Most workflows don't need AI. They need clean inputs, deterministic logic, and an integration that fires on a trigger. For those, AI is overkill at best and unstable at worst.
But there's a class of workflows where rule-based automation has been quietly losing for the last two years. We've rebuilt enough of these on Claude (and occasionally GPT-4) to see the pattern clearly: anywhere the work depends on language, judgment, or unstructured context, an LLM-driven approach now outperforms the rules-based version we used to build.
Five workflows where this is true, plus the cost ceiling to watch for.
1. Unstructured intake
If your team handles email, support tickets, document submissions, or freeform forms, the first 30% of the workflow is usually a human reading and routing.
The traditional approach: write extraction rules. Match keywords, parse patterns, fall back to manual handling for anything weird.
The Claude approach: feed it the document and ask, "what does this person want, what's their priority, and where should this go?"
In our builds, the LLM approach gets to 90%+ accuracy on the first pass and handles edge cases better than the rules-based version ever did. It also doesn't break when someone phrases things differently — which is the failure mode rules-based extraction lives with permanently.
2. Classification with subtle context
"This support ticket is from a high-value customer with three tickets this month and is using language that suggests churn risk."
That's the kind of classification that takes humans seconds and rules-based systems forever, because the signal is in tone, history, and pattern, not keywords. Sentiment analysis APIs were the old answer; they were never very good.
Claude reads the ticket plus the customer's history and returns a structured JSON: priority, risk score, recommended owner, suggested first response. We've built this for a SaaS support team where the impact was reducing average first-response time on at-risk tickets by about 70%.
3. Drafting and tone-matching
Drafting first-pass replies — to leads, to customers, to internal stakeholders — used to be the part of automation we couldn't touch. Generic templates are obviously generic. Customers can tell.
Claude with a few examples of your team's voice now produces drafts that need editing, not rewriting. The compounding win is that the draft is in front of the operator instead of a blank email, which collapses response time for the team that does the editing.
Human-in-the-loop: AI drafts, human reviews, AI doesn't send. We've seen too many teams try to take the human out and end up with something off-brand on the way to a customer.
4. Multi-step research and synthesis
Sales teams spend hours pre-call: pulling LinkedIn, the company website, their last interaction in HubSpot, the product they expressed interest in, recent news mentions.
Traditional automation can pull all of that and put it in a doc. The doc still requires a human to read and synthesize.
Claude can pull, read, and synthesize, returning a one-page brief with the things that actually matter for the call. We've built this for a sales team where the prep time per call dropped from 45 minutes to 6, and the close rate went up because reps showed up better-prepared, not less.
5. Customer-facing language at scale
Localizing transactional emails. Rephrasing notifications for different audiences. Generating personalized subject lines. Anything where the volume is high, the variations are subtle, and the cost of being wrong is low.
This is where AI wins on cost too: hiring writers to handle the long tail of customer communication is expensive and slow. Claude does the long tail at near-zero marginal cost, with a human reviewing the first version of each new template.
The cost ceiling
LLMs are not free. At our typical usage (Claude Sonnet, structured outputs, mid-context), the per-call cost is somewhere between half a cent and a few cents. That's nothing for low-volume workflows. It's something for high-volume ones.
A workflow that fires a million times a month at 2 cents per call is $20K/month in API costs. That's a real number, and it changes the math on whether AI is the right tool versus a clever rules-based system.
We use a simple test: estimate the per-call cost, multiply by expected volume, and compare against the human cost it's replacing. AI wins easily when you're replacing 10 hours of senior time per week. AI loses when you're replacing 1 minute of junior time per day. The middle is where it gets interesting.
The pattern
The clean rule we use internally:
If the work involves reading or writing language and the rules can't be cleanly enumerated, reach for an LLM. If the rules are clean, reach for n8n.
Anywhere those overlap — and they overlap a lot now — the right build is usually a hybrid: n8n handles the triggers, integrations, and flow control; Claude handles the steps where language or judgment lives.
The best automation systems we've built in the last year aren't AI systems or rules systems. They're operating systems where the two are wired together, each doing the part it's actually good at.