What stays human: the escalation boundary explained
Human in the loop AI means an agent that hands back rather than guesses. The four triggers for escalation, and what an owner actually has to do.
Every AI employee Replican builds escalates rather than guesses. It hands back to a person when a rule doesn't cover the situation, when money or a relationship is materially at stake, when the decision is a genuine judgement call, or when its own confidence is low. That boundary is designed in from the start, not added after something goes wrong, and it's the feature that makes an agent trustworthy, not a caveat that limits it.
Why is an agent that guesses worse than one that stops?
An agent that guesses when it shouldn't causes damage that a person then has to notice and clean up, often after the fact, which is worse than an agent that simply says "I don't know, here's what I found" at the moment it hits a limit.
Guessing feels efficient in the moment and is expensive afterwards. An agent that presses ahead on an invoice dispute it doesn't fully understand might send an aggressive reminder to a client who's actually paid, or go easy on one who genuinely hasn't. An agent that drafts a reply about a refund policy it's misremembered creates a promise the business now has to honour or awkwardly walk back. In both cases, the cost isn't just the immediate mistake. It's the trust cost of a client discovering that "the system" got something wrong and nobody caught it before it reached them.
Stopping costs almost nothing by comparison. A held task with a note explaining why it's held is a five-minute read for a person, resolved and moved on. A guess can cost hours of cleanup and real relationship damage; a stop costs five minutes. That asymmetry is the entire argument for building the boundary in deliberately rather than hoping the agent rarely gets it wrong.
What are the four triggers that hand a task back to a person?
An AI employee hands a task back to a person on exactly four triggers: the rule doesn't cover the situation, the money or the relationship involved is material, the decision is a judgement call rather than a procedure, or the agent's own confidence in its answer is low.
- 01The rule doesn't cover it. Every job has rules covering the normal cases. When a situation doesn't match any of them, a client asking for something nobody anticipated, an invoice with a discrepancy the rules never described, the agent stops rather than inventing a rule on the spot.
- 02The money or the relationship is material. A threshold set during mapping: a value above which a payment, a discount, or a commitment needs a person's sign-off, or a client whose history means any unusual contact should be reviewed first, not sent automatically.
- 03It's a judgement call. Some decisions are, by nature, not procedural: deciding whether to waive a fee for a difficult but valued client, or how firmly to chase a payment from someone going through a rough patch. These belong to a person even when the agent could technically execute the mechanical steps.
- 04Confidence is low. When the agent isn't sure it has understood the situation correctly (an ambiguous email, conflicting data between two systems), it says so rather than acting on a best guess. Low confidence, honestly reported, is more useful than false certainty.
Any one of these firing is enough to stop the task and hand it to a person. They aren't a checklist an agent works through in sequence; any single trigger ends the automated path immediately. The table below sets out what each trigger looks like in practice.
| Trigger | What it looks like | What the agent does |
|---|---|---|
| Rule doesn't cover it | A situation with no matching rule in what the agent was built against | Stops and flags it rather than improvising a rule on the spot |
| Money or relationship is material | A value above an agreed threshold, or a client whose history calls for care | Holds the action for sign-off instead of proceeding automatically |
| It's a judgement call | A decision that depends on discretion, not procedure, however clear the mechanics | Leaves the decision to a person, even if it could execute the steps |
| Confidence is low | The agent isn't sure it has read the situation correctly | Reports its uncertainty instead of acting on a best guess |
How is escalation designed in, rather than added afterwards?
The escalation boundary for a given job is set during the mapping stage, before the agent is built, by working out with the person who currently does the job exactly where their own judgement kicks in.
This is deliberate, not an afterthought bolted on once something has already gone wrong. Mapping the job properly means asking the person doing it now: what do you do without thinking, and what makes you pause? The answers to the second question become the rules for when the agent stops. A good bookkeeper doesn't send every overdue reminder identically; they know which clients need a gentler touch and which invoices look wrong before they even check the numbers. That instinct is exactly what gets captured as an explicit boundary, so the agent inherits the caution along with the task, not just the task.
