ARTICLES · DUE DILIGENCE, ON US
Read the guides before you commit.
Decision guides and engineering notes on choosing, scoping, building and running AI systems that keep working in your daily operations, long after the launch.
- Read guides to choosing and building useful AI systems.
- Explore the engineering work behind reliable operation.
- Test a proposal before you commit budget or people to it.
Strategy
You Are Approving A Build Nobody Has Checked
The proposal has a number, a timeline and a team who believe in it. Most AI projects are lost in that room, because nobody independent has checked the decision it changes, the data it runs on, or the systems it must plug into.
The Control Your Auditor Cannot Re-Perform
A control that is a person's check may run every time and still fail an audit, because nobody can re-perform it. Design the evidence into the system from the start, captured at the moment of each action, with its reasoning attached.
It Went Live And The Savings Never Arrived
The system works and the return still has not shown up in your numbers. Every dollar is committed before launch and every dollar of return is earned after it, once the project budget has closed and nobody owns the system.
The Process That Was Too Small To Fix
For years your broken reconciliation was too small to justify a project. The cost of a small, well-specified piece of software has fallen far enough that the answer has expired, and a whole class of that work is now worth doing.
Nobody Wrote Down What Success Would Look Like
You approved a project to fix a workflow, and months later nobody in the room can say whether it worked. That is a writing failure, not a technology failure: nothing in the specification could be graded pass or fail by an outsider.
Measure AI Impact By Decisions, Not Accuracy
Model dashboards show accuracy and uptime while the business sees nothing change. This article shows where AI value actually escapes, and how to measure the 2 levels that decide it: outcomes and their business impact.
Who Signed Off, And What The Record Shows
After an incident, you need to know who approved the action and what happened. A mail thread and somebody's recollection of a call are not a control. Build those answers into the record before adding AI to a process.
The Work Does Not Carry Its Own Context
Before anyone can act on a piece of work, they go hunting for facts in the CRM, the billing system, old mail and a colleague's memory. Give people and AI agents the context with the work, so they spend less time searching for it.
Why Healthcare AI Fails At The Hospital Connection
Healthcare AI products stall at the connection to hospital systems, not at the model. FHIR fixes the shape of the data and decides almost nothing about the workflow or the auth surface. Here is what EHR and FHIR work requires.
Your Delivery Metrics Look Healthy While Risk Grows
Your DORA metrics, deployment frequency and recovery time, look healthy while teams quietly lose their hold on the systems they ship. AI makes this drift faster and harder to spot, so pair delivery metrics with comprehension.