Most RPA failures in small and mid-market teams share a root cause: the project was scoped like an enterprise deployment — six months of discovery, a Center of Excellence, a half-million-dollar license — before anything got shipped. Below is what we've learned from a dozen engagements where the brief was smaller and the answers had to arrive faster.
Start with one process. One.
The first automation should pay for the rest. Pick the process with the highest volume × time-per-run × error rate × executive attention. It's usually invoice processing, employee onboarding, or a finance reconciliation. Ship that in six to eight weeks, end to end, so the organisation can see what working looks like.
Pick the platform the team will own
- Microsoft Power Platform — if the team already lives in Microsoft 365 and needs Power Automate plus Power Apps plus Power BI to talk to each other.
- UiPath — when you need attended bots, heavy document processing, and an orchestrator you can grow into.
- Make or Zapier — for SaaS glue, marketing ops, and anything that's mostly moving structured data between cloud apps.
- Custom code — when the above bend too far. A small Python worker with a queue and monitoring costs less than a bent license.
Pick whichever platform the team can operate after you leave. The best automation is one your people can debug on a Tuesday afternoon without paging a consultant.
Instrument from day one
Every bot gets three things from the start:
- A structured log per run (timestamp, inputs, outputs, duration, outcome).
- A weekly digest emailed to the business owner.
- A human escalation path when the bot hits an unknown case.
You need this to prove the ROI, tune the bot, and catch drift. Teams that skip this ship a bot, celebrate, and discover six months later that it's been failing silently for weeks.
Don't automate a bad process
If the underlying process is broken, RPA just makes it break faster. Before scripting anything, spend a week watching the manual workflow. Nine times out of ten you'll find shortcuts, duplication or rework you can fix before the bot writes a single line of code. That redesign is often worth more than the automation itself.
The 80/20 rule on exceptions
A bot that handles 80% of cases cleanly and escalates the rest to a human is usually the right target. The last 20% of edge cases typically cost as much to automate as the first 80% combined — and they change. Route them to a person, log what happens, feed the learning back.
Measure what the executive cares about
Time saved is the obvious metric and the wrong one to lead with. Lead with:
- Cycle time — how long the process takes from start to done.
- Error rate — how often a human has to fix a mistake.
- Cost per transaction — fully-loaded labour + licensing.
- Coverage — what fraction of the eligible volume the bot actually handles.
Report these monthly. Your executive sponsor will stay engaged because they can see the business change.
What we've shipped
In the last twelve months we've automated processes in accounts payable, customer onboarding, claims triage, inventory reconciliation, HR offboarding and sales-handoff routing. Most came in between 8 and 14 weeks. All are still running. None are interesting — which is exactly what you want from infrastructure.
If you want to talk through a first candidate for automation in your own team, drop us a line. First call is free and we're direct about whether RPA is actually the right tool for the job.