NO. 02 · THE SYSTEMS
Agents Are Not Headcount.
The most common question about AI agents is how many people they replace. It is the wrong question, and firms that ask it build the wrong systems.
The wrong question
When an executive asks how many people an agent replaces, they are pricing AI as cheaper labor. Cheaper labor is a margin story, and margin stories cap out fast: the best case is your current business, slightly less expensive.
The firms compounding with AI ask a different question: what work was economically impossible at human cost that is now trivial at machine cost? Aethelgard’s sourcing agent does not replace an analyst. It reads the entire deal universe every night, work no analyst headcount could ever have covered. The return is not a salary saved; it is a capability that did not previously exist.
Where agents pay back
Three conditions, all required. The task is narrow: a job description you could write for a competent temp, not “do our marketing.” The task is repetitive judgment: the same decision shape, hundreds of times, where rules almost work but not quite. And the task is instrumented: you can measure whether the agent did the job, automatically, without asking it.
Follow-up triage. Lead enrichment and scoring. Invoice matching. Anomaly escalation. Pre-screen qualification. Status reporting. None of these is glamorous. All of them are running in production for our clients right now, paying back monthly, which is more than can be said for most of what is on stage at AI conferences.
Where they don’t
Agents fail where the job description is vague, the failure modes are unbounded, or nobody will own the telemetry. They fail as “digital employees” given broad mandates and judged on vibes. They fail spectacularly when bolted onto a process that was already broken. Automation is an amplifier, and it amplifies dysfunction with perfect fidelity.
This is why our scans look at the process before the technology. An agent on top of a nine-hour follow-up culture simply generates apologies faster. Fix the system; then staff it with software.
The ownership test
One question separates an AI vendor from an AI partner: when the engagement ends, who owns the system? If the answer is a subscription you rent forever, at a price that mysteriously tracks your dependence on it, you have not bought leverage. You have leased a different kind of headcount.
Every agent we build is delivered as a system the client owns: the code, the prompts, the evaluation data, the dashboards. Our continued involvement is a choice the telemetry has to keep earning, not a hostage situation. We put this in writing because it is rare enough that saying it is a differentiator, which says something unflattering about the industry.
A buying checklist
Before commissioning any agent, demand five artifacts in writing: the job description, the measured leak it closes (with monthly cost), the evaluation method, the escalation path to a human, and the ownership terms. A builder who has all five is an engineer. A builder missing two is a demo. A builder offended by the question is a press release.
“Automation is an amplifier, and it amplifies dysfunction with perfect fidelity.”
FROM THE VAULT
Agentic Engineering
The field manual: job descriptions for agents, grounding architecture, evaluation harnesses, escalation design. Free, by email.
RELATED · NO DEAD ENDS