What AI Actually Does (And Doesn't)

How AI Actually Works

Part 1: What's Really Happening When You Type a Prompt

In the spring of 2023, the marketing team at Heinz ran an experiment. They asked several AI image generators to produce 'ketchup.' Every single tool, without being told the brand name, generated an image of a red bottle that looked unmistakably like Heinz. The team was delighted, but also a little unsettled. How did the AI know? Nobody programmed it with the rule 'ketchup equals Heinz.' The AI had absorbed millions of images and captions from the internet and, through sheer pattern repetition, had learned an association so strong it reproduced it on demand. Heinz turned this into a marketing campaign. But the moment also exposed something most professionals don't fully grasp: AI doesn't think the way we do. It recognizes patterns at a scale no human can match.

The Heinz team's creative director, Cristina Kenz, described the result as 'the ultimate proof point' for brand recognition. But what she was really witnessing was a statistical mirror. The AI had seen enough human-generated content, recipe blogs, grocery store photos, Instagram posts, Wikipedia articles, that it could reconstruct the world's most associated image for a given word. This is the core mechanic behind every AI tool you will use as a manager: massive pattern recognition trained on human-generated data. It's not magic. It's not consciousness. It's a very sophisticated, very fast system for predicting what comes next based on what it has seen before.

That distinction matters enormously in practice. When you understand that AI tools like ChatGPT, Claude, or Microsoft Copilot are pattern-matching engines rather than reasoning minds, you start making better decisions about when to trust them, when to push back, and when to add your own judgment. The managers who get the most out of AI aren't the ones who treat it like a search engine or a magic oracle. They're the ones who understand the basic mechanics well enough to work with the grain of the tool rather than against it. This lesson gives you that foundation, no technical background required.

The Pattern Machine: Where AI's Power Actually Comes From

Picture a new hire who has read every email, report, proposal, legal brief, news article, and novel ever published, in multiple languages, before their first day on the job. They haven't lived any of it. They haven't made decisions, felt consequences, or built relationships. But they've seen so many examples of how professionals communicate that they can produce a passable first draft of almost anything you ask for. That's roughly the situation with a large language model (LLM), the technology behind ChatGPT, Claude, Gemini, and Copilot. The term 'large language model' just means a system trained on enormous amounts of text to predict what words should follow other words. GPT-4, the model behind ChatGPT Plus, was trained on an estimated 1 trillion words of text.

Here's where it gets practically useful. Because these models have seen so many examples of professional writing, they've absorbed implicit rules about structure, tone, and format that most people never consciously articulate. Ask ChatGPT to write a performance review for a mid-level employee who missed targets but showed growth, and it will produce something that follows the unwritten conventions of HR writing, balanced language, specific-sounding examples, a constructive close. It learned those conventions by reading thousands of real performance reviews. It doesn't know your employee. But it knows the genre. That's a genuine and significant capability you can put to work.

The limitation is the flip side of the same coin. Because AI is predicting based on patterns rather than reasoning from facts, it can produce outputs that sound authoritative but are wrong. This is called 'hallucination' in the industry, a slightly dramatic word for what is essentially confident confabulation. The AI fills in gaps with plausible-sounding content. A legal team at a New York law firm learned this the hard way in 2023 when an attorney used ChatGPT to research case precedents. The AI invented several convincing but entirely fictional court cases. The attorney submitted them to a federal judge. The resulting sanctions made international news. Pattern recognition without verification is a liability.

Click to enlarge

A Hospital System Learns to Work With the Grain

In 2023, UC San Diego Health piloted an AI tool, built on GPT-4, to help physicians draft responses to patient messages in their MyChart portal. Before the pilot, doctors were spending an average of 24 minutes per day just on inbox management. Patient messages asking about medication interactions, test results, and appointment logistics were piling up. The AI draft tool didn't answer patients directly. It generated a suggested response that the physician could review, edit, and send. The key design insight: the AI handled the genre work (professional, empathetic, medically cautious language) while the physician provided the factual accuracy and clinical judgment.

