The Flag System
Calibrating scepticism: not every AI output needs the same scrutiny, and treating them as if they do is its own problem
This essay sits primarily under the Are we ready? question of the practice’s methodology: calibrating the firm’s scepticism so AI outputs receive review proportionate to their risk, rather than uniform alarm or uniform trust.
This essay sits underneath the Trust Tool. The Trust Tool tells you which quadrant an output belongs in. The flags tell you what to check once you are in a verify quadrant. It is a second check, not a separate go/no-go system. Project-level gates belong to the AI project lifecycle.
The core idea
Not every AI output needs the same level of scrutiny. A brainstormed list of marketing angles for the firm’s website does not carry the same risk as a legal citation, a medication interaction, or a tax position. Some outputs are low-risk. Some should make you stop and verify before acting on them.
Treating everything as high-risk leads to fatigue: people stop checking the things that actually matter because everything has been treated as if it matters equally. Treating everything as low-risk leads to professional negligence. The flag system gives your team a quick way to calibrate their response. It is a second check on the output in front of you, not a traffic light for whether to greenlight a project.
It works for AI output, but it also works for anything else that crosses your desk: a report, a memo, a colleague’s claim in a meeting. The underlying skill is the same: reading with appropriate scepticism.
Green: low scrutiny appropriate
Internal-only, low-stakes, non-binding. Reversible if wrong. Examples:
- Brainstorming meeting agendas, internal newsletter angles, marketing taglines
- Restructuring an internal SOP draft for clarity
- Generating discussion questions for a team meeting
- Summarising a public document to decide whether to read it in full
The check here is light: does it read sensibly? Use it.
Yellow: verify before acting
A yellow flag means you need to verify before relying on the output, not that the output is wrong.
| Signal | What to do |
|---|---|
| AI cites a source, statute, case, paragraph, or guideline | Open the primary source and read it yourself before quoting or relying on it |
| The output names dates, numbers, or definitions you have not personally seen elsewhere | Cross-check against an authoritative source |
| The output summarises the contents of a long document you provided | Spot-check against the original; summaries omit and reframe more than people expect |
| The reasoning is plausible but you cannot follow each step | Ask the model to show the steps; verify the reasoning yourself |
| The output applies a general rule to your specific facts | Check that the rule actually applies; exceptions are usually where the work lives |
The common thread is missing provenance. The information might be perfectly accurate, but something about how it reached you (an absent source, a generated number, an applied rule) means you cannot yet trust it enough to act on it. Yellow flags mean check before you proceed.
In regulated practice, almost any output that names a section, a case, a guideline, a dose, a rate, a threshold, or a deadline starts as yellow until verified.
Red: high risk; do not act without independent verification
These are patterns associated with confident, plausible, professionally dangerous output.
| Signal | Why it matters |
|---|---|
| AI provides a citation that fits the argument too neatly | Models confabulate citations that look correct. Verify the case, paragraph number, statute, or DOI exists and says what is claimed |
| AI offers a clinical, legal, or financial recommendation that crosses into advice | Beyond a draft, this becomes a professional act subject to your registration. Treat the AI’s output as an input to your judgement, not as the recommendation itself |
| The output is unusually confident on a contested or jurisdiction-specific question | Australian regulatory framing is a common gap; many models default to US analogues |
| The output makes a positive claim about a person, business, or matter you cannot independently verify | Defamation, complaints, and insurer issues live here |
| The output recommends an action that is irreversible or has external consequences | Sending, lodging, advising, prescribing: anything you cannot pull back |
Red flags are qualitatively different from yellow flags. Yellow flags say the evidence is incomplete. Red flags say the cost of being wrong is high enough that incompleteness is not acceptable. Most professional negligence claims arise from confident, plausible output that nobody checked because nobody felt the need to, not from outright fabrication.
A note on misinformation flags
The flag system also catches a category of output that is technically coherent but reasoning-compromised: claims that “they” are hiding the truth, lone-genius framings against established consensus, sweeping statements about whole groups, extraordinary claims with no evidence. These are rare in professional AI output, but worth recognising: they share a common structure of asking you to trust more and verify less. That inversion is the hallmark of misinformation in any source, AI or otherwise.
The short version
The flag system is about proportional scepticism. Most output is fine. Some requires checking. A small amount should trigger real alarm. Knowing which is which is a skill, and it is a skill that AI will not develop for you. Your firm has to bring it.
Calibration like this is what agency looks like in everyday practice. The flag system is the active stance of deciding, again and again, what the firm will and will not accept on its name. The partner’s job is to make this calibration visible: to model it, to teach it, and to make space for staff to slow down on the work that needs slowing down.
About this work
borck.consulting helps regulated Australian small firms build proportional scepticism into their AI workflows from the start, rather than retrofitting it after the first incident. Engagements begin with the AI Readiness Diagnostic.