AI at Work
What AI gets wrong at work: the honest limits
A bookkeeper asked Claude to draft a response to a client about their overdue invoice. Claude drafted a professional, warm, reasonable message. It also included a specific payment deadline that was not real, a reference to a call that never happened, and the wrong invoice amount.
She caught it. Corrected it. Sent the right version. But if she had skimmed and clicked send, she would have sent her client a confident-sounding email full of wrong information.
That is the real AI failure mode in an office context. Not the tool being unhelpful. The tool being helpful-sounding and wrong. This article covers the failure modes that actually matter so you know what to watch for.
Hallucinated specifics
This is the most common and most dangerous error. AI fills in gaps with plausible-sounding specifics that it made up.
You ask it to draft an invoice. You did not give it the line items. It will draft line items. They will be specific, they will be formatted correctly, and they will be wrong.
You ask it to write a proposal. You did not give it your actual pricing. It will write pricing that sounds reasonable for the category and is incorrect for your business.
You ask it to summarize a contract. It will summarize. Some of what it says will accurately reflect the document. Some will reflect what contracts in this category usually say, which is not the same as what your contract actually says.
The fix is not to distrust AI output. The fix is to read every specific: every number, every date, every name, every claim that requires factual accuracy. Those are the places where it invents things. Verify them before anything goes out.
Confident tone on wrong information
AI does not qualify uncertainty well. A human who is unsure about a fact tends to hedge: "I think it's around $400, but let me check." AI tends to state things directly and confidently regardless of whether it knows they are true.
This is a problem because the confident tone is part of why AI output looks trustworthy. A document that sounds certain is more likely to be sent without verification than one that sounds uncertain. The writing quality actively works against careful review.
The practical response: treat every AI-generated specific as a question mark until you have verified it, regardless of how confident the sentence sounds.
Wrong tone for the relationship
AI defaults to professionally pleasant. That is the right tone for many business communications. It is not the right tone for all of them.
A message to a long-standing client you have worked with for eight years should sound like it comes from a person who knows them. AI will write it like it is going to a stranger. A message to a difficult customer who needs to be handled carefully will get the same professional-pleasant treatment as every other message. A message that needs to be firm about a boundary will come out soft.
The fix is to tell AI the relationship context before it drafts. "This is a client I have worked with for eight years. Keep the tone warm and direct, not formal." "This customer has disputed invoices before. The tone needs to be professional but not accommodating." Two sentences of context changes the output significantly.
Even with good context, read the draft for tone before you send it. You know the relationship. The tool does not.
Missing context about your actual situation
AI knows nothing about your business unless you tell it. It does not know your prices, your current job backlog, your staff availability, your customer history, or your standing policies. If you ask it a question that requires knowing those things, it will answer using what it knows about businesses like yours in general.
That is sometimes close enough. Often it is not.
A scheduling question ("what should I tell this customer about availability?") requires knowing your actual schedule. AI does not know it. A pricing question ("is this quote reasonable?") requires knowing your costs and margins. AI does not know them. A customer-history question ("should I offer this client a discount?") requires knowing the relationship and the history. AI does not have it.
The fix is to give it the relevant context before you ask. Paste in the schedule. Give it your actual pricing. Tell it the customer history that is relevant to the question. The more specific the input, the more useful the output.
Errors in specialized domains
AI is a generalist. It knows a lot about many things. In specialized domains, it knows enough to sound right while being wrong at the level of detail that matters.
Legal documents. Financial calculations. Technical specifications for a trade. Medical or clinical guidance. In these areas, AI can help you understand something or produce a first draft. It cannot substitute for a qualified professional reviewing the result before it is relied on.
The confident tone compounds this problem. A paragraph about contract terms that sounds authoritative may still be wrong about the specific jurisdiction, the specific contract type, or the specific clause that applies to your situation. AI is not a lawyer. It just writes like one.
Outdated information
AI models have a knowledge cutoff. Anything that changed after that cutoff is not in the model. Tax rules, regulations, software versions, pricing for services or materials. If the question depends on current information, AI may give you an answer that was accurate a year ago and is wrong today.
For anything time-sensitive, verify against a current source. Treat AI output on current information as a starting point that requires a check, not a final answer.
The compounding problem: sending without reading
Every error above is manageable. The error that makes them unmanageable is sending AI output without reading it first. Each individual mistake is catchable with a one-minute read. Skipping the read is the only way they reach your customers.
Read before you send. Every time. This is not a limitation of AI. It is the correct division of labor between a drafting tool and the person responsible for what goes out under their name.
The limits are real. They are also manageable. The most useful relationship with AI in an office context is to treat it as a fast first-draft generator and yourself as the editor who checks the specifics. That combination is faster than starting from scratch and more reliable than sending without review. Both halves need to be there.
Work smarter. Check what matters.
The AI Field Guides for business operators cover the real workflow: which tasks to hand to AI, which prompts produce reliable results, and where you have to stay in the loop.
Guides for business owners → See all guidesGo deeper
AI Field Guides from $9.99
Practical field manuals for real operators. Each guide covers one tool or workflow in full: the setup, the prompts, the limits, and the results. Available on Amazon and Kindle Unlimited.