Every automation project I've walked into that failed had the same fingerprint: someone automated the loudest process, not the right one. The invoice team complained the most, so the invoices got a bot. Six months later the bot was quietly switched off, and nobody could quite explain why the promised savings never showed up.
Automation doesn't fail because the technology is weak. It fails because the process underneath it was a bad candidate — and no amount of RPA, LLMs, or clever orchestration fixes a process you picked for the wrong reasons.
The trap: automating the pain instead of the payoff
Loud pain and high value are not the same thing. The task your team grumbles about might be annoying and infrequent — automate it and you've spent three months of engineering to save four hours a quarter. Meanwhile the boring, invisible work that runs 4,000 times a month sits untouched because nobody complains about it.
Before you score a single candidate, separate two questions that get tangled constantly: Is this process painful? and Is this process valuable to automate? The second one is the only one that pays your invoice.
A scoring lens that actually predicts ROI
In my engagements, I've found four dimensions do most of the sorting. Rate each candidate honestly — a gut "high/medium/low" is enough to start.
- Volume × frequency. A task done 20 times a day beats one done twice a month, even if the monthly one feels heavier. Automation returns compound on repetition.
- Rule stability. If the logic changes every quarter, you'll spend more maintaining the automation than you saved. Stable, boring rules are gold.
- Input consistency. Structured, predictable inputs (a form, an API, a fixed template) automate cleanly. A process fed by free-text emails and PDF scans from 40 different vendors will punish you.
- Cost of error. High-stakes, low-tolerance steps — anything touching payroll, compliance, or a customer's money — need a human in the loop, which changes the economics entirely.
The best first automation is high-volume, rule-stable, structured-input, and low-blast-radius if it's wrong. Find that corner and start there.
Where AI changes the old math
Classic automation demanded clean, structured inputs. That single constraint disqualified most real-world processes, where the input is a messy email or a scanned document. Large language models loosen that constraint — they tolerate ambiguity, read unstructured text, and handle the "it depends" cases that used to force a human handoff.
But loosening isn't removing. An LLM that's right 95% of the time is a gift in a support-triage flow and a liability in a wire-transfer approval. The discipline shifts from "can we structure the input?" to "what's the cost of the 5% it gets wrong, and can we catch it?" Match the tool to the tolerance:
- Deterministic rules for anything where being wrong is expensive and the logic is clear.
- LLM-assisted, human-approved for judgment calls with real downside — a person reviews before it commits.
- Fully autonomous LLM only where errors are cheap, visible, and easily reversed.
Start small, instrument everything
Pick one process. Not a platform, not a "transformation" — one process, ideally one where you already know the baseline: how long it takes, how often it runs, what it costs today. If you can't measure the before, you can't prove the after, and unmeasured automation quietly becomes shelfware.
Ship a thin version, keep the human in the loop for the first few weeks, and watch three numbers: time saved per run, error rate versus the manual baseline, and how often a human still has to intervene. That last one is the honest signal. If people are constantly fixing the bot's output, you didn't automate the process — you added a step.
Where to start Monday
List the ten processes your team touches most this week. For each, mark volume, rule stability, input structure, and cost of error. The winner won't be the one that annoys everyone — it'll be the quiet, high-frequency, structured one nobody thinks about. Automate that first, prove the number, then earn the right to the harder ones.
The best automation candidate is rarely the one shouting for attention. Which of your quiet, repetitive processes is costing you the most while nobody's watching?