The results were striking. Physician satisfaction with inbox management improved significantly, and patients reported that response quality felt more thorough, likely because doctors had more time to focus on the substance rather than the mechanics of writing. This example illustrates a principle worth internalizing: AI performs best as a first-draft engine, not a final-answer machine. The physicians at UC San Diego weren't replacing their expertise. They were offloading the cognitive labor of formatting and phrasing so they could apply their expertise more precisely. That's the model that works across industries.

What Different AI Tools Are Actually Built For

A Sales Director Discovers the Difference Between Tools

Rachel Garrett, a sales director at a mid-sized B2B software company in Austin, started using AI tools in early 2023. Her first instinct was to use ChatGPT for everything, drafting proposals, summarizing competitor research, writing follow-up emails. It worked well for drafts. But she noticed that when she asked ChatGPT to summarize a 40-page RFP (Request for Proposal) that a client had sent over, she had to paste the document in chunks, losing context between sections. When her company moved to Microsoft Copilot through their existing M365 subscription, the workflow changed. She could open the RFP in Word, ask Copilot to summarize the key requirements, and get a structured output in under two minutes, without leaving the document.

Rachel's experience points to something the comparison table above captures: the right tool depends on where your work lives. If your team runs on Microsoft 365. Outlook, Word, Teams, SharePoint. Copilot is almost certainly the highest-leverage tool because it works inside your existing environment. If you're a Google Workspace shop, Gemini makes the same argument. ChatGPT and Claude are more powerful for complex reasoning and long-form drafting, but they require you to bring the content to them. Neither is universally better. The best tool is the one that fits your workflow with the least friction.

Why 'Talking to AI' Is a Skill Worth Developing

Most professionals who feel disappointed by AI tools have one thing in common: they're treating them like a search engine. They type a short, vague query and expect a precise, useful result. 'Write me a proposal' returns something generic. 'Summarize this' returns something too broad. The frustration is real, but the cause is usually the instruction, not the tool. AI tools respond to context, specificity, and structure. The more clearly you describe who you are, what you need, who the output is for, and what format you want, the more useful the response becomes. This isn't a technical skill. It's a communication skill, one that most managers already have in other contexts.

Think about how you brief a junior colleague on a task. You don't just say 'write a report.' You say: 'Write a two-page summary of last quarter's sales performance for the CFO. Focus on the three regions that missed targets. Use plain language, she doesn't want jargon. Include a bullet-point summary at the top.' That same specificity, applied to an AI prompt, produces dramatically better results. The industry term for this is 'prompt engineering,' but that phrase makes it sound more technical than it is. It's really just clear briefing. You already know how to do this. You just need to apply it to a new kind of collaborator.

The practical implication is that your value as a manager doesn't decrease because AI can draft documents quickly. It increases, because the quality of what AI produces is directly proportional to the quality of the direction you give it. A vague brief returns a vague draft. A sharp, specific brief, with context about the audience, the purpose, the tone, and the constraints, returns something you can actually use. The professionals who will get the most out of these tools are the ones who are good at clarifying what they actually want. That's a management skill, not a technical one.

Key Takeaways from Part 1

1. AI tools like ChatGPT, Claude, and Copilot are pattern-recognition systems trained on vast amounts of human-generated text, they predict useful responses, they don't reason the way humans do.
2. This makes AI highly capable at genre-based tasks: drafting professional documents, reformatting content, summarizing long text, and generating structured outputs in a specific style.
3. The same mechanic creates a real risk: AI can produce confident, plausible-sounding content that is factually wrong. Always verify facts, statistics, citations, and names before using AI output professionally.
4. Different tools are optimized for different environments. The best tool is usually the one already integrated into where your work lives. M365 users should explore Copilot; Google Workspace users should explore Gemini.
5. Prompt quality determines output quality. Specific, context-rich instructions, describing your role, the audience, the format, and the purpose, consistently produce more useful results than short, vague queries.
6. AI works best as a first-draft engine. Your job is to provide direction, add facts and context the AI can't know, and apply judgment the AI can't exercise. The combination of AI speed and human judgment is where the real productivity gain lives.

