AI can save you ten minutes and still make a task more expensive.
That happens when the draft is fast but the review is not. A routine internal summary may be a good candidate for automation. A legal filing, audit conclusion, board memo, disciplinary communication, or public statement is different: the cost of one confident error can be far larger than the time saved producing the first draft.
This article and the calculator below use four practical questions to decide how much AI should be involved: who will read it, how many facts must be verified, how much context lives outside the document, and what happens if it is wrong.
Short version: automate drafting where errors are cheap and easy to catch. Keep humans close where facts, relationships, or accountability make mistakes expensive.
The real cost is often verification
Generative AI produces fluent text quickly. Fluency is useful, but it creates a trap: a sentence can sound finished before anyone has checked whether it is true.
For a brainstorm, that is not a serious problem. If the model suggests twelve campaign ideas and three are nonsense, you discard them. For a report containing dates, regulations, quotations, financial figures, technical specifications, or named sources, every factual sentence creates review work.
Call that the verification tax. The more factual and consequential the document, the more of the apparent drafting speed you give back during checking.
A real example: fake case citations
In Mata v. Avianca, lawyers submitted court filings containing nonexistent judicial opinions generated by ChatGPT. The Southern District of New York imposed sanctions after the fake cases and quotations were presented to the court. The lesson is not that lawyers should never use AI. The court itself noted that using technology for assistance is not inherently improper. The failure was submitting material without fulfilling the human duty to verify it.
You can read the court’s sanctions order in Mata v. Avianca.
That example is extreme, but the mechanism is ordinary. A model generates something plausible, the reviewer recognizes the general shape, and confidence in the prose replaces verification of the details.
Risk factor 1: audience sensitivity
Ask who receives the output and what they reasonably expect from you.
An internal note saying “the meeting moved to 3:00” does not need the same level of authorship as a message explaining layoffs, a response to a regulator, or a proposal to a client who is paying for your judgment.
This is not about hiding AI use. It is about understanding that some communications carry relational information that is not in the prompt: history with the client, a previous disagreement, an executive’s preferences, a promise made verbally last month, or a phrase that is technically harmless but politically disastrous in that organization.
Risk factor 2: information integrity
Count the facts that would matter if they were wrong.
Consider two documents of the same length:
- a first draft of workshop ideas;
- a procurement comparison containing prices, contract dates, security certifications, and legal requirements.
Both might be 1,000 words. The second document is far more expensive to verify because the important unit is not word count; it is the number of claims that must be traced back to reliable evidence.
Example: If an AI turns a 40-page PDF into a five-bullet summary, you saved reading time only if somebody checks that the five bullets preserve the exceptions, dates, and conditions that actually matter.
Risk factor 3: strategic nuance
Some work depends on facts that are not written down anywhere accessible to the model.
A manager may know that Finance rejected the same proposal last year because of a conflict between two departments. A consultant may know that a client uses “pilot” to mean “no budget commitment.” An analyst may know that a data series changed methodology in 2024 even though the column name stayed the same.
AI can help structure the material, but it cannot reliably infer private institutional context that you did not provide. If the success of the message depends on that context, the human contribution is not cosmetic editing; it is the core of the task.
Risk factor 4: accountability
Finally, ask who owns the error.
If an AI-generated meeting summary misses a minor action item, you correct it. If an AI-generated compliance memo invents a requirement, the organization does not get to outsource responsibility to the model.
The higher the legal, financial, safety, employment, or reputational consequence, the less useful “but the AI wrote it” becomes as an explanation.
How to use the AI Reputation Risk Calculator
Score each dimension from 1 to 5. The calculator is deliberately simple. It is a triage tool, not a scientific risk model.
- Audience sensitivity: 1 for routine/internal; 5 for a client, executive, regulator, public audience, or sensitive recipient.
- Information integrity: 1 for mostly creative material; 5 for dense factual or technical claims.
- Strategic nuance: 1 when context is explicit; 5 when unwritten history and relationships matter.
- Accountability stakes: 1 when an error is easy to correct; 5 when an error can create legal, financial, career, or safety consequences.
AI Reputation Risk Calculator
What the score should change
A low score does not mean “send whatever the model produces.” It means the downside is limited enough that AI can do more of the mechanical work.
A middle score suggests a hybrid workflow: human creates the facts, position, and constraints; AI helps organize or compress; human verifies and rewrites the final version.
A high score means speed is probably not the main objective. Use AI for narrow support tasks if useful—outline alternatives, identify questions, improve formatting—but keep factual verification and final judgment with a person who understands the consequences.
One final test before you press send
Ask: Could I defend every important sentence if the recipient asked where it came from?
If the answer is no, the document is not finished. The problem is not whether it “sounds like AI.” The problem is that nobody has taken ownership of the claims yet.
Reference: U.S. District Court sanctions order, Mata v. Avianca, Inc. (2023).
