A year of shipping automation across different industries, different tech stacks, and different levels of technical sophistication on the client side leaves a fairly consistent set of lessons behind. Some confirmed what we already believed. A few overturned assumptions we walked in with. Here's the honest version, not the highlight reel.
What held up
- Boring beats clever. The simplest tool that solves the problem outlasts the impressive one, because the client's team can actually maintain it.
- Scope discipline saves relationships. Every project where we wrote down what was explicitly out of scope avoided the awkward conversation later.
- Monitoring is not optional. Every serious incident this year traced back to a gap in monitoring, not a gap in the core logic.
- Ownership matters more than price. Clients who cared about owning their source code and credentials had dramatically better outcomes a year in than clients who just wanted the cheapest quote.
What we got wrong at the start of the year
We underestimated how much of a project's success depends on the handoff, not the build. Early in the year, a technically excellent system sat half-used for two months because the client's team didn't feel confident touching it. We fixed that by investing more time up front in documentation and runbooks than felt necessary at the time, and it consistently paid off later. We also learned to push back harder on unscoped 'AI-powered' requests, where a client wanted intelligence applied to a process that was actually a data problem: messy inputs, inconsistent formats, no clear source of truth. No model fixes that; it has to be cleaned up first.
The pattern underneath all of it
Every project that went well this year had the same shape: a clearly scoped process, explicit handling of exceptions, monitoring that made failure visible immediately, and a handoff that left the client capable of running it without us. Every project that went sideways was missing at least one of those four things, almost never because the underlying technology couldn't do the job. Going into next year, that's the checklist we're holding every new engagement against before a single line of code gets written.
Put this to work
3 linksRelated reading
AP automation in 2026: what the research says, and what a 20-person business should do about it
New surveys put the share of finance teams still keying invoices by hand at about 2 in 3, and SMB adoption of AI at its highest yet. A read of the 2026 numbers, where they mislead, and a 3-step plan that fits a business without a finance department.
Intake → Model reads → Checks → Review tab → Your Sheet
How to choose an automation partner (without getting locked in)
The cheapest quote often becomes the most expensive system. A short checklist for hiring someone to automate your business, ownership first.
Short cycles → Full source → Your accounts → Runbooks
When automation is worth it (and when it isn't)
Not every manual task should be automated. A simple framework for deciding what to automate first, and what to leave alone.
How often? → Rules-based? → Error cost? → Automate it
Continues · Got a workflow like this?