An AI employee is built in four stages: map, build, deploy, run. Replican maps the repetitive work and where a human must stay in control, builds the agent against your real tools and rules, deploys it onto a private managed VPS you own, and then keeps running it: monitoring, tuning and extending the agent as your business changes. No stage is skipped, and no stage happens without the one before it.

This page describes each stage in the depth an owner deciding whether to commit actually needs, without promising how long any of it takes. The shape of the process is consistent. The duration depends on the job, the tools involved and how much access takes to arrange, none of which Replican will guess at in advance. For a closer look at the mapping stage specifically, including how the rules actually get captured, see how an AI employee is built, or browse the full guides index for the rest of the detail behind this process.

What happens during the map stage?

Mapping means Replican sits down with the actual repetitive work in your business and works out which parts are worth automating, which parts need a person to stay involved, and which parts aren't worth automating at all. This is the stage where most of the honest work happens, and it's the one competitors skip because it doesn't look like progress.

Concretely, mapping involves walking through a specific job (invoice chasing, inbox triage, first-line lead outreach, whatever's currently eating time) and breaking it into steps. Some steps follow a pattern that repeats reliably: the same categories of email, the same structure of invoice, the same shape of enquiry. Those are candidates for automation. Other steps require a judgement call that depends on context an agent can't reliably read: whether to chase a client who's normally reliable but three weeks late, whether a complaint needs a personal call rather than a templated reply. Those stay with a person, and Replican says so during mapping rather than building around them and hoping.

Mapping also produces something concrete: a description of where the escalation boundary sits for this specific job. What does the agent do when it hits a case the rules don't cover? Who does it hand back to, and how? That boundary is agreed before a line of the build starts, not bolted on afterwards. If, at the end of mapping, the honest answer is that a job isn't worth automating, because it's too small, too variable, or the risk of getting it wrong outweighs the time saved, Replican says that plainly. See what we won't automate for the full reasoning behind that boundary.

What's asked of the client at this stage is mostly access and honesty: a walk-through of how the job actually gets done today (not how it's supposed to get done), and visibility into the tools involved.

What happens during the build stage?

Building means the agent is built against your actual tools and your actual rules, not a generic template adapted after the fact. If your books run in Xero, the agent works inside Xero. If your inbox is Outlook and your CRM is HubSpot, the agent is built against those specific systems and however you've configured them, not a hypothetical "typical" setup.

This matters because most repetitive work isn't actually simple once you look closely. An inbox triage job sounds uniform until you notice that supplier emails need one kind of handling, client emails need another, and anything from a specific handful of people needs to go straight to a named person regardless of content. Those rules come from the mapping stage, and the build stage is where they get encoded, along with the escalation points already agreed, so the agent hands back rather than guesses whenever a case falls outside what was mapped.

Concretely, building involves connecting the agent to the relevant tools with the access it needs (and no more than it needs), encoding the rules that came out of mapping, and testing the agent against real examples from your business rather than synthetic ones. A bookkeeping agent gets tested against your actual bank feed and your actual invoice formats. An inbox agent gets tested against a real slice of your real inbox, with your permission.

What's asked of the client here is access to the relevant tools, a point of contact who can answer questions about edge cases as they come up during testing, and patience with the fact that real businesses have messier data than demos ever show.

What happens during the deploy stage?

Deploying means the agent goes live on a private managed VPS that belongs to the client, not to Replican. This is a deliberate architectural choice, not a technical detail. A private managed VPS is, in plain terms, a server set up specifically for your business, running your agent and nothing else, that you own, rather than a shared subscription platform where a vendor holds your account, your data and your access.

The practical difference shows up if you and Replican ever part ways: you keep the server and everything on it. Nothing disappears because a subscription lapsed, because that isn't how this is structured. The full explanation of what this means for data residency, GDPR and what "managed" involves in day-to-day terms is on infrastructure you own. It's the single biggest structural difference between how Replican deploys agents and how most AI vendors do.

What's asked of the client at deployment is minimal: sign-off that the agent behaves as expected against the real cases it was tested on, and agreement on how the server itself gets provisioned and paid for. Replican handles the technical setup.

What happens during the run stage, and why does the agency stay involved?

Running means Replican keeps monitoring, tuning and extending the agent after it goes live, rather than handing over a finished build and moving on. This is the stage most agencies don't offer, because it's ongoing work rather than a one-off project, and it's the reason Replican describes itself as running AI employees rather than delivering AI projects.

Running covers three things: monitoring, tuning and extending. Monitoring means checking that the agent is behaving as expected and catching drift before it becomes a problem: a supplier changes their invoice format, a client's enquiry patterns shift, a tool gets upgraded and something in the integration needs adjusting. Tuning means refining the rules as real cases surface ones the original mapping didn't anticipate, and narrowing or widening the escalation boundary as trust in the agent's judgement is earned or found wanting. Extending means adding scope to the same agent, or building a second one, as the business identifies more work worth automating once the first agent has proven itself.

What's asked of the client during the run stage is feedback: flagging when something looks wrong, and being available for the occasional conversation about whether a rule needs adjusting. The agency stays involved so that responsibility for the agent's ongoing behaviour doesn't quietly become the client's problem to solve alone.

Where does this process stop, and what happens if a job isn't a fit?

It stops at the mapping stage, honestly, whenever a job turns out not to be worth automating, and Replican will say so rather than build it anyway. Some work changes too often to be worth encoding into rules. Some work is genuinely a judgement call that belongs to a person. The full reasoning behind where that line sits is on what we won't automate, and it's worth reading before a first conversation, because it explains what Replican will turn down as much as what it will build.

For work that doesn't fit a named role at all (a single workflow rather than an ongoing job), see bespoke automation for one stubborn workflow. For owners who'd rather learn to build this kind of thing themselves, there's one-to-one AI training instead. If you're still weighing an AI employee against hiring, a VA or a DIY tool, the comparison hub works through each of those pairings honestly.

If you have a specific job in mind, send a brief describing it and we'll tell you honestly which stage it's likely to stop at. If you've got questions before that, the AI employee FAQ answers the ones that come up most, or email hello@replican.ie directly.

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.

Send a brief Email hello@replican.ie