is the same one that governs every outbound channel. When a new approach opens up, the early movers see genuinely exceptional returns — high open rates, high reply rates, pipeline at a fraction of the cost of a human team. Those returns are outlier performance. They are X₁. They are partly attributable to the novelty of the approach and partly to the fact that early adopters are operating in an uncrowded space.
As the tooling democratises and adoption widens, every company optimising against the same channel simultaneously degrades the signal-to-noise ratio for all of them. The mean response rate falls. The market regresses.
What AI SDRs change is not the existence of this dynamic but the speed at which it plays out. Because AI enables every company to reach the same volume floor almost simultaneously, the saturation that previously took years now takes quarters. The Poisson process of market adoption that I described in the first edition of this series — where new categories take time to develop and that timing is essentially memoryless — applies here too. The early movers’ head start is shorter than they think, because the tools that gave them the head start are available to everyone else almost immediately.
If you’re finding that:
- Your AI SDR’s reply rates were stronger in the first two quarters of deployment than they are now, without obvious changes to targeting or messaging
- Deliverability has become an increasingly active maintenance task — domain rotation, inbox warming, bounce rate monitoring
- Prospects are replying to point out that your “personalised” email is clearly automated
- The benchmarks your vendor quoted at onboarding no longer match your actuals
Then you are observing market-level regression in real time. The mean has moved, and it will continue to move in the same direction as adoption continues to broaden.
So what do you actually do with this?
Regression to the mean has always had the same practical implication: it tells you where and when to intervene, not whether to abandon the approach entirely.
- Measure qualified pipeline rate, rather than meetings booked. Meetings booked is a top-of-funnel proxy that can look healthy long after the underlying quality has deteriorated. The metric that exposes the bottleneck migration is the proportion of AI-sourced meetings that progress past discovery. If that number is softening while volume is holding, you have found where the regression went.
- Treat your setup as a living brief. The sequences, ICP criteria, and messaging logic you deploy at launch reflect your sharpest thinking at that moment — but the market keeps moving. What worked in Q1 may be stale by Q3, not because it was wrong, but because your buyers, competitors, and channel dynamics have shifted. Build in deliberate reviews every two to three quarters to stress-test whether your ICP assumptions still hold and update the messaging to match. This is also where human SDRs earn back their place in the loop: for the interpretive work of noticing when something that used to land has quietly stopped working.
- Put humans in at the point where signal is richest. The more sophisticated vendors have effectively admitted where AI breaks down through their own product roadmaps by building approval checkpoints, tiered escalation logic, and testing frameworks to manage the boundary between what the AI should and should not handle. The core issue is that the AI doesn’t always flag its own uncertainty. A novel objection, an unusual use case, a soft buying signal. The system routes them to the nearest known category and optimises confidently from there. A prospect who raises an unusual objection and gets a thoughtful human response has a qualitatively different experience than one whose reply gets auto-categorised and triggers the next sequence step. A recent study in Industrial Marketing Management found that human salespeople outperformed autonomous AI agents in B2B inbound contexts on both relational and financial outcomes — not because of execution quality, but because of the connection buyers form with a person that they don’t form with an agent. The practical implication isn’t to pull back on AI at the top of the funnel. It’s to be deliberate about where the handoff and interference happens.
- Ask for the distribution, not the case study. When a vendor presents their headline AI SDR results, they are presenting their X₁ — top decile customers in ideal conditions during the early adopter phase of their own platform. Regression to the mean is a property of populations. The average deployment looks like the median. If the distribution isn’t forthcoming, that is itself informative. If they do share the distribution, use it to set your own expectations honestly. Find the median, not the mean — vendors will often report averages that are skewed by their best accounts. Then locate where you are likely to sit: your industry, your ACV, your ICP breadth, your team’s capacity to maintain the strategic inputs. If your realistic scenario lands in the bottom two quartiles of their distribution, that is not a reason to walk away, but it is a reason to reprice the ROI and sequence the investment differently.
The insight hasn’t changed since Galton was measuring the heights of Victorian people: extreme outcomes are not the natural resting state of a system. They are the product of conditions that are unlikely to persist indefinitely. AI SDRs change the mechanism by which your top of funnel operates. The mathematics of how performance evolves — at the company level and across the market — remains in force regardless.
Enjoyed this? I’m always happy to discuss math theorems and how they are applied to the day-to-day business world. Find me on LinkedIn or drop me a line at gc@oxx.vc.
For further reading on regression to the mean and its implications:
- Kahneman, D. — Thinking, Fast and Slow, Chapter 17: “Regression to Mediocrity”
- Barnett, A.G., van der Pols, J.C., Dobson, A.J. — “Regression to the mean: what it is and how to deal with it” (International Journal of Epidemiology, 2005)
- The original post on sales team regression in my series on the maths behind the business.
