AI Business Professional common mistakes
AI Business Professional · Understand the traps and how to avoid them.
Treating generative AI output as verified fact without checking it against a source or running a calculation, then building a recommendation on it.
Candidates over-trust fluent, confident-sounding text and assume that because an answer is well-formed it is accurate; they also confuse 'the model said so' with evidence.
What to do instead
For any factual, numeric, or citation-bearing claim, require an independent check (primary source, recalculation, or a second tool) before it enters a deliverable, and state the verification step explicitly.
Describing AI as 'thinking', 'understanding', or 'knowing' and then reasoning about its capabilities from that anthropomorphic framing.
Everyday language about AI is mentalistic, so candidates import assumptions about intent, memory, and understanding that do not hold for statistical next-token prediction.
What to do instead
Re-describe the system in mechanism terms (training data, pattern learning, probability of next token, no persistent memory unless engineered) and answer capability questions from that description.
Confusing model training, fine-tuning, and retrieval-augmented generation, and recommending fine-tuning for a problem that is actually a missing-context or freshness problem.
Vendors and marketing blur the terms, and candidates learn 'fine-tuning' as a generic fix for 'the model doesn't know our stuff'.
What to do instead
Map the problem first: is it style/format (fine-tuning or prompting), private or current facts (retrieval/grounding), or task instruction (prompting)? Choose the intervention that matches the cause and name the trade-offs.
Writing a single prompt, judging the output, and concluding the tool 'can't do this' or 'works great' without revising the prompt or testing across varied inputs.
Candidates treat prompting as a one-shot query rather than an iterative specification, and they generalise from one lucky or unlucky sample.
What to do instead
Iterate: specify role, task, constraints, format, and examples; then test on a small set of representative and edge-case inputs and report results as a range, not a single anecdote.
Pasting confidential customer data, credentials, or unreleased financials into a public consumer chatbot to 'save time'.
The tool feels like a private workspace and the convenience is immediate; candidates underestimate that inputs may be logged, retained, or used for training depending on the plan and settings.
What to do instead
Apply a data-classification check before any prompt: if the data is confidential or personal, use only an approved enterprise instance with the right contractual terms, or redact/anonymise first.
Treating 'the AI is biased' or 'the AI is fair' as a verdict about the model in the abstract, rather than about specific data, labels, and use context.
Fairness is taught as a property of systems, and candidates look for a single pass/fail answer instead of tracing bias to training data, proxy variables, and deployment context.
What to do instead
Locate the harm: which population, which decision, which data or label produces the disparity, and which mitigation (data, features, thresholds, human review) addresses it. State residual risk.
Proposing an AI initiative without a baseline metric, a defined business owner, or a way to measure value beyond 'it saves time'.
Candidates focus on the technology demo and skip the business case; 'efficiency' feels self-evidently valuable so it is never quantified.
What to do instead
For every proposal, state the current baseline, the target metric, the measurement method, the accountable owner, and the cost to run it, then compare against the do-nothing option.
Assuming a pilot's success proves the workflow will scale, ignoring process redesign, change management, and exception handling.
Pilots are run on clean, willing teams with hand-holding; candidates extrapolate without accounting for the messy edge cases and adoption friction that appear at scale.
What to do instead
Separate the model's capability from the operating model: specify who handles exceptions, how the process changes, how staff are trained and measured, and what volume/edge-case testing is required before rollout.
Treating data quality, labelling, and governance as a one-time cleanup before an AI project rather than an ongoing operational requirement.
Data work is invisible in demos and candidates want to get to modelling; governance feels like paperwork rather than a control that prevents drift and rework.
What to do instead
Build data quality into the workflow: define ownership, freshness and accuracy checks, labelling standards, and review cadence, and treat degradation as a monitored risk with an owner.
Answering an ethics or compliance question with a general principle ('we will be responsible') instead of naming the specific control, owner, and escalation path.
Candidates know the vocabulary of responsible AI but have not practised translating principles into operational controls, so answers stay at the level of values.
What to do instead
Convert each principle into a testable control: what is checked, by whom, at which stage, with what evidence, and what happens when it fails. Name the accountable role and the review gate.