Where AI Wins Right Now

What AI Can Do Well Today

What made this possible wasn't magic, it was a specific category of work that AI handles exceptionally well. Klarna's customers were asking questions that followed predictable patterns: where's my order, how do I return this, why was I charged twice. The answers required understanding natural language, retrieving accurate information, and responding clearly. They didn't require physical presence, emotional nuance at the level of a grief counselor, or creative judgment calls. The task was bounded, language-based, and high-volume. That combination is AI's sweet spot, and understanding it is the single most useful frame you can carry into any conversation about deploying AI in your organization.

Klarna's story also contains a warning that we'll return to later in this lesson. The company later walked back some of its automation, acknowledging that customer experience suffered in certain edge cases, complex disputes, emotionally charged situations, queries that fell outside the training distribution. The lesson isn't that AI failed. It's that AI succeeded within a specific envelope of tasks and showed its limits at the edges. Knowing where that envelope ends is just as valuable as knowing where it begins.

The principle Klarna's experience reveals is this: AI performs best when tasks are language-based, repeatable in structure, and don't require physical-world judgment or accountability for novel decisions. That sounds abstract, so spend the rest of this lesson watching it play out across industries and roles. Each example will add a layer to your understanding of what 'performing well' actually means, and what it silently excludes.

The Five Things AI Does Exceptionally Well Right Now

Morgan Stanley manages over $1.3 trillion in client assets. Its financial advisors spend a meaningful chunk of their working week doing something that has nothing to do with financial strategy: searching for information buried in thousands of internal research documents, compliance guidelines, and market reports. In 2023, the firm deployed a GPT-4-powered tool, built with OpenAI, that lets advisors ask questions in plain English and get synthesized answers pulled from that entire document library in seconds. An advisor preparing for a client meeting about emerging market bonds no longer reads three research reports. They ask the AI, review the output, and verify the key claims. The work that took 45 minutes now takes 8.

This is AI's first major strength: information retrieval and synthesis. Not just searching, but reading, connecting, and summarizing across large bodies of text. ChatGPT, Claude, and Perplexity can all do versions of this. Perplexity specifically is built around it, it searches the web in real time and returns synthesized answers with citations, rather than a list of links. The distinction matters because the bottleneck for most knowledge workers isn't finding information. It's processing it. A lawyer reviewing discovery documents, a consultant scanning competitor annual reports, a product manager reading user research transcripts, they all face the same problem: too much text, not enough time. AI breaks that bottleneck.

The second major strength is drafting and editing. This is where most professionals first encounter AI's usefulness, and it's broader than it looks. It's not just 'write me an email.' It's transforming a rough bullet-point brain dump into a structured proposal. It's taking a legal contract and producing a plain-English summary for a non-lawyer. It's rewriting a technically accurate but dense paragraph so a general audience can absorb it. Claude 3.5 Sonnet and GPT-4o are particularly strong here, they maintain coherence across long documents, follow specific style instructions, and adjust register (formal, casual, technical) on demand. GitHub Copilot does the same thing for code: it drafts, explains, and refactors based on your intent expressed in natural language.

Click to enlarge

A Different Industry, The Same Pattern

Move from finance to healthcare and the same pattern appears. Physicians at major hospital systems spend, on average, 37% of their working time on documentation, according to research published in the Annals of Internal Medicine. That's clinical notes, referral letters, prior authorization forms, discharge summaries. It's time not spent with patients. Ambient AI documentation tools like Nuance DAX (built on Microsoft Azure OpenAI) and Suki AI listen to physician-patient conversations and generate structured clinical notes automatically. Doctors review and approve the notes; the AI drafts them. Mass General Brigham reported that physicians using DAX reduced documentation time by 7 minutes per patient encounter, significant when you're seeing 20 patients a day.

Notice what the AI is doing here: it's listening to spoken language, converting it to text, then structuring that text into a standardized format with the correct medical terminology in the right fields. That's pattern completion across language, applied to a domain-specific template. The physician's clinical judgment, what to prescribe, what to diagnose, what risks to flag, remains entirely human. The AI handles the transcription and formatting layer. This distinction between 'generating the artifact' and 'making the judgment' runs through almost every successful AI deployment you'll examine. The artifact production is automated; the accountability stays human.

Click to enlarge

What AI Handles Well vs. Where It Struggles

Click to enlarge

The Marketing Team That Shipped 10x More Content

Jasper AI, a content platform built on top of large language models, published case studies showing marketing teams producing content at volumes that would have been impossible before. One mid-sized e-commerce brand reported going from publishing 4 blog posts per month to 40, using a workflow where marketers brief the AI on topic, audience, and key points, then edit and fact-check the output. The time per piece dropped from 6 hours to under 90 minutes. This isn't the AI writing autonomously and publishing unreviewed content. It's a new division of labor: the AI handles the first draft, the human handles strategy, accuracy, and final judgment.

