Automation was supposed to lighten the load. In revenue teams, it’s quietly moving the load somewhere heavier — and nobody’s measuring it.
There’s a story we tell ourselves about AI in revenue work. You plug in the tools, the busywork evaporates, and your team gets to spend its time on the high-value, human stuff. More pipeline, less grind. Easy bonuses all around.
I believed a version of that for a while. Then I started paying attention to the people actually running these systems — the RevOps leader babysitting the lead-routing logic, the SDR manager reviewing hundreds of AI-drafted emails, the marketer who now “just edits” the AI’s first draft. None of them are working less. Several of them are working noticeably harder, and a few of them look exhausted in a way I haven’t seen before.
So, I’ve been working through a theory, and it’s a bit uncomfortable:
AI doesn’t remove the work. It moves the work up the stack, onto fewer people, and into a kind of effort we’re not trained to see and are even worse at measuring.
I want to be careful here, lest I come off as anti-AI. I believe the efficiency gains are real. The point is that the gains come with unrecognized cost, with the bill landing firmly on the human in the loop.
The Work Doesn’t Magically Disappear
So, what actually happens to a task when AI takes it over?
Before, much of your work was generative: Writing the email, building the list, drafting the recap. You made the thing from scratch.
Now, it’s no longer about “making”; it’s more about judging. The AI takes over the generative part, and makes the thing, and your job becomes deciding whether the thing that it’s made is any good, fixing what isn’t, and owning the result.
A 2025 Microsoft Research study of 319 knowledge workers found that generative AI shifts the nature of the work toward information verification, response integration, and task stewardship. From producing the work to overseeing it (Lee et al., Microsoft Research / CHI ’25).
Many will say this sounds like a promotion, and, that’s at least partially true. Verification and stewardship are higher-order skills than first-drafting. But, as with a promotion, there’s a catch: judging is often cognitively more difficult than doing. When you write an email yourself, you only have to be as good as you are. When you review an AI’s work product, you must be good enough to catch what it got subtly wrong. A misread of the account, the confident-but-stale stat, the tone that’s feels about 10% off can cost you real relationships. Nuance is a very high bar.
So, you’re no longer doing the easy 80%, yet you’re permanently on the hook for it, plus the hard 20%, over and over, all day, every day.
BCG’s 2026 workforce research lands in the same place: As repetitive tasks get automated, the remaining work concentrates in problem-solving, decision-making, and judgment — and cognitive load intensifies (BCG, 2026). The easy parts leave. The hard parts stay and compound. That’s the first half of the paradox.
The Human in the Loop Is the New Bottleneck
Here’s where it gets structural. When the AI is fast and the human is the reviewer, the human becomes the constraint on the entire system.
While the machine can draft a thousand emails in a minute, it cannot decide whether sending them is a good idea. That decision (and, most importantly, the accountability for it) still lays with a human.
We like to call that person a “safety net.” That’s the wrong metaphor. A safety net is passive; it just hangs there in case something falls. The human in an AI revenue workflow is doing continuous, active labour: Monitoring outputs, catching errors, deciding what escalates, absorbing the blame when something goes wrong. They’re not the net. They’re the bottleneck the whole throughput depends on. Problem is, we’ve been staffing that role as if it were the easy part of the job.
Customer-service research has already flagged the failure mode: Continuous monitoring produces oversight fatigue. Agents wear down from constantly deciding whether to trust the AI, and vigilance degrades exactly when volume is highest (Cobbai, 2026). There’s an older, crueler version of this from automation research. Decades ago, researchers described the irony of automation: When you automate the routine work and leave only exception-handling to the human, you strip away the everyday practice that kept their judgment sharp, so the person is least prepared at the exact moment you need them most (cited in Lee et al., 2025).
For revenue teams this is not abstract. The SDR who only reviews AI drafts stops developing the instinct for what a great cold email feels like. The RevOps analyst who lets the model build every report stops noticing when the numbers smell wrong.
We are gradually conditioning out the work that built the intuition that makes us good judges.
The bill comes due the day the AI is confidently, catastrophically incorrect and the human who was supposed to catch it has quietly lost the judgement to do so.
The Gains Are Real. The Load Is Just Hidden.
I want to be fair to the optimists, the productivity numbers are genuinely good.
BCG found consultants using generative AI completed tasks 25% faster and produced work rated 40% higher in quality (via SERVSIG, 2026). Those gains are real and worth having. But the same body of research keeps surfacing a second finding that doesn’t make it into the keynote: the gains don’t scale linearly, and past a point they reverse.
“I was working harder to manage the tools than to actually solve the problem.” — an engineer in BCG’s study
The pattern in the data: Productivity gains flatten and then decline after roughly three simultaneous AI tools, as the overhead of managing the tools starts to outweigh what they produce (SVA Consulting, 2026).