When AI Gets It Wrong. And Why That Matters for Your Team

What happened here cuts to the heart of how AI language models actually work. The model didn't lie in the human sense, it didn't know the cases were fake. It generated text that looked statistically like what a real legal citation should look like, because that's exactly what it was trained to do. It had seen thousands of legal briefs and knew the pattern: case name, court, year, ruling summary. So it produced that pattern, fluently and confidently, even when the underlying facts didn't exist. This is called a hallucination, and it's one of the most important concepts any professional using AI tools needs to understand.

The principle extracted from this story isn't 'don't use AI for legal work.' It's more nuanced and more useful: AI is extraordinarily good at producing text that looks correct, but it has no internal alarm that fires when something is factually wrong. Confidence in AI output is stylistic, not epistemic. The model writes with the same assured tone whether it's summarizing a real report or inventing one. Your job as a professional isn't to be suspicious of everything AI produces, it's to know which outputs require verification and which are safe to use as-is.

The Pattern Machine: How AI Learns to Sound Like an Expert

To understand why hallucinations happen, you need a working mental model of how these systems learn. Imagine you're training a new hire by giving them access to every business book, every industry report, every email thread, every news article, and every academic paper ever published, and then asking them to read all of it before their first day. They'd emerge with an extraordinary grasp of language, concepts, tone, and structure. They'd know what a good executive summary sounds like. They'd know how marketing plans are typically structured. They'd be able to draft a performance review that hits all the right notes. But they wouldn't have lived any of it. They'd be pattern-matching at enormous scale.

That's essentially what large language models do during training. They process billions of words and learn to predict which words follow which other words across an almost incomprehensible range of contexts. The result is a system that has absorbed the shape of human knowledge without experiencing the world directly. It knows that capital cities are usually mentioned alongside country names. It knows that financial reports contain phrases like 'year-over-year growth' and 'EBITDA margin.' It knows that apology emails tend to start with acknowledgment before moving to remedy. This structural knowledge is genuinely powerful, and it's why AI tools can accelerate so much professional writing and thinking work.

But here's the catch that professionals need to keep front of mind: the model's knowledge is frozen at a point in time. ChatGPT-4's training data has a cutoff. Claude's does too. Unless the tool is explicitly connected to a live search engine, like Copilot in Microsoft 365, which can pull from the web, or ChatGPT with browsing enabled, the AI is working from memory, not from today's reality. Ask it about a law that changed last quarter, a competitor's new pricing, or your company's current headcount, and it will either tell you it doesn't know (if it's well-designed) or fill in the gap with something plausible-sounding (if it's not being careful). Knowing this shapes how you use these tools wisely.

A Hospital System That Got AI Adoption Right

In 2022, a regional hospital network in the American Midwest began piloting AI writing tools for their administrative and communications teams, not clinical staff. The goal was simple: reduce the time nurses and unit coordinators spent drafting internal memos, patient communication templates, shift handover summaries, and staff scheduling updates. Within six months, the communications team reported saving an average of four hours per week per person on drafting tasks. The key to their success wasn't just the technology, it was the clear boundaries they set around what AI could and couldn't touch.

The hospital's approach offers a template worth stealing. They categorized tasks into three buckets: AI-first (drafts AI produces that get light human review before sending), AI-assisted (AI provides a starting point, but a human substantially rewrites), and AI-free (anything involving patient records, clinical recommendations, legal documents, or personalized care communications). That taxonomy didn't require anyone to understand machine learning. It just required leadership to think clearly about where accuracy was non-negotiable versus where speed mattered more than perfection. Every manager on your team can apply that same logic starting this week.

Click to enlarge

Comparing the Major AI Tools: What Each One Is Actually Good At

A Sales Director Who Stopped Writing Proposals From Scratch

Marcus runs a seven-person sales team at a mid-sized B2B software company. Before AI tools, his team spent three to five hours crafting each client proposal, pulling together company research, tailoring messaging, structuring the document, writing the executive summary. It was high-effort, repetitive work that pulled senior reps away from actual selling. Marcus started using Claude Pro to draft initial proposals. He'd paste in the prospect's LinkedIn profile, their company's recent press releases, and his team's notes from the discovery call. Then he'd ask Claude to produce a first draft structured around the client's stated priorities.

