Section 1
The problem was never defined
The most common failure looks like success for the first month. A team adopts a tool, uses it enthusiastically, and cannot say what business number it was meant to move. When the enthusiasm fades, nothing remains. The fix is a sentence written before any build: we lose time or money because this specific work waits for a person to collect, read, format, or route information. If the sentence cannot be written, there is no project. Then attach a baseline. How long the work takes now, how often it breaks, what the delay costs. Teams resist this because measuring the baseline is tedious and delays the fun part by a week. That week is what separates a capability from an experiment. The structural version of this discipline is in [Building an AI Automation Roadmap for Your Startup](/blog/building-an-ai-automation-roadmap-for-your-startup).
Section 2
The data is not in a usable state
The second failure arrives during the build. The information the automation needs is spread across a CRM with inconsistent fields, a shared drive of PDFs, three spreadsheets, and one person's inbox. The pilot works on the clean sample and falls apart on the real thing. The fix is to scope the first automation around the data you already have in a usable state, rather than around the workflow you most want to fix. This feels like a compromise and it is the correct one. Cleaning a data source to enable an automation is a legitimate project, but it is a separate project with its own timeline, and pretending otherwise is how a six-week build becomes a six-month one.
Section 3
The people who use it were told, not asked
The third failure is quieter. The automation works and the team routes around it, because it changed how their work is judged and nobody discussed that. Ask the people doing the work what wastes their time before deciding what to automate. They will usually name something different, smaller, and more valuable than what leadership had in mind.
Section 4
It shipped and nobody owned it
The fourth failure happens after launch. An automation runs for months producing slightly wrong output that people manually correct without reporting, because reporting it is more effort than fixing it. The savings evaporate and nobody notices. The fix is an owner per automation and a correction rate that is actually tracked. Give reviewers a one-click way to mark an output as wrong, and look at that number monthly. A rising correction rate almost always means the underlying workflow changed: a supplier altered a form, a regulation moved, a system was updated. The automation did not break. Its world did, and nothing was watching for it.
Section 5
The risk was discovered late
The recurring failure here is running the risk assessment after deployment rather than before, which inverts the sequence NIST recommends. By then the workflow has dependencies and switching it off has a cost, so the risk gets accepted rather than addressed. Do it in an hour at the start. What can the system access, what can it change, what must it never decide alone, who is accountable when an error reaches a customer or employee. Keep sensitive data out of prompts that do not need it, log decisions and escalations, and disclose when AI is involved. Where the rules are sector-specific, [Regulatory Challenges in Implementing AI Automation](/blog/regulatory-challenges-in-implementing-ai-automation) covers the ground in more detail.
Section 6
The wrong things were counted
The final failure is measurement. A team reports automations launched, hours saved, and messages processed, none of which tell you whether the business improved. Hours saved in particular is only real if the hours went somewhere, either into more output or out of the cost base. Track cycle time, rework, conversion, cost per transaction, adoption, exception volume, and correction rate instead. Then hold a monthly review with the same five questions: what moved faster, what improved in quality, what risk appeared, what humans still had to fix, and what should be retired. Retiring things is a sign the process works. On reporting results honestly, [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation) is a useful companion.