I didn't set out to build a framework.
I was trying to understand why every improvement I made seemed to create another bottleneck. Somewhere between building dashboards, automating workflows, reading The Goal, and asking thousands of questions, I stopped asking "What can AI automate?" and started asking "What prevents value from flowing?" This is where that question lives.
The AI Throughput Lens
A five-part lens for deciding where AI belongs, where it doesn't, and what should improve next. The center never changes: today's constraint.
Today's
Constraint
1. Systems
See the whole
2. Constraint
Find the limit
3. Socratic AI
Ask better questions
4. Build
Fit the work
5. Evolve
Measure and repeat
AI creates capability. Throughput creates value.
Automating a non-constraint may save time locally while producing little measurable improvement to the organization.
Research Lab
The open questions I haven't resolved yet.
Theory of Constraints applies to knowledge work with almost no translation required.
The vocabulary changes — constraint becomes bottleneck becomes "that one approval," inventory becomes queued decisions — but the underlying mechanics look the same on a factory floor and in a Slack thread.
A real before/after measurement of decision latency, not just a description of it.
I have plenty of anecdotes about work sitting in a queue. I don't yet have a clean measurement of how long value actually waits at a decision point, and I'm suspicious of my own hypothesis until I do.
Does operational accounting for AI need a new statement, or just a new line on an existing one?
Traditional cost accounting wasn't built to notice digital inventory. Still working out whether that means a new operational report, or whether existing throughput accounting already has room for it.
Socratic-style AI prompting produces better decisions than direct-answer prompting, for the same problem.
This is close to the center of the whole site, and I still don't have a fair side-by-side test of it — only the felt experience of it working, which is not evidence.
How does this framework interact with Lean and DevOps, which already claim similar territory?
Theory of Constraints, Lean, and DevOps all converge on 'find where flow breaks,' from different origins. I haven't mapped where they actually disagree versus where they're using different words for the same idea.
What are the consequences of knowledge production outrunning an organization’s ability to evaluate it?
The biggest open question on the site and the one with the least developed thinking. It’s next.
Research Log
Where the Research Lab holds long-running open questions, this is the dated log underneath it — one entry per experiment, added as they happen. Hypothesis, method, result, lesson, next step.
Investigative AI vs. Summarization
Hypothesis
Determine whether AI can reconcile complex business records instead of simply summarizing data.
Method
Provided multiple reports and iteratively refined prompts, asking AI to identify discrepancies and match transactions.
Result
AI performed well while working with the supplied data, but reached a hard limit when additional evidence was locked inside systems it could not search.
Lesson Learned
The next generation of business AI depends on secure access to operational systems. Better prompts cannot replace missing data.
Next: Explore AI agents or workflows capable of securely searching business systems and tracing records end-to-end.
Rapid Dashboard Response
Hypothesis
A targeted dashboard can resolve reporting questions faster than lengthy explanations.
Method
Used AI to rapidly isolate a subset of reporting data into a focused HTML dashboard. Iterated on formatting and layout until the output was presentation-ready.
Result
The dashboard redirected the discussion toward evidence and highlighted that the underlying data required attention rather than the report itself.
Lesson Learned
Interactive visual evidence reduces debate and accelerates problem solving.
Next: Build reusable utilities that generate focused dashboard views in minutes.
Full write-up: The Pretty Picture →Teaching AI Through Demonstration
Hypothesis
People adopt AI faster when they see real workflows instead of feature lists.
Method
Demonstrated practical AI use cases to a colleague and encouraged exploration through questions rather than instructions.
Result
The demonstration revealed significant untapped potential — many people already have access to AI tools but don’t yet know how to collaborate with them.
Lesson Learned
AI adoption is as much about changing how people think as it is about learning new software.
Next: Document practical workflows that help others move from curiosity to confident experimentation.
Journey
Automation
Started where most of this starts — automating the parts of the job that were repetitive and obviously wasteful. Each win felt unambiguous.
Operations
Moved from single automations to connected workflows. The wins kept coming, and so did a pattern I didn’t have language for yet: solving one problem reliably surfaced the next one.
Systems
Started treating the organization as one connected system instead of departments to optimize separately. Personal experiments — Battle of the Bots, Project Sagebrush — became a second lab for testing the same ideas.
Strategy
Recommended The Goal for an unrelated reason and got hit by how directly a forty-year-old factory novel mapped onto the questions already forming. "You’re my Jonah" happened somewhere in here.
Research
Writing the questions down instead of just carrying them. This site, the Field Notes, and the working manuscript are the current form that takes.
Predictions
AI will lower the cost of building targeted internal tools.
Organizational understanding will become more valuable than software ownership.
AI may create digital inventory faster than organizations can evaluate or use it.
Decision latency may become a more important constraint than knowledge production.
The strongest AI implementations may be defined by the quality of questions rather than the quantity of output.