What deserves automation, and what does not
Three tests, all required: the task is repetitive enough that rules or examples can describe it, it runs often enough for the hours to matter, and a mistake is cheap or catchable. Sorting incoming email passes all three. Writing a sensitive client proposal fails the third and should stay human, with AI drafting at most. And a broken process fails an implicit fourth test: automating it makes it fail faster at scale. Fix the process first, then automate the fixed version. This ordering is why our levers put diagnosis before automation.
Finding candidates: follow the copy-paste
The best automation candidates hide in plain sight because everyone has stopped seeing them. Look for re-keying, where the same information is typed into two systems; recurring documents assembled by hand every week or month from the same sources; routing, where someone reads incoming requests only to decide who handles them; and monitoring, where someone checks a source on a schedule. One question surfaces most of them: “what do you do every week that a well-trained new hire could do with a checklist?” The answers, with hours attached, are the shortlist a diagnostic quantifies.
The orchestration layer: n8n, Make, and the no-silo rule
Workflow tools like n8n and Make are the plumbing: they watch for a trigger, then execute steps across your existing tools, with AI called where judgement-like work is needed. Two principles matter more than the brand. First, no new silo: the workflow should connect the CRM, mailbox and drive you already run, not introduce another place where data lives. Second, ownership: self-hosted n8n keeps data and logic fully under your control, which is why it is our default for clients with confidentiality stakes, while Make trades some of that control for a gentler learning curve. The right choice depends on your volume, your data and who maintains it.
Where AI fits inside a workflow
Classic automation handles the deterministic steps: move the file, fill the field, send the message. AI slots in exactly where the old blocker was, the step that needed reading or judgement: classify this request, extract the amounts from this invoice, summarise this thread, draft this reply. A useful mental model is a conveyor belt with a reader stationed at one point. Most workflows need zero or one AI step; if a design needs five, the process was probably not understood first. The deterministic parts stay deterministic, which keeps the whole thing debuggable.
Human in the loop: guardrails where mistakes are expensive
The design question is not whether the AI is reliable enough to act alone, it is what a mistake costs at each step. Cheap, reversible, internal: let it run, log everything. Expensive or client-facing: the workflow prepares, a human approves, and approval is one click in the tools your team already uses. Well placed, the human step costs seconds and keeps trust; and as confidence builds in the logs, approval can be relaxed step by step, in that direction only. Loosening a guardrail is easy; rebuilding trust after an automated blunder reached a client is not.
Every automation ships with its own measurement
Baseline first: hours per month and error rate of the manual process, written down before anything is built. Then the workflow counts its own work as it runs: volume processed, exceptions raised, failures. Hours saved times loaded hourly cost gives a payback your accountant can verify, and quick wins measured this way typically pay back in two to six months. Our reporting case study shows the format on a real deployment. One simple dashboard, not a reporting machine; the measurement exists to steer, not to impress.
Automations are living software
Every workflow depends on tools that change without asking: an API updates, a mailbox format shifts, a source moves. Unmaintained, an automation does not announce its death; it fails silently while everyone assumes it is running, which is worse than the manual process it replaced. The minimum is error alerting that reaches a named human, plus a monthly glance at the counters. Beyond a handful of workflows, dedicated monitoring becomes cheaper than the risk, which is what our copiloting format exists for. Budget maintenance from day one; the alternative is rediscovering it during an incident.
Your first automation, concretely
Pick from your shortlist the candidate that combines real hours, cheap mistakes and an enthusiastic owner, and resist the temptation of the impressive one. Write on one page what triggers the workflow, what it does, where the human approves, and how you will know it worked, with the baseline measured. Build small, run it in parallel with the manual process for a couple of weeks, compare, then cut over and widen. The pattern is the same at every scale, and it is how a full integration strategy starts: one verified win, then the next.