How I Work With AI

An AI-enabled workflow is only as dependable as the problem definition, context, data, rules, permissions, evaluation, and human responsibility around it.

Before scaling a use case, I look for what it depends on: clear business definitions, relevant and permitted information, reliable source systems, appropriate access, realistic evaluation, ownership, maintenance, and a way to recognize and respond when something goes wrong.

More data isn't automatically better. More automation isn't automatically progress.

A faster workflow built on conflicting definitions, unclear permissions, or no accountable owner can scale confusion faster than a manual process ever could.

When the foundation isn't ready, I say so. The right next step may be to repair data, clarify the workflow, document a decision rule, establish ownership, test a lower-risk version, or defer the technology until the organization can use it responsibly.

I want the work to leave your team stronger: clearer on what its systems mean, more confident in how they operate, and better able to make the next decision without starting over.

AI is useful when it helps people spend less time on repetitive, low-risk work and more time on judgment, creativity, service, investigation, and decisions requiring human context.

Appropriate uses often include categorizing or summarizing information, searching an approved knowledge base, drafting from reliable source material, supporting research, organizing routine administrative work, and assisting bounded workflows with clear context and feedback.

For these, a strong harness matters. Define the approved sources and context, the expected output, the limits of what the system may do, the quality checks, the error conditions, and the feedback process. Use initial human review, sampling, evaluations, thresholds, or ongoing monitoring according to the risk.

AI doesn't earn its place by producing an answer quickly. It earns its place when it improves a defined workflow and makes the people involved more capable.

AI should support human capability. It should not be used to make human accountability disappear.

The level of review should match the consequence of the work. Low-risk tasks can be bounded, tested, monitored, and improved through feedback. Higher-impact systems need stronger safeguards and meaningful human judgment.

Meaningful review isn't a person clicking approve at the end of a workflow. It requires:

1. Subject-matter expertise to recognize a weak, misleading, incomplete, or harmful output
2. The ability to add context the system may not have
3. Authority to override, reject, or change the output
4. A documented escalation path when the issue can't be resolved inside the workflow

A person who cannot challenge the output is not providing meaningful oversight. They are being asked to approve a system they do not control.

I won't recommend AI as the final decision-maker in work that materially affects a person's life, livelihood, safety, rights, health, legal standing, financial standing, employment, or access to essential services. AI may inform a qualified person's decision. It doesn't replace that person's responsibility.

I also don't treat AI as a substitute for work depending on empathy, intuition, human connection, or hands-on assessment of a real situation.

I expect AI systems to be understandable at the level their role and risk require.

That means being able to explain what the system is for, what information it uses, what it can't reliably determine, what it's allowed to do, how its output is tested, where its boundaries are, how important outputs can be traced or challenged, and who owns the result.

For a consequential workflow, a polished demo isn't enough. The design should make purpose, source data, business rules, permissions, assumptions, failure conditions, review points, acceptance criteria, maintenance, and escalation visible before anyone relies on it.

The goal isn't false certainty or perfect automation. It's a system that can be tested, questioned, corrected, maintained, and governed by the people responsible for its consequences.

Technical performance matters. Accuracy, reliability, latency, security, cost, and operational resilience matter. So do adoption, trust, accessibility, maintained quality, and whether the people doing the work experience a real improvement.

A system people route around hasn't created much value. A system people use enthusiastically but can't challenge or safely correct isn't a success either.

I look for evidence at three levels:

A small, bounded workflow that earns trust and teaches the organization how to operate AI well can be worth more than a large pilot producing a slide deck and no durable change.

I use AI in client work deliberately, not casually.

When AI is relevant to an engagement, I use approved tools and apply controls appropriate to the work. I minimize the information shared, use redaction, anonymization, or de-identification where practical, and don't put client-confidential information into public models.

For material use of client data in an AI-enabled workflow, I seek client-specific approval and consider the relevant tool terms, data-processing commitments, training use, subprocessor access, data location, security controls, retention, deletion, logging, and change-notification practices.

I apply purpose limitation and least-privilege access. Data should be used only for the agreed purpose, and people and systems should get only the access the work requires.

I'm happy to disclose when AI materially assists a deliverable. Whatever the tools, I remain responsible for the work and the advice.

This site lets AI crawlers read it, including the ones that collect pages to train models. That is a decision, not a default, and the reasoning is written into the site’s robots.txt file as well as here.

Search and retrieval crawlers are how search engines and AI answer tools find and cite these pages. Blocking them would make the site harder to find for the people it’s written for.

Training crawlers are a separate question, and I allow them on purpose. Everything here is published to be read and learned from, and client material is never published here. Search visibility doesn’t depend on that choice either way, so it can be reversed without cost.

One crawler is blocked: it states no purpose and documents no way to control it.

I review this twice a year. The next review is March 2027, and if the decision changes, this section changes with it.