How to Use AI Without Losing Accountability

A practical framework for using AI where it is strong, avoiding unreliable answer-first workflows, and keeping professional accountability intact.

Most approaches to generative AI fall into one of two traps: trusting it completely and getting burned by fabrications, or avoiding it entirely and missing what it does well.

For accountable work, ask AI to implement your instructions in a form you can verify.

The Problem with Answers

AI is good at:

• Translation: turning natural language into code, intent into implementation, one format into another
• Pattern detection: finding structure in data volumes no human could process
• Generation: producing candidates, drafts, variations, possibilities
• Iteration: refining and expanding without fatigue

Two weaknesses limit how we can use those capabilities:

• Consistency: ask the same question twice, get different answers
• Reliability: outputs sound plausible whether they’re correct or not

A plausible output may be wrong, and repeating the request does not guarantee convergence on the truth.

Verifying a direct AI answer requires the expertise to produce it yourself. With that expertise, AI can save time but adds no capability. Without it, you have to take the model’s word for the result.

This is the fundamental problem with asking AI for answers.

Thorough dialogue improves answers: you can probe assumptions and ask the AI to challenge your reasoning. It still cannot provide a verifiable account of how it produced the answer. The answer may be correct, but the conversation cannot establish that, and asking again tomorrow may produce a different result.

Executable Instructions

Ask AI to implement instructions in an executable artifact.

You can inspect the implemented steps, check them against your intent, and trace the result to inputs you control.

This changes what expertise means. You don’t need the expertise to produce the answer. You need the expertise to specify what you want and to verify that the method is sound. Specifying and verifying take less time than producing, assuming you know enough to do all three.

For consequential work, anything someone pays for or needs to defend, this becomes a concrete pattern I call Executable Knowledge Architecture:

1. The expert specifies intent
2. AI translates that intent into an executable artifact
3. The human verifies the artifact
4. The artifact produces reproducible results

An executable artifact records the expert’s judgment as instructions a machine can follow. You can inspect and test the method, then reproduce the output from the same inputs.

Different stakeholders have different obligations to these artifacts. The expert who specifies bears different responsibility than the operator who runs them. Capability governance maps who verifies what, who bears what risk, and what track record justifies what autonomy.

Current AI benchmarks compare AI answers with human answers. They do not measure method quality, defensibility, or reproducibility, so they do not assess this approach.

The Foundations

You do not need the theory to use this approach. The underlying claims about knowledge and combined human and AI capabilities explain why it works and where it leads.

I define knowledge as the ability to produce outcomes, rather than as mental states or beliefs. The definition makes human and AI contributions commensurable: both have capabilities, and together they can produce outcomes neither achieves alone.

Tacit knowledge is knowledge we have but cannot articulate. Tacitness depends on the agents involved: what one agent cannot express, another may be able to articulate. Human difficulty expressing something does not establish a permanent limit.

AI creates new ways to transfer knowledge. Expertise that could previously be transferred only through apprenticeship or years of practice is becoming encodable, reducing the domain of tacit knowledge.

Software and Organizational Knowledge

The pattern scales beyond individual artifacts.

Two barriers have prevented most expertise from becoming software: you needed to know how to code, and you needed to understand how systems work. Agentic AI removes both barriers. When they fall, the intermediary layers we call “applications” face pressure.

As AI capabilities improve, work that once required custom development requires only specification. The design approach changes: developers build around AI from the start.

The DIY knowledge worker made sense for forty years. AI offers the autonomy of doing the work yourself with the additional capacity of delegation.

Each engagement produces verified, executable records of methods that previously required human expertise. Others can inspect and reuse them, reproducing the method at zero marginal cost. Organizational knowledge accumulates as working code that people keep current through use; documentation often goes stale.

At scale, AI becomes the primary interface between human intent and executable systems. People specify what they want, AI translates the specification, and the artifacts execute it.

This change will happen. You will either encode expertise or have your expertise encoded.