Why 40% of agent projects get cancelled, and the four controls that save them
Gartner expects most agentic AI projects to be scrapped by 2027. The failures are operational, not model quality, and four controls keep a project alive.
Read
09 · Blog
Notes on running AI agents in production.
Practical guides and playbooks for deploying, monitoring, and safely automating AI agents inside your own environment. Proof over slideware.
- Governance
- Automation
- Assurance
- Deployment
The EU AI Act and US state laws for autonomous agents: what is enforceable now
A dated, sourced map of AI agent rules in mid-2026: which EU AI Act and US state obligations apply today, which are delayed, and what to build for now.
Read
What actually kills agent deployments (and how to not be a statistic)
Gartner expects 40% of agentic AI projects cancelled by 2027. The failures are operational, not model quality. Six documented patterns, the control for each, and a pre-mortem.
Read
When your AI vendor retires the model you rely on, your work can break overnight
When a vendor retires or quietly changes the AI model your work depends on, the instructions you tuned, the tests you trust, and your costs can break overnight, with no change on your side. Here is how to stay ahead of it.
Read
AI models you can download and run yourself: when they are the right call
A sourced guide to running AI models you can download and run yourself (Llama, Mistral, Qwen) inside your own environment, whether your work is regulated or not: what that really means, how to host them privately, and a clear rule for when to choose them.
Read
When your AI assistant shows someone a file they were never allowed to see
Build an AI assistant on your own documents the easy way and it can surface a salary letter or a sealed contract to the wrong person. The fix: the agent only sees what that person is already allowed to see, and personal details are masked.
Read
What you can and cannot prove about the AI models you rent from an outside lab
An honest account of what you cannot check about the leading AI models run by outside labs, why they keep it private, and what you can still control: your own tests, a record of every step, and locking the exact version you trust.
Read
Why an AI agent costs more than the bill suggests: retries, lookups, and dead ends
The price an AI vendor charges per word hides what real work actually costs. To finish one task an agent retries, looks things up, and hits dead ends, and each one multiplies the bill. Measure the cost to finish the task instead.
Read
Where your prompts go: what the big AI vendors keep, and what changed
A 2024-2026 map of what Anthropic, OpenAI, and Google keep from your prompts: retention windows, training use, consumer versus enterprise tiers, and the policy changes regulated teams must verify in a DPA.
Read
A one-page record for every AI agent before it ever touches real work
An AI agent nobody has written down is an agent nobody can answer for. One short record per agent names who owns it, what it is allowed to touch, what it can spend, how risky it is, and how to undo it, all before its first real run.
Read
Cost per finished job: the one AI number your CFO will actually trust
What an AI vendor bills per word tells you almost nothing about real work. The cost to finish one actual task, a cleared invoice or a resolved ticket, is the number finance can underwrite and the one that tells you when to automate more.
Read
Running AI agents inside your own walls, where your data is allowed to live
A field guide to keeping your AI agents inside your own environment: your prompts, documents, and a record of every step stay on systems you control, and every request passes through one checkpoint you can see.
Read
How to know when an AI agent is safe to let run on its own
A step-by-step path where an agent earns more responsibility only as it proves itself, with the proof required at each step and the high-stakes work, like a claim or a payment, that should never skip a step.
Read