The result wasn't a finished proposal, it was a strong 70% draft that his team could refine in under an hour instead of building from zero. More importantly, the AI-generated drafts were often better structured than what his team produced under time pressure, because the model had absorbed the pattern of thousands of effective proposals. Marcus still reviews every proposal personally before it goes out. He checks the facts, adjusts the tone, and adds the specific relationship context that only a human who knows the client can provide. His team's proposal output increased by roughly 40% in the first quarter. The AI didn't replace his salespeople, it removed the part of the job they liked least.

Putting AI to Work in Your Actual Workflow

The professionals who get the most out of AI tools aren't the ones who use them most, they're the ones who use them most deliberately. They've identified three or four specific tasks in their week that AI handles well and built a small habit around those tasks. A marketing manager might use ChatGPT to draft social media copy every Monday morning. An HR director might use Claude to summarize candidate feedback after every interview round. A teacher might use Copilot to generate differentiated quiz questions from a lesson plan. These aren't dramatic workflow overhauls, they're precise insertions of AI into moments where it saves real time.

The quality of what you get back from AI tools depends heavily on what you put in. This is what practitioners mean when they talk about 'prompt engineering', but that phrase makes it sound more technical than it is. A better frame: think of it like briefing a smart contractor. The more specific context you give, your role, your audience, the purpose of the document, the tone you want, any constraints, the more useful the output. A vague request gets a generic response. A specific request, loaded with real context, gets something you can actually use. The difference between 'write me a project update email' and 'write a project update email to my CFO explaining a two-week delay due to vendor issues, keeping it under 150 words and avoiding blame language' is enormous.

There's also a compounding benefit that most people discover after a few weeks of regular AI use: you get better at knowing what AI can do. You develop intuition for which requests will land well and which will produce something you need to heavily edit. You learn to spot hallucinations faster. You start writing prompts that are more precise because you've seen what precision produces. This isn't a technical skill, it's professional judgment applied to a new tool, the same way you learned over time how to write an effective email subject line or run a meeting that doesn't waste everyone's time.

Click to enlarge

Key Principles from Part 2

1. AI hallucinations happen because models generate plausible-sounding text, not verified facts. Confidence in tone does not mean accuracy in content.
2. AI knowledge has a training cutoff. For current events, recent data, or live information, use tools with web access enabled (Copilot, ChatGPT with browsing) or verify manually.
3. Different AI tools have different strengths. Copilot and Gemini integrate into existing software; ChatGPT and Claude are more powerful for general tasks; Grammarly refines rather than creates.
4. Categorize your tasks: AI-first, AI-assisted, or AI-free. The category depends on how much accuracy risk is involved, not how complex the task is.
5. Treat AI output as a strong 70% draft. Your value-add is the context, judgment, relationships, and verification that AI cannot supply.
6. Better prompts produce better outputs. Give AI your role, audience, purpose, tone, and constraints, the same way you'd brief a capable colleague.
7. Regular use builds intuition. The more you use AI tools deliberately, the faster you develop judgment about when to trust, when to edit, and when to verify.

When AI Gets It Wrong, and What That Means for You

In 2023, a partner at a New York law firm submitted a legal brief citing six court cases to support his client's argument. The cases looked authoritative, proper citations, plausible summaries, convincing legal language. Every single one was fabricated. ChatGPT had generated fictional case names, fictional rulings, and fictional quotes with complete confidence. The partner had not verified them. The judge was not amused. The firm faced sanctions and the story became front-page news. This wasn't a story about a rogue AI. It was a story about a professional who misunderstood what AI actually does, and paid the price for it.

The lawyers involved weren't careless people. They were experienced professionals who assumed that because the AI's output looked authoritative, it was accurate. That assumption is the central trap. AI language models generate text by predicting what words should come next based on patterns learned from training data. They don't look things up. They don't check facts in real time. They produce fluent, confident-sounding language even when the underlying content is wrong. Understanding this one fact changes how every manager, marketer, and consultant should use these tools.

