When Your Boss Says AI-Generated Plans Are Better: Unpacking the Cognitive Bias

1. The manager’s illusion starts with a polished demo

The author acknowledges why a leader might be fooled: feeding a vague request to ChatGPT yields a well‑structured, nicely formatted proposal in seconds, which looks superior to a colleague’s rough work, leading the manager to proclaim full automation.

However, real business proposals must pass several rigorous checkpoints:

Legal review (compliance risks)

Financial calculation (numeric accuracy)

Business implementation (feasibility of execution)

Audit trail (accountability for errors)

Thus a demo only shows “does it look right”, while a product must answer “is it actually correct”.

2. Breadth is sweet, depth is poisonous

Many mistakenly think large models know everything. In reality they can chat about anything but do not truly understand.

Their strength lies in statistical pattern matching over massive text corpora, providing breadth for information retrieval, formatting, and draft generation. When depth is required—understanding causal chains, finding optimal solutions under constraints, or reasoning about completely new scenarios—the models quickly expose flaws.

The most dangerous failure mode is the model’s confident hallucination that sounds professional.

3. The 80% fallacy

Leaders often say “80% accuracy is good enough”. The author breaks down the math: assuming a single‑step accuracy of 95%, a workflow with multiple decision nodes compounds the error.

5 nodes: end‑to‑end accuracy = 0.95⁵ ≈ 77%

10 nodes: end‑to‑end accuracy = 0.95¹⁰ ≈ 60%

Running AI end‑to‑end on 100 proposals would leave 20‑40 of them with hidden mistakes—misplaced decimal points, wrong policy years, or concealed legal risks—without any visual cue of the error.

4. Practical rebuttal scripts

If you constantly face pressure to go “full AI”, the following lines can help protect your role:

Script 1: Highlight the multiplicative effect

"Leader, a 95% single‑step accuracy drops to about 77% over five steps. With thousands of documents a year, can we accept that error rate?"

Script 2: Shift responsibility

"If AI makes a mistake, who is liable? The model provider, the approver, or me? Verifying each output negates the efficiency gain."

Script 3: Use real cases to expose flaws

"Let’s compare a complex case handled by a human versus AI and highlight the seemingly reasonable yet actually wrong details—like a 2 million‑yuan loss hidden in the recommendation."

Script 4: Redefine AI’s role

"AI assists, not replaces. It can draft, format, and fill gaps, but final decisions, accountability, and quality checks must remain human responsibilities."

Providing a compromise gives the leader a path forward while preserving the author’s job security.

5. Final thoughts

The author does not dismiss the value of large models. They dramatically lower the cost of creating an initial draft, which is their true productivity boost.

However, low startup cost does not equal low verification cost. Bridging the gap from a draft to a usable product requires domain expertise, business experience, responsibility, and risk judgment—areas where current models are still lacking.

The real danger is not AI’s weakness but the overconfidence sparked by a well‑crafted demo that claims to shoulder the full weight of production environments.