Start from the workflow, not the model

At Axioware I have designed and deployed AI-powered automation across more than 15 business workflows, and it has cut manual operations by around 70% across client accounts. The single biggest lesson: the model is rarely the hard part. The workflow around it is.

Map where the information lives

Before choosing a model or writing a prompt, I map the workflow as it actually runs today. Where does the information start? Who touches it next? Which tools does it pass through: a CRM like GoHighLevel or Salesforce, a call platform, a shared document? Where does it get copied by hand?

The handoffs between tools are where the time goes, and they are also where an automation earns its keep.

Let the LLM do the one thing only it can do

Once the workflow is mapped, the LLM’s job is usually narrow: summarise a call, classify a lead, draft a reply, extract fields from a document. Everything else is plumbing: triggers, API calls, retries, routing and writing results back to the right place. In n8n that plumbing is most of the canvas, and that is fine.

The AI Sales Call Analysis pipeline is a good example. One discovery call goes in. LLM summarisation runs once, and two documents come out: a client-facing summary and an internal sales brief, so the rep walks into the next call already prepared. The model call is one node. The value is that nobody has to write either document by hand on a busy day.

Ship the thin slice first

I aim to get one path working end to end, in production, early. A thin slice surfaces the real problems while they are still cheap to fix: an API that rate-limits, a field that is empty half the time, a step that someone does differently on Fridays. A polished design built on assumptions breaks on contact with all three.

If it is done by hand twice a week, it is a workflow

That rule of thumb has found more value than any brainstorm. Repetitive, rule-bound tasks with clear inputs and outputs are exactly where automation pays back quickly, and that is where most of the 70% came from.

Accuracy is a feature

When an automation writes into a client’s systems, mistakes compound. On the enterprise data pipelines I work on, which connect GoHighLevel, Salesforce and the Meta Graph API with OpenAI and Claude, we hold 99.5% processing accuracy. A number like that does not happen by accident: inputs need validating, outputs need checking before they are written, and failures need to be visible instead of silent.

In short

1. Map the workflow before you pick the model.
2. Give the LLM one clear job and build solid plumbing around it.
3. Ship one working path early, then widen it.
4. Measure accuracy, and treat silent failures as bugs.