My Stanceon AI

00/Position

01/In practice

Firstly, I must define my terms: when I say “AI,” I mean generative AI - the large language models that have come into the mainstream since November 2022. I do not refer to machine learning, natural language processing, robotic process automation, or the like. While the market often lumps those all together, I consider the former to be mature, reliable technologies, and generative AI to be useful, but ripe for misuse.

I use AI internally and externally.

Internally, it’s an aid to my work: I use it to speed up writing and reviewing code, to organize operations and documents, and to think through plans. I use AI every day to supplement my workflow.

Externally, it shows up as a component inside the workflows I build - e.g., scanning a bill of materials, classifying unstructured documents, drafting first-pass categorizations for human review.

I avoid using AI on public-facing creative products such as writing and design. That is where human touch is the differentiator.

My external stance follows the opening line. I apply the same discipline to AI that I apply to any capital decision: it has to beat the alternative on the outcome, not on novelty. Most of the time a different solution is better, cheaper, and more trustworthy - that’s what I ship.

This means that I am not an “AI-implementation service” - I provide infrastructural and retained solutions for data and finance. Thus, I evaluate each problem and opportunity on its face and find the right fit. Sometimes that solution is best served by a large language model (AKA “AI” or “genAI”) and in those cases, it becomes part of my recommendation and implementation plan. But most often, AI is the cherry on top of the real work: people interviews, foundational data engineering, scaffolding workflows, and automation.

A strong data foundation lets me surface insights and run processes that are often beyond the reach of traditional finance (more on those capabilities here). Those finance insights are, in turn, often out of reach in traditional organizations, where the technical teams and the business teams are siloed, or don’t share enough vernacular to know what’s possible. Once the base is solid, adding AI as a layer on top is both a much more straightforward lift and far more effective once in place.

In my experience, and in my peers’, mainstream engineering combined with business fluency solves nearly every organizational need around analytics, reporting, and automation - for nearly every client. More importantly, it avoids the risks that AI implementations pose: models that aren’t properly scaffolded and produce answers that are extremely confident and flatly wrong, to say nothing of the cybersecurity exposure and the added cost.

I’ve likewise found that AI, when it is used, works best as one part of an automated workflow. I might use AI to scan a bill of materials, but the result of that scan is fed into a much more structured, deterministic automation, reducing the risk of hallucinations and other invisibly incorrect outputs. Importantly, using deterministic automation keeps the workflow auditable - each number and output can be traced, which is a constant challenge with large language models and deep learning systems.

I do recognize that AI is a useful and remarkable tool. That’s why I stay current - in technical proficiency and in the latest developments - across both the frontier and local, self-hosted models.