AI belongs inside the workflow, not instead of it
The version that fails hands a whole process to a model and hopes. Survival looks different: a deterministic pipeline, with a model call only where a human would have had to think.
The short version
5 things that decide this
- 01A process rebuilt entirely as an AI agent loses the parts that worked: reliable routing, fixed steps, an audit trail.
- 02The pattern that holds up in production keeps a deterministic pipeline and inserts a model call only at the step that needs judgment.
- 03Classification, drafting, and extraction are model work. Routing, sequencing, and delivery are not.
- 04Little Tree Confections' meeting-to-task automation runs this way: n8n handles the pipeline, AI refines the transcript, ClickUp receives a finished task.
- 05The question worth asking before any build is not whether to use AI, but which single step actually needs it.
Replacing the process is the mistake
A common first move is to hand an entire process to an AI agent. Read the inbox, decide what matters, take the action, tell someone. It demos well because a demo has no edge cases.
In production, the process had parts that were never the hard part. Routing a task to the right department. Checking a due date. Writing a record to the right system. A deterministic pipeline already did those reliably, with a clear log of what happened and why. Handing all of it to a model trades that for a system that has to re-decide the easy parts every run. Sometimes it gets them wrong.
The tell is a workflow with no fixed steps left. If every part of the process depends on what the model decides this time, nobody can say what the automation actually does.
- 01Routing, sequencing, and writing to a system of record do not need judgment. A model deciding them adds a failure mode with no upside.
- 02A process with no deterministic steps left has no audit trail, because every run can behave differently.
- 03The parts that already worked reliably are usually the parts an AI-first rebuild throws away first.
Keep the pipeline. Insert the model where judgment lives
The version that holds up keeps the pipeline in place. It adds a model call only at the step a human used to read, judge, or interpret. Everything before and after that step stays exactly as reliable as it was.
Three kinds of steps are usually the right place for a model. Classification: deciding which category, department, or priority something belongs to, where the input is messy but the output is a fixed set of options. Drafting: turning a rough input into a first-pass written artifact a human still reviews. Extraction: pulling structured fields out of unstructured text, like a name, a date, or an amount buried in a paragraph.
Everywhere else in the pipeline stays rule-based. A model call has a cost, a latency, and a chance of being wrong that a fixed step does not. Spending that cost only where a rule genuinely cannot do the job is the whole discipline.
- 01Classification: sorting messy input into a known set of categories.
- 02Drafting: producing a first-pass version of something a human still checks.
- 03Extraction: pulling structured fields out of free text.
- 04Everything else: routing, delivery, logging, retries. Rules, not model calls.
Little Tree: n8n routes, AI refines, ClickUp receives
This is the pattern behind Little Tree Confections' meeting-to-task automation. Fireflies records every meeting, and n8n owns the pipeline. It pulls the transcript and sends it out for refinement. The result becomes a ClickUp task, with an assignee, a due date, and a link back to a Notion summary.
One model call does the whole job inside that pipeline: turn a rough transcript into a clean action item. That is drafting and extraction, the work a person used to do by re-reading the recording. It does not decide which department a task belongs to. n8n reads the live ClickUp structure and routes on that instead. Routing keeps working when a department changes, and it never depends on a model guessing right.
A second, separate workflow runs on a daily schedule to build the CEO brief and send deadline reminders. It is deterministic on purpose. A daily summary does not need a model deciding whether to run.
Questions this raises
01How do I know which step in my process actually needs AI?
Ask which step currently requires a person to read something messy and make a judgment call, rather than follow a rule. That is a classification, drafting, or extraction step, and it is the only kind of step a model earns its place at. If a rule can decide it today, a model does not improve it.
02Isn't a full AI agent simpler to build than a mixed pipeline?
It looks simpler at the design stage, because one component appears to replace many. In production it is harder to debug. A failure could be anywhere inside one opaque decision, instead of at one traceable step in a pipeline with fixed stages.
03Does this mean n8n workflows only need one AI step?
Most need one or two, not more. Little Tree's real-time pipeline calls a model once, to refine the transcript. Everything before it and after it, ingestion, routing, task creation, logging, is deterministic n8n logic with no model in the loop.
Related
- Business process automation →The service behind builds like this one.
- n8n in production →Our review of the tool that runs the deterministic half.
- AI agents vs RPA →Where a model belongs in an automation, and where a fixed rule still wins.
- Little Tree Confections: meeting-to-action automation →The production workflow this piece draws from.
