Value
Would better execution materially change cost, speed, capacity, risk, or customer experience?
Field notes from implementation
Practical thinking for changing how work gets done.
Decision frameworks, operating patterns, evaluation methods, and tool intelligence for leaders turning AI ambition into safe, measurable change.
Transformation design · Agent operations · Tool intelligenceHow to choose the first implementation wave for a company-wide AI program.
Map the company direction first, then choose a department where value, feasibility, ownership, and the path into later waves are strongest.
Use the first-wave scorecardResearch tracks
Three decisions that shape a production system.
System design · Decision framework
Build, buy, or orchestrate?
How to decide what to preserve, buy, build, and connect inside one operating system.
Open field note →Agent operations · Human judgmentThe exception queue is the product
How to design escalation, uncertainty, fallback paths, and learning before the happy path scales.
Open field note →Evaluation · Deployment patternFrom impressive demo to production evidence
How representative cases, shadow operation, and controlled authority turn a promising demo into operating evidence.
Open field note →First-wave scorecard
Five signals that a department is ready.
The scorecard does not promise automation. It identifies departments worth observing, baselining, and testing before architecture begins.
Frequency
Does the work recur often enough for improvement to compound and for evaluation to stay meaningful?
Boundaries
Can acceptable actions, escalation, permissions, and failure be defined?
Evidence
Can we assemble enough historical cases and source evidence for representative evaluation?
Ownership
Is there an operator who can teach the work and an accountable leader who can own the result?
Field note · Tool decisions
Build, buy, or orchestrate?
Most production systems combine existing software, selected AI products, and custom agents. The decision is where each belongs, who owns it, and how the parts work together.
Connect
Preserve the systems that still earn their place.
Keep reliable software and data flows, then connect them so context and action can move through the work without discarding what employees already trust.
Buy
Use a proven product when it closes the gap faster.
Choose an existing product when the capability is established, integrations are sufficient, and the workflow can adapt without losing important operating knowledge.
Build
Engineer the logic that is specific to the company.
Build when proprietary context, unusual controls, or a distinctive workflow creates value a general product cannot preserve.
Field note · Agent operations
The exception queue is the product
A production agent is defined less by the happy path than by what happens when evidence conflicts, a system fails, or judgment is required.
Context
Prepare before escalating.
Retrieve source evidence, explain the conflict, and show the allowed options so the operator receives a decision-ready case instead of another research task.
Priority
Make waiting visible.
Materiality, customer impact, age, deadline, and confidence should determine what reaches a person first.
Learning
Use overrides as evidence.
Every correction should improve the evaluation set, policy model, routing rule, or permission boundary before authority expands.
Field note · Evaluation
From impressive demo to production evidence
Quality scores matter only when they predict useful, safe behavior inside the operation the system will actually touch.
Offline
Start with representative cases.
Include ordinary work, edge conditions, policy boundaries, ambiguous inputs, and outcomes that would be unacceptable.
Shadow
Compare without authority.
Let the agent prepare recommendations beside employees before it can change a source system or communicate externally.
Controlled
Expand one permission at a time.
Release a bounded action, monitor traces and overrides, and widen responsibility only when operating evidence stays strong.