What AI assistants can and cannot do: knowing when to trust and when to verify

1. What you will learn

Most owners have now tried a chat assistant, and many staff already use one quietly for emails and captions. The owners who get real value are not those who use AI the most, but those who know exactly which jobs it does well, where it fails, and how much checking each job needs. By the end of this lesson you will be able to:

• list the task types where current AI assistants are reliably strong;
• list the failure patterns you must plan for, with the reason each happens;
• place any task on a trust-and-verify matrix using two questions: how costly is an error, and how easy is it to check;
• set a verification method for each task before anyone uses AI on it.

2. The idea explained

Where AI assistants are strong. Because they are trained to produce fluent, well-organised language, assistants perform well at:

• Drafting — first versions of emails, posts, proposals, job descriptions and SOPs.
• Rewriting — changing tone, length, reading level or language (English, Hindi, Hinglish, other Indian languages with varying quality).
• Summarising — condensing long documents, transcripts or message threads into key points.
• Structuring — turning messy notes into tables, checklists, outlines or step-by-step procedures.
• Classifying — sorting messages or reviews into categories such as complaint, enquiry or order.
• Extracting — pulling named fields (date, amount, invoice number) from text or documents.
• Generating options — names, angles, headlines, objections a customer might raise, interview questions.
• Explaining — describing a concept, a formula or a spreadsheet function in plain language.

Where they fail, and why.

• Invented facts (hallucination). The model produces what is plausible, not what is verified. It can invent statistics, laws, case names, product specifications and quotations.
• Out-of-date knowledge. Training stops at a cut-off date; rates, rules, prices and competitor details move on.
• Arithmetic and counting. Text prediction is not a calculator. Long sums, percentages across tables and word or item counts can be wrong unless the tool runs actual calculations and you check them.
• Fake or wrong citations. Asked for sources, a model may give real-looking but non-existent references or links that do not support the claim.
• Missing context. It does not know your customer's history, your verbal promises or your unwritten rules unless you provide them.
• Over-agreement. Assistants tend to accept the framing of a question. Ask "why is this plan good?" and you get reasons it is good, not a balanced view.
• Bias. Output can reflect stereotypes present in training data, for example in job adverts or customer profiling.
• Judgement and accountability. It cannot take responsibility for a hiring, credit, legal, medical or pricing decision. A person must.

The trust-and-verify matrix. For any task, ask two questions.

1. Error cost: what happens if the output is wrong? (Low: a weak caption. High: a wrong price in a quotation, a wrong tax figure, a harmful health claim.)
2. Ease of checking: can a knowledgeable person check it quickly? (Easy: a summary of a two-page letter you have read. Hard: a market-size statistic with no source.)

This gives four zones:

• Low cost, easy to check — use freely with a quick read. Social caption drafts, internal notes.
• High cost, easy to check — use with mandatory expert check. Quotations from your own price list, customer replies about orders, SOP drafts.
• Low cost, hard to check — use for ideas only. Brainstorming, possible customer objections; do not state these as facts.
• High cost, hard to check — do not rely on AI output. Legal interpretations, tax positions, medical or financial advice to customers, unsourced statistics in investor or bank documents. Use qualified professionals and primary sources.

Verification methods. Match the method to the output: compare with the source document; recalculate numbers in a spreadsheet; check rules on the official portal; ask the model to quote the exact passage it relied on and then read that passage yourself; have a second person review Tier 3 outputs.

3. Let us work through it

Step 1 — List ten tasks you or your team might hand to AI this month.

Step 2 — Rate error cost for each: low, medium or high, writing one line on what could go wrong.

Step 3 — Rate ease of checking: easy or hard, and who could check it.

Step 4 — Place each task in a zone of the matrix.

Step 5 — Write the verification method for every task outside the "use freely" zone, naming the source and the checker.

Step 6 — Share the list with the team as the first draft of your "AI allowed uses" note, which becomes part of your AI usage policy in Lesson 15.

5. Common mistakes and how to fix them

• Asking AI for facts and pasting them straight into documents. Fix: treat unsourced facts as leads and confirm from primary sources.
• Trusting totals and percentages in AI summaries. Fix: recalculate every number in a spreadsheet.
• Asking leading questions. Fix: ask for arguments for and against, and for the assumptions.
• Using AI for decisions that need accountability. Fix: AI may prepare options; a named person decides.
• Accepting citations without opening them. Fix: open every source and confirm it says what is claimed.
• Having no verification plan until something goes wrong. Fix: write the check method for each task before use.