On 20 October 2026, Keith Richman shared a clip of Josh Elman on X. He set it against the flashy claims people make about personal AI agents: "It booked my hotel room!" "It cancelled my subscription!" Elman's framing is quieter, and we think more useful. Richman put it in one line: "do the busywork, then leave the decision to us."
That sentence works as a design rule for agent workflows. If you are setting up agents in a marketing or ops team, it tells you where the agent stops and where a named person takes over. This note is about drawing that line in practice.
Busywork and decisions
Busywork is collecting material, drafting, checking output against a fixed list, reformatting and routing things to the right place. An agent that reads five sources and writes a summary is doing busywork. So is an agent that pulls data into a report template or checks a draft against a style guide.
A decision is anything that gets published, gets sent, spends money, or judges quality or truth. Once an output leaves the company, reaches a customer, moves budget or declares something correct, a person should own it.
As a rule of thumb, agents prepare and people decide. If a step produces options, evidence or a draft, an agent can own it. If a step concludes something, a person owns it.
Why judging stays outside the agent
The first rule has a second one inside it: never let an agent grade its own work, and be careful about letting AI judge work at all.
Two examples show the risk. On 21 October 2026, Jenny Lorraine Nielsen reported on X that the AI detector Pangram scored papers she wrote in 2016 and earlier as 100% AI generated. Her conclusion: "Do not trust AI to evaluate whether something is written by AI." That is one person's report, not a controlled study. Research points the same way, though. Liang and colleagues showed in 2023 (arXiv 2304.02819) that several widely used GPT detectors consistently misclassified writing by non-native English writers as AI-generated, while they identified native English samples accurately.
We saw a smaller version of this in our own work. We ran three open-source marketing skills against our own site and wrote the test up in a separate note. The skills found a real problem with our offer. But each "this may be broken" guess, first an image path and then a contact page, turned out to be wrong once a simple fetch checked it. The agents raised the flags, and the person checking decided what was true. We needed both steps.
In all three cases the weak point was the judgment. So keep the verdict with a person, or at least with a checker that is separate from the agent that did the work.
Maker, skeptic, person
On 22 October 2026, @shmidtqq described a research setup on X that separates these roles. Five agents research one question from five angles in parallel. A sixth, read-only agent acts as a skeptic and checks each finding: is it correct, is it current, does the source exist? A writer agent ranks what survives. The person is the last step. In the author's words: "nothing ships without your yes."
The skeptic in that setup does not decide what ships. It tests findings against checkable criteria and passes on the ones that survive. The final call needs context and someone accountable, so it stays with the person.
Our note "What is agentic marketing?" describes the full control model we build into workflows, so we will not repeat it here. This note makes a narrower point: the checking role has to be separate from the doing role, and the deciding role stays with a person.
The decision inventory
Before you automate anything, walk through one workflow from start to finish and mark every step as busywork or a decision. A content research workflow might look like this:
- Pull sources on a topic. Busywork.
- Summarize each source. Busywork.
- Verify each claim: does the source exist, does it say what the summary says, is it current? Busywork, done by a separate checker agent or tool, never by the agent that wrote the summaries.
- Rank the findings by relevance to your audience. Borderline. An agent can propose the ranking, and a person confirms it, because relevance depends on context the agent does not have.
- Approve the brief for production. Decision, with a named owner.
- Publish or send. Decision, with a named owner, behind an approval gate.
Three rules come out of this exercise.
Every decision step gets a named owner. Write down a person, because a step that belongs to "the team" tends to end up as the agent's call.
The maker and the checker are different. The agent that produced the work does not evaluate it. Use a second agent, a different tool, or a deterministic check such as the fetch in our site test.
The approver gets evidence along with the verdict. When something reaches a person for approval, they should see the sources, the fetch result and the log entry, and a green checkmark alone is not enough. If the approver can only click "yes", the gate becomes a rubber stamp.
Where to start
Pick one workflow you already run and do the decision inventory on it this week. Hand the busywork steps to agents with a separate checker, put a name on every decision, and give that person the evidence to decide quickly. That line works as the test for each step: does this step do the work, or does it make the call?