Building the boundary in from the start also means it's testable before anything goes live. Replican tests a new agent against real past cases specifically to check that it stops on the situations it should, not just that it acts correctly on the ones it shouldn't stop on. An escalation boundary that's never been tested against real edge cases isn't a real boundary. It's a hope. See how an AI employee is built for how that testing actually works.
What does the business owner actually see, and what do they have to do?
When an agent escalates, the owner or the named person responsible sees a specific, held task with an explanation of why it stopped, not a vague alert, and not a pile-up of unresolved items with no context.
Concretely, that means a note describing what the agent was doing, what it found, and what specifically triggered the hand-back: "Invoice #4021 is 45 days overdue and the client has disputed a previous invoice this quarter, holding rather than sending the standard second reminder." The person reviewing it has enough information to decide in a minute or two, without having to reconstruct the situation from scratch. They approve it, correct it, or handle it themselves, and the agent picks the pattern up for next time where it's genuinely repeatable.
This is meant to be a small, occasional task, not a second inbox to manage. If escalations are arriving constantly, that's a sign the boundary or the rules need tuning, exactly the kind of ongoing adjustment that happens during the run phase, not a sign the whole approach has failed. See what we don't automate for the fuller list of work Replican deliberately leaves with a person from the outset.
Why is the boundary a feature, not a caveat?
The escalation boundary is what makes it possible to trust an agent with real access to real systems in the first place. Without it, giving software access to your inbox, your accounts and your CRM would be reckless regardless of how capable the underlying technology is.
Framed as a limitation, "the AI can't handle everything" sounds like a weakness to be improved away. Framed accurately, it's the mechanism that makes real access safe: an agent with an honest, tested boundary can be trusted with genuine authority over genuine systems, because everyone involved knows exactly where that authority stops. An agent sold as fully autonomous, with no defined stopping point, is either overstating what it can safely do or hasn't been tested against the cases where it shouldn't act. Either way, that's the version worth being wary of, not the one that occasionally says "I need a person to look at this."
This is why the boundary gets stated plainly on every page describing what an AI employee does, rather than mentioned once in the small print. It's the answer to the real question underneath "is an AI employee worth it": not "how much can it do" but "how do I know it will stop when it should."
Send a brief
If you want to see where the escalation boundary would sit for a specific job in your business, send a brief describing it and we'll map it honestly, including telling you if the judgement involved means it should stay with a person entirely. You can also email hello@replican.ie.
Frequently asked
questions.
Not for the great majority of the work, which is exactly the routine part the agent handles without intervention. Escalations are meant to be occasional: a small number of genuinely uncertain cases per week, not a constant stream requiring review.
That's a signal to tune the rules, not a sign the approach is broken. During the run phase, Replican adjusts the boundary as patterns emerge. If a category of "uncertain" case turns out to have a clear answer, it gets folded into the rules and stops needing a person.
Yes, and it usually should be, as the job and the business change. A boundary set during mapping is a starting point based on what was true then, not a fixed setting. This is part of what "running" an agent, rather than just deploying it, actually means.
The person who currently does the job, working with whoever maps it, sets the boundary based on where their own judgement genuinely kicks in. It isn't a generic setting applied the same way to every business.
No. A low-confidence hand-back is the agent working correctly: recognising the limit of what it can reliably judge and saying so, rather than acting on a guess. That's the behaviour the boundary is designed to produce.
Every one. It's set during mapping as a standard part of building any agent, not an optional extra added for higher-risk jobs only. This boundary is exactly what separates an AI employee from a plain automation with no judgement in it, and it's the same principle behind every entry in the full role catalogue. For the fuller test on whether a specific task belongs on the automated side of that boundary in the first place, see what jobs can you automate with AI. And when the boundary is crossed anyway, by mistake, see when an AI employee gets it wrong for how that gets caught and corrected.
Describe the job.
We’ll tell you honestly whether it fits.
No pricing games, no sales call before you’ve said what you need. Send a brief and a person reads it, not a bot.