For most of the 2010s, “process automation” had a credibility problem.
Companies bought RPA licenses, hired consultants, drew swim lanes, and ended up with brittle bots that broke every time someone changed a column header in Salesforce. Half the projects were quietly switched off. The other half just shifted the bottleneck somewhere else.
Then something happened: the underlying tools got dramatically better. Modern API ecosystems, observable infrastructure, cheap LLMs, and orchestration platforms like n8n and Temporal have collapsed the cost of building automations that don’t suck.
The leverage is real now. The question is no longer “should we automate?” — it’s “which workflows actually deserve to be automated, and which ones do we leave alone?”
Here’s the lens we use.
The four-question filter
Before we touch a single integration, we run every candidate workflow through four questions. If a workflow doesn’t score well on at least three, we leave it alone.
1. Is it boring?
Boring is a feature, not a bug. The best automation candidates are workflows where the human cost is high but the cognitive value-add is near zero: copying the same data between two systems, formatting the same report, sending the same nudge.
If a step requires real judgement, that’s where you keep the human in the loop — possibly with AI assistance, but not on autopilot. The biggest mistake we see is teams trying to automate the parts that are interesting and leaving the boring drudge work to people. Backwards.
2. Does it run on a predictable trigger?
“Every time an order comes in.” “Every Monday at 9am.” “When a ticket sits unanswered for 4 hours.” Predictable triggers are easy to wire up reliably.
What’s hard? Workflows that run on fuzzy human judgement: “when it feels like the right moment to follow up.” Those are the ones that look automatable on paper and become a nightmare in production.
3. Are the inputs and outputs well-shaped?
Structured data in, structured data out, with a clear definition of “done.” This is where most of the implementation cost lives. If the inputs are 12 different email formats from 12 different vendors, you can absolutely automate it — but you’re also signing up to maintain a parser.
This is where modern LLMs change the math. Five years ago, parsing semi-structured input (invoices, support tickets, contracts) was a pain so expensive most teams wouldn’t bother. Today, a small Claude or GPT-4o-mini call handles it in seconds for a few cents — and that unlocks a class of automations that simply weren’t economical before.
4. Is the cost of getting it wrong recoverable?
Some workflows are forgiving: a misrouted internal Slack message is a five-second fix. Some are not: a misfired refund or a wrongly-sent customer email is genuinely costly to undo.
The fix isn’t to avoid automating the high-stakes ones. It’s to design them with confidence thresholds, dry-run modes, and human escalation built in. Anything below a defined confidence score gets queued for human review. This is the same principle behind autonomous driving — and it works.
The shortlist nobody talks about
Once you’ve run the filter, what tends to make the cut isn’t the glamorous stuff. It’s the unsexy middle of the org chart:
- Lead routing and enrichment. New form submission → enrich → score → route to the right rep with context. Nobody’s job, everybody’s headache.
- Returns and refunds processing. The 80% of cases that follow your standard policy can run themselves with full audit logs.
- Inventory reconciliation. Sync between ERP, Shopify, and 3PL — with discrepancy alerts when the numbers diverge.
- Onboarding handoffs. Sales-to-success, account creation, welcome sequences, doc provisioning.
- Internal reporting. Daily/weekly digests pulled from N tools, summarized, posted in the right channel.
- Vendor and bill reconciliation. Statement matches, exception flagging.
None of these will impress your board. All of them will return your team 10–25 hours per week, every week, forever.
The real ROI is compounding, not single-shot
The big mental shift: don’t evaluate an automation by the time it saves on day one. Evaluate it by the time it saves every week, forever, while requiring near-zero maintenance.
A workflow that takes 20 hours to build and saves 4 hours a week pays for itself in five weeks. After that, it’s pure compound interest — and your team gets to redirect those 200+ hours a year into work that actually moves your numbers.
That’s the bet. After a decade of false starts, the technology is finally good enough to make it consistently pay off.
The only thing left is picking the right workflows. That’s where most of the value still gets lost — and where most of our work begins.