Automation should follow judgment: what an AI may do on its own

An AI may act on its own when a step is cheap to get wrong and easy to undo: reading information and drafting text for review. Sending and changing records should wait for approval. Spending money and making commitments stay with a person. The test is the cost of a mistake and whether it can be reversed, not capability.

How do you decide what an AI may do on its own?

Ask two questions about each step, not about the system as a whole. What does it cost if this step is wrong? Can it be taken back? A step that fails quietly and can be reversed can run unattended. A step that reaches a customer or moves money cannot.

The idea is older than language models. In a 2000 paper in IEEE Transactions on Systems, Man, and Cybernetics (opens in a new tab)^[1], Parasuraman, Sheridan and Wickens proposed that automation can be applied to "four broad classes of functions": information acquisition, information analysis, decision and action selection, and action implementation. Each can sit anywhere from fully manual to fully automatic, and the authors list the costs of decision and action consequences among their secondary criteria for evaluating a design. NIST's AI Risk Management Framework (opens in a new tab)^[2] (version 1.0, January 2023) says human-AI configurations "can span from fully autonomous to fully manual", and that some systems need human oversight while others do not.

OpenAI's guide to building agents (opens in a new tab)^[3] offers a working rule. Rate each tool low, medium or high by "read-only vs. write access, reversibility, required account permissions, and financial impact", then pause for checks or escalate to a human before high-risk functions run.

Start from a trigger anyone on your team can read. "A contact changes Stage", "a visitor stops replying" and "every Monday at 9am" can be inspected and argued with. A trigger that fires whenever the AI senses a good moment hides a judgment inside the automation.

Which actions can run automatically, which need approval and which stay human?

Six kinds of action cover most business automations, ordered from easiest to hardest to take back.

Draft is the row most teams skip: the AI does the slow part while a person keeps the decision. Drafts still inherit the failure modes in why AI replies go wrong, so review them before they go out. In regulated fields such as financial advisory, expect more rows to sit with a person, and treat the split as a question for qualified review.

These defaults are a starting position, not a standard. Move a row only after you have read enough runs to know what the step does when it is wrong.

What does "keep judgment human" mean in practice?

Plandora's principles include "Keep judgment human." and "Automate with intent." The second adds that "Action should follow a clear decision." In practice the decision is made before the automation exists: you choose the trigger, the tools it may use and which rows of the table need a person. The automation then carries it out.

A case that leaves the script belongs with a person in the Inbox, which is the subject of human handoff for AI conversations. For the wider picture of how a Digital Mind turns expertise into action, see the expertise operating system.

How do Plandora Automations fit this framework?

An Automation is an agent that does a task when an event happens or at a time you choose. A Specialist is different: your Mind consults it in the middle of a live conversation, and it sees only the person it is talking to. An Automation can look across your whole audience, for example to find or count contacts, and that wider reach is a reason to give it narrower permissions.

Automations may be available through a guided rollout or self-serve, depending on your plan and rollout, so confirm what your account has before you plan around it. The Automations guide describes the setup like this:

• You choose exactly one trigger: an event (a new message, no reply, a new contact, a stage change, a file, an order or a link click) or a schedule from hourly to monthly.
• You describe the task and select Draft with AI, pick a template or start from scratch. A new Automation starts Off.
• Under Context you write its Goal, Role & style and Boundaries. Then you write the steps in order, typing @ to add information or tag your team and / to use a tool.
• The Tools card lists everything it can use: Contacts, Inbox, Message, Team and Store, plus integrations agents can use. Each is set to Read or Read / Write, and Message is Write because it sends a template. Drive and Social media are always on, and Web is optional.
• Runs shows what each run did, Improve suggests changes you accept or reject one by one, and a switch beside the status stops future runs.

The guide describes an approval choice for messages to several contacts. It does not describe an equivalent approval step for record changes, so count on the Read or Read / Write setting and your own review of Runs, and ask the team what your account adds.

Why isn't an approval step enough on its own?

People tend to trust what a system hands them. Goddard, Roudsari and Wyatt's 2012 systematic review (opens in a new tab)^[4] in the Journal of the American Medical Informatics Association defines automation bias as "the tendency to over-rely on automation". In a small pooled analysis of four healthcare studies that compared clinical decision-support groups with control groups, the risk ratio for an erroneous decision following incorrect advice was 1.26 (95% CI 1.11 to 1.44): the decision-support groups followed wrong advice more often than the control groups did. These were clinical settings, so read the figure as a direction, not a forecast for customer messaging. The authors list training and emphasising user accountability among the mitigators.

An approval queue that a reviewer clicks through is therefore oversight in name only. Show the reviewer what the automation saw and did, make rejecting as easy as approving, check a sample of approved items for errors, and name one accountable person. NIST asks for the last of these in plain terms. GOVERN 3.2 (opens in a new tab)^[2] reads: "Policies and procedures are in place to define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems."

The EU AI Act writes the same worry into law for one category of system. Article 14 (opens in a new tab)^[5] requires high-risk AI systems to be designed so that people can oversee them effectively. Point 4(b) says overseers should be able "to remain aware of the possible tendency of automatically relying or over-relying on the output". Article 14 applies only to high-risk AI systems, and the European Commission's AI Act page (opens in a new tab)^[6] says the rules for high-risk use cases in the Annex III areas apply from 2 December 2027, a date set by the AI Omnibus amendments that entered into force on 27 July 2026 (page last updated 19 October 2026). Whether a given automation counts as high-risk is a legal classification, so ask a qualified adviser rather than assuming either way.

What should you check before you turn an automation on?

1. Write the trigger and the single task in one sentence each. If the task needs "and", split it into two Automations.
2. Classify every step with the table. Anything that sends, spends or promises needs approval or a person.
3. Give it the fewest tools. The OWASP Excessive Agency entry (opens in a new tab)^[7] for LLM applications (2025) lists excessive functionality, excessive permissions and excessive autonomy as root causes. Set tools to Read unless a step must write, and leave Web off unless your knowledge can't answer.
4. Check the target. A schedule can cover This Mind only, every contact, a stage or a group, and contact-based schedules run once per matching contact, so a large group means many runs.
5. For messages to several contacts, begin with Send after approval. The same OWASP entry advises: "Utilise human-in-the-loop control to require a human to approve high-impact actions before they are taken."
6. Write Boundaries, and say what to do when information is missing and when to ask a person. The Specialists guide gives that advice, and Automations use the same Context fields.
7. Check consent and channel rules. WhatsApp outreach may need an approved template, and you remain responsible for notices, consent, opt-outs and lawful outreach. Ask a qualified adviser about your jurisdiction.
8. Where you can choose the target, start with a small group. Turn it On, read Runs, use Improve, then widen the target. Valid edits go live as they save, so switch it Off while you rework the steps.
9. Name an owner who reads Runs and knows where the switch is.

How often should you review an automation?

NIST's AI RMF asks that processes for human oversight be defined, assessed and documented (MAP 3.5), without prescribing an interval. Review more often where a mistake costs more, and more often early.