What we won't automate
Replican won't automate work that changes every time, work where the judgement is the actual job, anything where a mistake is expensive or unrecoverable, jobs too small to justify building, or relationships that are the product. The mapping stage exists specifically to catch these and say no before any work starts.
Why does an AI agency have a page about what it won't build?
Because saying no to the wrong jobs is what makes the jobs Replican does build trustworthy. Most AI marketing implies that everything can and should be automated, given enough cleverness. That's not true, and pretending otherwise is how businesses end up with an agent quietly making bad calls in a part of the job nobody thought to check.
This page exists to be read before you send a brief, not after. If you recognise your own bottleneck somewhere below, that's useful information: it means the honest answer to "can Replican automate this?" might be no, and better to hear that now than after a build starts.
Work that changes every time isn't worth automating
An AI agent works by recognising patterns and following rules built around them. Work that has no stable pattern (genuinely different from one instance to the next, with no reliable structure underneath the variation) can't be encoded into rules that hold up, no matter how the agent is built.
The test Replican applies during mapping is simple: if you asked five people who've done this job to describe how they handle it, would they describe roughly the same steps? If yes, it's a candidate. If every instance is its own problem with its own logic, an agent built for it will either fail unpredictably or spend so much time escalating that it saves no time at all. That's not a reason to distrust AI in general. It's a reason to be precise about which specific job is being proposed.
Work where the judgement is the job stays with a person
Some roles exist specifically because a business needs someone to weigh things up and decide, not to follow a procedure, but to use discretion that comes from experience, context and stakes a rule can't fully capture. Hiring decisions, pricing exceptions for a difficult client, deciding whether to take on a risky piece of work: these are judgement calls, and the judgement is the entire value of the role, not a step in a larger process.
Automating the paperwork around a judgement call is fine: an agent can prepare the information a person needs to decide. Automating the decision itself, where the decision is the job, isn't something Replican will build. If the pitch for a role sounds like "let the AI decide," that's usually the signal to stop and re-scope it as information-gathering instead.
Anything where being wrong is expensive or unrecoverable needs a person in the loop
Some mistakes cost a few minutes to fix. Others cost a client relationship, a regulatory problem, or money that can't be clawed back. The cost of being wrong should shape how much autonomy an agent is given, and in some cases the honest answer is that no amount of testing makes the risk acceptable to hand to an agent unsupervised.
This is why every AI employee Replican builds has an explicit escalation boundary, set during mapping: the agent hands back to a person when a rule doesn't cover the situation, when money or a relationship is material, or when its own confidence is low. For some jobs, once that boundary is drawn honestly, almost everything ends up escalating, which is itself the answer that the job isn't a good fit for automation in the first place. Read more about how that boundary gets set on how an AI employee gets built.
Work that only looks repetitive is a trap worth naming
Some jobs look uniform from a distance and turn out to have quiet exceptions baked in that only the person doing them currently knows about. An inbox that looks like "reply to enquiries" often contains a dozen unwritten rules about which senders get special handling, which phrasing to avoid with a particular client, which requests actually mean something different from what they say.
Mapping exists partly to surface these before a build starts, by asking how the job is actually done today rather than how it's supposed to be done on paper. A job that only looks repetitive from outside can still be automated, but only once the real pattern underneath it has been found, which sometimes changes the scope of the build significantly, or rules it out.
Some jobs are too small to be worth building
Automating a job that takes ten minutes a week isn't worth the setup, testing and ongoing attention it needs to run reliably. There's a threshold below which the honest advice is to keep doing it manually, or to fold it into a person's existing role rather than give it to an agent. Replican will say so during mapping rather than build something disproportionate to the problem it solves.
This is also where bespoke automation for one stubborn workflow sometimes fits better than a full AI employee. A single small workflow doesn't need an ongoing role built around it, just a fix.
Relationships that are the product stay with a person
If what a client is actually paying for is a relationship with a specific person (an advisor, a consultant, an account manager whose judgement and manner they trust), automating the contact itself removes the thing being sold. An agent can support that relationship: preparing information, handling scheduling, drafting a first pass. It shouldn't replace the conversation. See an AI employee for client support and onboarding for where the line sits on a role that's adjacent to this without crossing it: first-line queries and setup, not the relationship itself.
Accountability has to sit with a person, always
Whatever an AI employee does, a named person at your business remains accountable for the outcome: to clients, to regulators, to your accountant. Replican builds agents to reduce the volume of work a person has to do, never to remove the person who's answerable for it. An agent that pays a supplier, sends a client communication, or files something official does so because a person set the rules it's following and remains responsible for the result. This is a design constraint, not a legal disclaimer: it shapes where the escalation boundary sits on every role, including bookkeeping and accounts admin, where money is moving and accountability matters most.
Frequently asked
questions.
During the mapping stage, before any build starts. Replican walks through the actual job, looks for a stable pattern, and checks who's accountable for getting it wrong. If the pattern isn't reliable, if the judgement is the job, or if a mistake would be too costly, mapping ends with a recommendation not to automate it.
Yes, and it happens regularly. The honest answer sometimes costs Replican the work. It's said anyway, because a business that trusts the "no" is a business that trusts the "yes" on everything else.
It can support it (gathering information, preparing options, handling the admin around a decision) without making the decision itself. The distinction is between preparing a judgement call and making one.
It gets flagged during the run stage, and the rules or escalation boundary get adjusted, or, in rare cases, the agent's scope gets pulled back to what actually works reliably. See how an AI employee gets built for how the run stage handles this.
No, but there are jobs too small within any business to be worth automating individually. A ten-minute weekly task usually isn't worth building around, regardless of the size of the company doing it.
Because a list of what won't be built is more useful to a sceptical reader than a list of capabilities, and it's the fastest way to show how the mapping stage actually works before you commit to anything. If you've got a job in mind and aren't sure which side of this line it falls on, check the AI employee FAQ for the questions that come up most, or ask us directly whether it's worth automating: describe it in an email to hello@replican.ie and we'll give you a straight answer.
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.