Let’s look at a modern “AI-enabled” revenue stack: You have an AI SDR tool, an AI copywriting assistant, an AI meeting notetaker, an AI forecasting layer, an AI enrichment service. You’re already well past three. Each one was sold as a time-saver. Collectively they can produce a person whose whole day is tool-tending.
The reason this stays invisible is that we measure the wrong side of the ledger. Dashboards count outputs: Emails sent, meetings booked, reports generated, tickets closed. They don’t count the mental cost of producing those outputs. Time and labour demands go down, which is what the metrics see. Mental demand, monitoring pressure, and trust-related stress go up. The load didn’t vanish. It moved into a column nobody’s looking at.
The Trap Inside the Trap: Confidence
There’s one more turn of the screw, and it’s the most counterintuitive part. The better the AI gets, the more dangerous the load becomes. Because good AI lulls the human out of the very vigilance the system depends on.
The Microsoft study found a clean, slightly chilling relationship: Higher confidence in the AI was associated with less critical thinking, while higher confidence in oneself was associated with more (Lee et al., 2025).
Read that again. The more you trust the tool, the less you scrutinize it. So as your AI gets reliable enough to earn trust, your team’s guard drops; just as the rare errors become the expensive ones precisely because no one’s looking anymore.
This is why “just add a human in the loop” is not a strategy. A disengaged human in the loop is worse than being honest about the risk, because they provide the appearance of oversight without the substance. The hard part of revenue AI isn’t getting the model to produce. It’s keeping a human meaningfully engaged with output they’ve been conditioned to wave through.
So What Do You Actually Do About It?
None of this is an argument for ripping the tools out. It’s an argument for treating the human load as a real cost you design around, the same way you’d design around any other constraint. A few things I’ve come to believe matter:
- Set a span of oversight, not just a span of control. You wouldn’t ask one manager to supervise 40 people. So, don’t ask one person to meaningfully oversee eight AI tools. Audit how many systems each role is reasonably expected to properly monitor, and cap it.
- Embed AI in shared workflows, not stacked on individuals. The research is consistent that cognitive burden drops measurably when AI is woven into a team process with clear ownership at each step, rather than piled onto one contributor juggling the whole process (SVA, 2026).
- Assign individual accountability for oversight. Without a single person owning the oversight and outcome (rather than the composite process steps), individuals’ responsibility for scrutinizing their AI’s output will entropy. Named ownership for monitoring a given workflow sharply reduces the odds an error slides through (SERVSIG, 2026)
- Protect the judgment muscle on purpose. If your reviewers never do the underlying work anymore, build in deliberate reps (i.e. no skipping leg day). Task rotation, building from scratch, calibration sessions, etc., so the skills you’re relying on to catch errors don’t atrophy.
- Measure the hidden column. If your only metrics are outputs, you are flying blind on the cost side. Ask the people in the loop how the work feels, not just how much of it shipped. Oversight fatigue shows up in conversation long before it shows up in a dashboard.
Where I’ve Landed (For Now)
The seductive promise of AI in revenue work is that it lets you do more with less. That’s true…
But it’s a dirty truth. It turns out, the “less” isn’t free. You’re trading a large amount of easy, distributed, generative work for a smaller amount of hard, concentrated, evaluative work. Dropping that harder work onto fewer people, who are now permanently on the hook for everything the machine gets wrong, is creating a cognitive bottleneck.
The efficiency gain is real. The cognitive bill is also real. The teams that win won’t be the ones that ignore the bill — they’ll be the ones that design for it.
Done well, this is genuinely a better way to work: Humans freed from the repetitive grind, doing the judgment-heavy work that actually needs their intuition and experience. Done haphazardly (Tools stacked on individuals, oversight treated as a free safety net, the load left unmeasured) you get fried people, hidden errors, and a team that’s seemingly more productive, yet quietly falling apart.
The difference between those two outcomes isn’t the technology. It’s about being aware of the human load, understanding what to look for, and engineering around the pitfalls.
What I’m Still Noodling On
This is a working theory, and I'm looking for some pushback. A few things I keep circling:
- Where have you seen AI genuinely reduce your load, and where did it just relocate the work onto you in a form that’s harder to see?
- Is there a tool-count number where your own AI stack tipped from “helpful” to “management burden”?
- How do you keep a reviewer genuinely engaged with output they’ve been trained to trust, without making the review so heavy it kills the efficiency gain?
- Whose job is it to own the hidden cost — the IC drowning in it, the manager who deployed the tools, or RevOps who designed the workflow?
If you have a counter-example, a sharper framing, or a story that breaks my hypotheses, I want to hear it.
Until then, I’ll keep experimenting and learning.
