MAD 90 Editorial • October 4, 2026 • AI Enablement
A useful AI pilot begins with a specific task. “Use more AI” is too broad to measure. “Draft a first response to a recurring customer question, with a person checking it before sending” is a clearer starting point.
Write down who performs the task, what information they use, how often it happens, and what makes a good result. Include exceptions and the cost of errors.
Use approved information and tools. Keep credentials, customer records, and confidential material out of an unapproved pilot. Confirm how the tool handles inputs and who is allowed to use it.
Try representative examples and compare the pilot with the current method. Check time spent, accuracy, effort to correct the output, and usefulness to the person doing the work.
Assign a reviewer and make the approval step explicit. AI-generated text can sound confident while being wrong. A faster draft only helps when the final result is fit for purpose.
After a small pilot, choose whether to improve it, expand it, or stop. Record what you learned. Avoid scaling a workflow simply because the demonstration looked impressive.
Request a consultation at matthews@mad90.com.
This article offers MAD 90’s practical perspective. For guidance on AI risks and evaluation, see NIST’s AI Risk Management Framework and its Generative AI Profile.