← Research library

Escalation paths: agents should know when to stop and ask

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Marketing agents now draft, schedule, and change things across many systems, and most of the time they do it without a person watching each step. That makes one behavior more important than any single capability: knowing when to stop.

A capable junior employee does not guess when they are unsure about something important. They ask.

Agents need the same behavior, but they will not develop it on their own. By default, models tend to produce an answer rather than admit uncertainty. In our research, we treat escalation as something to design, with explicit triggers, a defined format, and a person on the other end who is expected to respond.

Why agents guess instead of asking

A language model is built to complete the task in front of it. Given a brief that asks for a product description and a source that lacks the product's price, the path of least resistance is to write something plausible. The model has no stake in the outcome and no sense that guessing is costly.

Instructions help only partly. "Escalate if unsure" asks the model to judge its own uncertainty, which is the thing it does least reliably. Our post on designing agents that handle ambiguity covers how agents can surface uncertainty. Escalation is the next step: deciding which uncertain cases go to a person, and making sure someone answers.

When agents should escalate

We group escalation triggers into five categories. Each one covers a different way an agent can reach the edge of what it should decide alone.

Missing information. The agent needs a fact that is not in its approved sources, such as a product detail, a customer permission, or a price. It should flag the gap rather than fill it.

Ambiguous instructions. The request can reasonably be read two ways, and the readings lead to different outcomes.

High consequence. The action would affect many people, cost meaningful money, or be hard to reverse, beyond the limits set for this agent.

Sensitive content. The task touches topics that need human judgment: news events, legal claims, complaints, crises, or anything involving vulnerable groups.

Conflicting rules. Two instructions or policies point in different directions, such as a brief that requests a claim the policy forbids.

Writing escalation rules

Escalation rules should be concrete enough that someone else could check whether the agent followed them. "Escalate if unsure" fails that test. Here are three rules that pass it.

  • Missing claims: "If a requested claim is not in the approved facts list, stop and list the missing claim."
  • Audience size: "If the audience exceeds 5,000 contacts, request approval before scheduling."
  • Sensitive inputs: "If the input mentions a complaint, refund, or legal matter, route it to the customer team and do not draft a reply."

Where possible, enforce the most important triggers in code, so they apply even if the agent misjudges. An audience size check is easy to run in the scheduling tool itself. A missing-claims check can compare the draft against the approved facts list before it moves on. Our post on guardrails as code covers how to build these.

The thresholds are yours to set. The 5,000 figure above is an example, and the right number depends on your list sizes and how reversible a send is.

A worked example

Consider a hypothetical lifecycle agent that drafts a reactivation email for lapsed customers. The brief asks it to mention a 20% discount. The approved promotions list shows that offer ended last week.

Without an escalation rule, the agent might write the email with the discount, since the brief asked for it. Or it might quietly drop the offer and write something vaguer. Both outcomes are guesses, and neither tells anyone that the brief and the policy disagree.

With a conflicting-rules trigger, the agent stops and sends an escalation to the campaign owner. It states that the brief requests an offer the promotions list shows as expired and offers two options: confirm a new offer or send without one. The owner answers in an hour, and the email goes out correct.

Make escalation easy to resolve

An escalation is only useful if someone responds quickly. The person receiving it should be able to decide without opening other systems. We suggest every escalation include four parts.

  • Goal: what the agent was trying to do.
  • Blocker: what specifically stopped it.
  • Options: the choices it sees, if any.
  • Request: what it needs from the person.

Route escalations to a named owner with an expected response time, and give that owner a backup. Our post on roles and permissions for a mixed human and agent team covers how to assign these roles.

If escalations pile up unanswered, agents stall and teams start overriding rules to keep work moving, which defeats the purpose. A response time target, measured and reviewed, keeps the path open.

Volume changes the design. One agent escalating a few times a week can send messages straight to its owner. Several agents escalating many times a day need a shared queue, sorted by urgency, with the same review tiers you use for approvals. Treating escalations as their own queue, with a service level, keeps them from getting lost among routine notifications.

Learn from escalations

Track how often each agent escalates and why. A rising rate may signal missing reference information, unclear instructions, or work that has drifted beyond the agent's scope. Each of those has a different fix, so the category matters as much as the count.

A rate of zero is also a warning. It may mean the agent is guessing instead of asking. If an agent handling varied, real-world inputs never escalates, test it with inputs that should trigger each rule and confirm that it does.

We also suggest tracking time to resolution and the share of escalations where the person's answer could have been found in an existing source. That last measure points to gaps in the agent's context that are cheap to close.

Escalation sits at the center of our orchestration run diagram, where the hub decides whether work proceeds, pauses, or goes to a person. The question we are studying next is how escalation rates should change as an agent moves up in autonomy. We expect them to fall at first and then settle, and we want to learn what a healthy settled rate looks like for different kinds of work.

How to apply this: write five concrete escalation rules for your most important agent, one for each category above, and enforce the most critical ones in code. Name who receives each type of escalation and how quickly they should respond.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.