Writing
Agents, and the systems they have to live in
Most writing about enterprise AI is about what models can do. Almost none of it is about what the systems underneath will actually permit. That gap is where programs die, and it is what I write about.
Start here
-
The Reversibility Ladder
Where agents may write in a system of record
Agent authority in enterprise systems should be governed by how hard a write is to undo, not by how capable the model is. A field guide from inside ERP.
-
Pilots Die at Integration, Not Intelligence
The demo ran on a spreadsheet. Production does not have one.
Enterprise AI pilots rarely fail because the model was not smart enough. They fail at the seam between the model and the system of record, and that seam is almost never in the pilot's budget.
More
-
Change Management Is the Actual Bottleneck
The system went live. The behavior did not change. Those are separate events.
Enterprise AI programs measure delivery and assume adoption. The gap between a working system and changed behavior is where most of the promised value is lost, and it is a design problem rather than a communications one.
-
Evals Are the Interview Question You Will Fail
Everyone says they evaluate. Almost nobody can describe what they measured.
Ask a team how they know their agent works and the answer is usually a demo and a feeling. A practical standard for evaluation inside enterprises, including the parts that are genuinely hard.
-
The Cost Curve Nobody Models
Routing between local and frontier models is an architecture decision, not a procurement one
Most enterprise AI business cases price a single model against a single workload. The real cost structure is a routing problem, and the variables that dominate it are rarely the ones in the spreadsheet.
-
What Manufacturing Taught Me About Agent Failure Modes
The shop floor has been running autonomous systems against physical reality for decades
Manufacturing solved the hard parts of autonomy long before agents existed: what to do when the system and the world disagree, and how to fail without stopping the line.
-
Migration Archaeology
Reading a system nobody in the building understands anymore
Legacy systems are not badly documented. They are documented in behavior, in the corrections people make around them, and in decisions whose authors have retired. A method for reading them before you replace them.
-
Seven Places an Agent Must Never Write
A list I would hand a program on day one, before anybody picks a model
An opinionated list of the write operations inside an ERP that should stay closed to agents regardless of model capability, and the specific reason each one is on it.
-
Master Data Is the Real AI Readiness Test
You do not have a model problem. You have four customer records for one customer.
Enterprises assess AI readiness by looking at infrastructure and skills. The binding constraint is usually master data, and it is measurable long before a model is chosen.
-
What Finance Actually Needs Before It Will Accept an Agent
The objection is not fear of the technology. It is a different question entirely.
Finance teams reject AI agents that engineering considers proven. They are not being conservative. They are applying a standard that accuracy does not satisfy, and that standard is worth learning before the meeting.