Superintelligence Red Lines: Decisions Your Company Never Delegates
Which decisions stay with accountable people as AI agents grow more capable? A practical way to write your company's red lines and check them.

A superintelligence red line is a short, written decision that your company never delegates to an AI agent, however capable that agent becomes. For most businesses the list covers dismissals, binding legal commitments, medical or clinical advice, money movement above a set limit, changes to access, and promises made to customers outside published terms. Writing these lines before deployment gives every agent a boundary that depends on accountable people, not on how well the agent performs.
Key takeaways
- A red line is a category of decision that stays with a named person, regardless of how well an agent performs.
- Most companies need a short list, usually covering firing, binding legal commitments, health advice, payments above a limit, access changes and customer promises.
- Write each line so that a person can test a real request against it in a minute or two.
- Agents can gather facts, draft options and recommend on either side of the line, but the accountable person decides and signs.
- Review the list when tools, permissions or applicable rules change, and read the logs for attempted actions that came close.
What a red line is, and what it is not
A red line is not a judgment about how intelligent the agent is. It is a statement about who owns a consequence. Your company already works this way. A junior employee can draft a contract but cannot sign it, and a support representative can offer the standard replacement but cannot waive a policy for a large account. Agent deployments blur this distinction because the agent can often press every button the software exposes. The red line restores the difference between what a system can do and what the company has authorized it to decide.
Red lines also differ from ordinary approval steps. An approval step is a checkpoint in a workflow, often used for routine items that still need a second look. A red line is a category that the workflow never completes automatically, even when skipping the approval would be convenient. Human approval gates for superintelligent agents describes where people sign inside a workflow. Red lines are the subset of those points that never move to automatic, no matter how long the agent has performed well.
Which decisions usually belong on the list
Start from categories, then tailor each one to your business. Most lists include some version of the following:
- Dismissal, discipline or any change to someone's employment status. An agent can summarize records and flag process concerns, but it does not recommend or communicate a termination.
- Binding legal commitments, including signing contracts, accepting terms that create obligations, answering a legal notice for the company, or agreeing to a settlement.
- Medical, clinical or health guidance given to patients, clients or staff. An agent can schedule, route and share approved material, but it does not advise on treatment.
- Payments, refunds, credits and price exceptions above a limit written into your policy.
- Granting, widening or restoring access for people, tools or data.
- Commitments to customers that are not in published terms, such as promised dates, guarantees, exceptions or statements about compliance.
- Deletion or alteration of records that people may need later, including the agent's own logs and memory.
Keep the list short. A document with forty items will be ignored within a week, and people will stop reading it. A handful of well-defined lines does more work than a long policy that nobody can apply at the moment a request arrives. If a category does not match a risk you actually face, remove it and say so in the record.
How to write a red line people can apply
A line such as 'no major decisions without oversight' fails the first time someone needs it, because nobody can tell whether a specific request crosses it. Write each red line with five parts:
- The action, in plain words, such as 'sending a refund above the amount in the refund policy'.
- The accountable role, named by function, such as 'the owner' or 'the operations manager on duty', not 'the team'.
- What the agent may still do, such as retrieving the order, calculating the amount and drafting the message for the owner.
- The evidence the person needs to decide, such as the request, the policy reference, the customer history and the agent's recommendation with its uncertainty.
- Where the decision and its reason are stored, and who can read them later.
Then test each line against three real past cases. Ask whether the new rule would have stopped the agent at the right point and given a person enough to decide. If a past case would have been handled automatically under the new rule, the line is either too vague or the workflow sends work to the wrong place.
An illustrative scenario: the refund that looks routine
Illustrative scenario: A small online shop gives its support agent read access to order history and a refund tool. Its written policy allows refunds up to a set amount without review. A customer asks for a refund well above that amount and cites a late delivery that the agent can confirm from tracking data. The red line for payments above the limit means the agent cannot complete the refund. Instead, it prepares a one-page summary with the order, the tracking events, earlier refunds to the same customer and two options for the owner. It tells the customer that a person will reply within the stated business hours and does not promise the refund. The owner approves one option in a few minutes, the decision and its reason are logged, and the refund is issued by the owner's action. The agent did the preparation. The company made the decision and can explain it later.
Where agents still help near the line
Red lines are often misread as bans on agent involvement. In practice the agent does most of the preparatory work: it collects facts, flags inconsistencies, checks the policy text, drafts options and reminds the accountable person of deadlines. The risk concentrates at the last step, when a draft becomes a commitment. Design the handoff so the person sees the evidence and the agent's uncertainty, not only a recommendation. A useful handoff contains the request, the rule that applies, what the agent verified, what it could not verify, and the exact action it is asking the person to approve.
Input channels deserve particular care. An agent that reads customer email may encounter text written to manipulate it into treating a message as an instruction. Prompt injection through email covers the defenses that matter for this case. The OWASP cheat sheet on prompt injection prevention is a practical reference for the input side, and it is worth sharing with whoever maintains your agent's tools.
Enforcing the list in tools, not only in policy
A written red line that the software does not enforce is only a suggestion. Enforcement comes from permissions. The agent should hold no credential that can send a termination notice, sign a document or move money above the limit, and any action that depends on context should route to a queue that a named person must clear. The NIST AI Risk Management Framework treats governance and accountability as continuing work rather than a single sign-off, and that view fits this pattern well. The OECD AI Principles likewise place accountability with the actors who build, deploy and operate AI systems.
Be Human's managed approach builds the list into the first engagement. That engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion together, so the red lines appear in the permissions and routing rather than only in a policy document. Managed operations then include monitoring, permission review and recurring evaluations, which check the list against what the agent actually attempted. When something goes wrong, the kill switch for business agents approach lets the company halt actions without losing the logs or the queued work needed to explain what happened.
Reviewing the list as tools and rules change
Red lines are not permanent. Review them on a regular schedule and also after specific events:
- A new tool, channel or data source is connected to the agent.
- Permissions change, including temporary grants that were never removed.
- The logs show an attempted action that came close to a red line, even if it was stopped.
- A rule changes in a jurisdiction where you operate or where your customers are.
- A person in an accountable role changes, so the new owner must accept the line explicitly.
For each review, check the list against the current permissions and the recent records, then remove, tighten or add lines with a short written reason. Who can stop a superintelligent agent is a useful companion question, because a red line is only as strong as the person who can enforce it on a given day. If you are not sure which of your current agent tasks sits closest to a red line, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can a capable AI agent ever make firing decisions? For most companies the answer should be no. The agent can summarize records and flag process concerns, but the decision and the communication belong to a named person who can explain and defend them. Record the reason and the person who made the decision.
Do red lines replace an AI use policy? They do not. A use policy describes what people and agents may do in general, while red lines name the few decisions that never move to automatic handling. Most companies need both, and the red lines should be referenced from the policy so that staff know where the hard boundaries are.
Who should own the red-line list? An owner or operations leader with authority over the affected functions should own it, with legal or compliance input where obligations are involved. The owner should be named in writing, and the list should say when it was last reviewed and by whom, so that it does not drift.
How many red lines does a small company need? Start with the handful that match the risks you actually face, usually between four and eight specific lines. A long list pushes everything into a queue, slows ordinary work and trains people to approve requests without reading them, which defeats the purpose of the boundary.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


