Responsible AI Use
JHarris Advisory uses commercially available AI tools in the operation of the practice. This page explains where those tools show up in the work, what they are not allowed to do, and who answers for the result.
Advising clients on AI governance while staying quiet about my own use of AI would be a strange way to run a practice. So here is the disclosure, along with the operating rules I apply to my own work.
Where AI shows up
Typical uses include:
Research. Locating authorities, regulatory text, enforcement actions, and secondary sources, and organizing what I find.
Analysis. Comparing frameworks, mapping obligations across overlapping laws, and pressure-testing a position before I commit to it.
Business evaluation. Vendor comparisons, market and product research, and internal decisions about how the practice runs.
Drafting and editing. First drafts, structural revisions, tightening, and formatting.
Presentations and visual materials. Slides, diagrams, tables, and client-facing explainers.
Quality assurance. Checking my own work for gaps, internal inconsistencies, missed authorities, and unclear language.
What AI does not do
AI does not determine JHarris Advisory's professional recommendations. It does not decide what a client should do, how a risk should be rated, whether a control is adequate, or what position to take with a regulator. Those are judgment calls, and they are mine.
No client-facing recommendation, risk determination, or deliverable is issued on the strength of a model output alone. A tool can surface an option. It cannot choose one.
Human review
Review means I read the whole thing, evaluate the substantive analysis, verify material factual and legal assertions as appropriate, and make the substantive calls myself. It does not mean a glance at the summary.
Confidential and client information
Client and confidential information is handled according to the security and contractual controls appropriate to it. In practice, that means a few things:
I evaluate a tool's data-handling terms and controls before using it with client information, including, where relevant, whether inputs are used for model training, applicable retention practices, processing locations, and subprocessors. I use business or enterprise configurations where appropriate for client work, rather than relying on consumer configurations without evaluating their data-use terms. I limit what goes into a tool to what the task actually requires, and I de-identify or generalize where the work allows it.
Where an engagement agreement, a client's own AI policy, a confidentiality obligation, or applicable law restricts AI use, that restriction controls. If a client prefers that no AI tools be used on their matter, that preference is honored. Just tell me.
Accuracy and appropriateness
AI output is treated as a draft from an unfamiliar source, not as a finding. Citations are checked against the primary source. Regulatory claims are checked against the actual text of the rule, guidance, or decision. Statistics and dates are verified or cut.
Appropriateness gets a second look on top of accuracy. A technically correct answer can still be wrong for the audience, wrong for the risk posture, or wrong for the specific facts of the engagement. Tools do not know the difference. That check is human.
Accountability
JHarris Advisory remains responsible for the work issued under its name. Use of an AI tool does not shift responsibility for the analysis, recommendation, or deliverable to the tool or its provider.
What "AI-assisted" means here
Some content published under the JHarris Advisory name is AI-assisted. Assisted is not the same as generated. The level and type of assistance varies by piece. The presence of AI assistance, or technical provenance indicating that an AI tool processed a file, should not be understood to mean that the underlying analysis, conclusions, or work product were predominantly generated by AI. Where applicable law requires a specific disclosure about AI-generated or AI-manipulated content, that disclosure is made separately.