A company cuts its corporate officer layer. AI is doing the work now, the thinking goes, “We don’t need as many decision-makers.” Costs come down. Shareholders cheer.
Then the legal letters arrive.
The people are gone. Accountability isn’t. Under Delaware law and the EU’s upcoming AI Act, fiduciary duty doesn’t disappear when the org chart gets flatter. It gets concentrated in whoever’s still around.
Mark Stouse and I got into this during our recent chat on The Causal GTM Leader. We came in planning to talk about what AI actually costs when companies treat it as free labor. We ended up somewhere thornier: what it costs when companies mistake efficiency for effectiveness. And why that mistake is starting to have legal consequences.
When OpenAI launched publicly, Mark did a word cloud of their launch materials. Efficiency was front and center. In business, efficiency has one meaning: cut costs.
The problem is that efficiency is a derivative metric. It tells you how cheaply you’re doing something. It says nothing about whether that something is worth doing.
You have to be effective first for any cost burden to be relevant, to be acceptable.
That’s not a philosophical point. It’s a decision-making sequence. If you don’t know whether a function, a role, or a workflow is producing commercial outcomes, cutting its cost doesn’t save you anything. It just makes a broken thing run cheaper.
There’s an economic frame that fits here: the Jevons Paradox. When you make something cheaper, and there’s already demand for it, consumption expands until total cost rises. You make content generation cheaper, so you generate more. Your total marketing cost goes up while output quality diffuses. More activity. No more effectiveness.
Gerard Pietrykiewicz and I made the same case from a different angle in AI Agents Are Cheaper Than Unmanaged Work, Not People. The per-prompt cost may be low, but the workflow cost is where it gets away from you.
This is the pattern GTM teams are running into right now. The tool can produce more, so the focus becomes MORE, rather than using the tool to make better decisions. The same argument is at the centre of The Decision Layer: effectiveness has to come before scale, or you’re just running the wrong system faster.
Large companies installed automated applicant tracking systems (ATS) to reduce recruiting costs. Fewer recruiters, faster screening, lower cost-per-hire. Reasonable on the surface.
The problem is the human configuring the system’s keyword filters and screening criteria. Research compiled by Select Software Reviews found that 88% of employers believe they are losing highly qualified candidates who are screened out because they didn’t submit ATS-friendly resumes. In other words, the system’s filters weren’t calibrated to what the role actually required.
Mark made it even more concrete. His wife, an ex-Accenture change-management specialist, applied for a role at a major consulting firm. Her CV was rejected by the ATS. When she connected with someone inside the firm later, they told her, “Oh, that happens all the time, and you’re perfect for this job.”
The talent you didn’t hire is invisible. The savings are on the ledger. That asymmetry shapes every one of these decisions.
If you are spending less money and it’s not effective, it doesn’t matter. You should be spending zero money if it’s not effective.
The ATS example is about recruiting. But the logic applies wherever AI is substituting for human judgment without a validated effectiveness baseline.
Automate what’s proven. Don’t automate the question of whether it’s working.
This is where our conversation went somewhere most AI-and-headcount pieces don’t.
Mark referenced a company (which shall remain unnamed) that significantly cut its corporate officer layer based on the AI efficiency argument. Shareholders have since brought legal action. Those officers held fiduciary duties. When they left, those duties didn’t leave with them.
The more you replace people with bots, the more you’re concentrating liability in the hands of a fewer and fewer number of people.
This is grounded in real law. Delaware’s 2022 amendment to the General Corporation Law affirmed that corporate officers have a duty of oversight and that officers of Delaware-domiciled companies can be held personally liable for negligence, not just bad faith. That covers roughly two-thirds of the Fortune 1000 and 90% of venture-backed companies in the US. As Mark put it in an earlier session: Saying “I didn’t know” won’t protect you.
The EU is moving in the same direction. Under Article 26 of the EU AI Act, deployers of high-risk AI systems — which explicitly includes employment and recruiting tools — are responsible for human oversight, monitoring, and audit logs. That obligation doesn’t sit with the vendor who built the system. It sits with the organization deploying it. Full enforcement kicks in August 2026.
A machine cannot bear legal accountability for a bad outcome. That’s not a limitation of current AI. It’s a structural fact. Someone always owns the decision. Remove the people, and the accountability doesn’t disappear. It redistributes upward to whoever’s left.
Removing the people does not actually remove the responsibility.
The efficiency argument for AI is easy to make. The numbers are visible. The savings are immediate. What’s harder to see is the cost of automating something that wasn’t working in the first place, and the exposure created when the accountability chain has a gap in it.
Before any AI-driven headcount or workflow decision, two questions are worth writing down.
1. Can you prove what’s effective here?
Not what the dashboard shows — what you can causally connect to a commercial outcome. If the answer is no, you’re not ready to automate. Keep investigating.
2. Who owns the decision if it fails?
For every workflow or role you’re considering automating, name the person who remains accountable if the system produces a bad result. If that name isn’t clear before the change, the change isn’t ready.
The decision layer doesn’t transfer to the tool. It can’t. Someone always owns the outcome.
Missed the session? Watch it here.
If you like this content, here are some more ways I can help:
Cheers!
This article is AC-A and published on LinkedIn. Join the conversation!