The principle here is not that AI is untrustworthy, it's that AI has a specific, predictable failure mode you can learn to manage. A confident tone is not a signal of accuracy. Length is not a signal of quality. Formatting is not a signal of correctness. Once you internalize this, you stop treating AI output as a final product and start treating it as a talented first draft from someone who occasionally makes things up without realizing it. That shift in mental model makes you a dramatically better user of these tools.

The Confidence Problem Across Industries

A marketing director at a mid-sized retail brand used Claude to research competitor pricing strategies for a board presentation. The output was detailed and well-structured, specific percentages, named campaigns, attributed quotes from executives. It looked like solid competitive intelligence. When her team fact-checked it ahead of the presentation, roughly 40 percent of the specific claims couldn't be verified, and two of the executive quotes didn't exist. She caught it in time. But she described the experience as 'deeply unsettling, it read like a real analyzt report.'

This pattern shows up across every professional context. An HR manager using AI to summarize employment law changes needs to verify those summaries against actual legal sources. A teacher using Gemini to generate historical facts for a lesson plan needs to cross-check dates and attributions. A sales consultant using Copilot to draft a proposal that references industry statistics needs to confirm those numbers came from a real study. The workflow is the same: use AI for speed and structure, then apply your professional judgment to verify what matters.

The good news is that AI's failure modes are predictable. Models are most reliable when generating structure, tone, and language, writing emails, reformatting content, brainstorming options, summarizing documents you provide. They are least reliable when asked to recall specific facts, recent events, statistics, or anything that requires external verification. Knowing which category your task falls into tells you exactly how much checking you need to do.

A Teacher Who Found the Right Balance

A high school history teacher in Chicago started using ChatGPT to build lesson plans and discussion questions. Early on, she made the same mistake as the lawyers, she used AI-generated historical details without checking them. A student caught an incorrect date in a worksheet. After that, she developed a simple personal rule: AI writes the structure and the questions, she verifies every specific fact. Her lesson planning time dropped from four hours a week to ninety minutes, and her materials actually improved because she had more time to add nuance.

Her workflow became a model for her department. Use AI to generate the skeleton, learning objectives, discussion prompts, activity formats, summary language. Then spend your expert time on what AI can't reliably do: verify accuracy, add context, apply judgment about what your specific students need. This division of labor is exactly how professionals in every field should think about AI. You bring the expertise. AI brings the speed.

Putting This Into Practice From Monday

The most effective professional AI users develop a two-step habit. First, they identify whether their task is structure-based or fact-based. Writing an email? Drafting a proposal structure? Brainstorming campaign names? AI is excellent here, use it freely. Researching a statistic, checking a regulation, citing a study? Treat AI output as a starting point and verify it through a credible source before it goes anywhere official.

Second, they give AI better inputs. The law firm partners gave ChatGPT a vague instruction and accepted whatever came back. The marketing director did the same. The teacher eventually learned to provide context, the curriculum standard, the student age group, the specific topic, and got dramatically better, more usable output. AI performs better when you treat it like a capable new colleague who needs a proper brief, not a search engine you can query with two words.

None of this requires technical skill. It requires a mental model shift: AI is a powerful drafting and thinking partner that has a specific weakness around factual recall. Build your workflows around that reality. Use AI for the parts it does well, apply your professional judgment to the parts it doesn't, and you get the speed benefits without the professional risk. That's the entire game.

Click to enlarge

Key Takeaways

1. AI language models predict plausible text, they do not retrieve verified facts. Fluent output does not mean accurate output.
2. Hallucinations are a structural feature of how these models work, not a bug that will disappear. Build verification into your workflow.
3. AI is most reliable for structure-based tasks: drafting, reformatting, summarising content you provide, and brainstorming.
4. AI is least reliable for specific facts, statistics, citations, legal references, and anything requiring current information.
5. Better inputs produce better outputs. Give AI context, constraints, and a clear brief, just like you would a capable human colleague.
6. The professional skill is not using AI, it's knowing which parts of your work to hand to AI and which parts require your expertise and judgment.
7. A 30-second verification habit before any AI-generated fact leaves your hands eliminates most professional risk.