The third major strength visible here is content scaling. AI doesn't get tired, doesn't have writer's block, and produces a competent first draft in seconds. For teams that need volume. SEO content, product descriptions, social media variations, email sequences, this is transformative. Notion AI, integrated directly into Notion workspaces, does this inside the tool where teams already work. You highlight a rough outline and press 'improve writing' or 'make this shorter' and the edit appears inline. The friction between thinking and producing is dramatically reduced. But the strategic question, what content to create, for whom, with what angle, remains a human decision. AI executes on briefs; it doesn't generate strategy from scratch reliably.

What This Means in Practice

The three examples above. Klarna's customer service, Morgan Stanley's research assistant, and the e-commerce content team, share a structural feature. In each case, AI is handling the high-volume, language-based production layer of work, while humans retain ownership of judgment, strategy, and accountability. This isn't a compromise or a limitation to apologize for. It's actually the correct design. The tasks that AI accelerates, drafting, summarizing, retrieving, formatting, are tasks that consumed professional time without requiring the expertise that makes professionals valuable. Freeing that time up is the point.

The practical implication is that the most immediate AI opportunity for most professionals isn't replacing their core expertise. It's eliminating the administrative and production overhead around that expertise. A consultant's value is in insight and client relationships, not in formatting PowerPoint slides or writing the first draft of a situation analyzis. A lawyer's value is in judgment and strategy, not in summarizing depositions or drafting the first version of a contract clause. A data analyzt's value is in interpretation, not in writing the boilerplate Python to load and clean a dataset, which GitHub Copilot or ChatGPT's code interpreter can produce in under a minute.

This reframing also explains why AI adoption tends to stick when professionals see it as an amplifier rather than a replacement. Klarna's initial framing, we replaced human agents, created organizational friction and eventually partial reversal. Morgan Stanley's framing, we gave advisors a research superpower, created adoption and satisfaction. The underlying technology is similar. The framing, the workflow design, and the human-AI division of labor are what determine whether deployment succeeds. The companies getting durable results from AI are the ones who asked 'what do our best people spend time on that AI could handle?' rather than 'what jobs can we cut?'

Key Principles From These Examples

1. AI excels at language-based, high-volume, structurally repeatable tasks, summarizing, drafting, retrieving, formatting, translating, and coding first drafts.
2. The highest-value AI deployments separate artifact production (AI's job) from judgment and accountability (the human's job), not one or the other, but both in the right place.
3. AI performance degrades at the edges of its training: novel situations, real-time data, physical-world tasks, and high-stakes decisions requiring genuine accountability.
4. Prompt quality drives output quality, vague inputs produce generic outputs; specific, context-rich inputs produce usable, professional-grade results.
5. Framing AI as an amplifier of existing expertise, rather than a replacement for it, produces better adoption outcomes and better results, as the contrast between Klarna's and Morgan Stanley's approaches illustrates.
6. The immediate opportunity for most professionals is eliminating the production overhead around their core expertise, not replacing the expertise itself.

What to Take Away From Part 1

• AI's current strengths cluster around five areas: information synthesis, drafting and editing, content scaling, code assistance, and structured question-answering from a knowledge base.
• Real deployments. Klarna, Morgan Stanley, Nuance DAX, show both what AI can do at scale and where its limits appear under production conditions.
• The task categories where AI performs well share a common structure: language-based, high-volume, pattern-amenable, with human review as a quality gate.
• AI does not perform well on tasks requiring real-time data (without tools), physical-world interaction, persistent memory across sessions, or novel ethical judgment.
• Your first move is identifying which parts of your own recurring work fall into the high-performance task categories, the task above gives you a method to do that immediately.

When AI Becomes a Research Department

In 2023, a mid-sized consulting firm in Chicago. Huron Consulting Group, quietly started routing its market research requests through AI before any human analyzt touched them. The problem they were solving wasn't about cutting headcount. It was about the crushing bottleneck between a client asking a question on Monday and getting a structured answer by Thursday. Junior analyzts were spending 60-70% of their time on information gathering: pulling earnings reports, summarizing competitor moves, building first-draft frameworks. The actual insight work, the part clients were paying for, was getting compressed into the last 20% of the timeline. AI, they found, could collapse that gathering phase from days to hours.

Their approach was methodical. They didn't replace the research process, they split it into stages and assigned each stage to whoever handled it best. Perplexity AI handled initial information triage: pulling recent sources, summarizing publicly available data, flagging contradictions across reports. ChatGPT-4 handled structural drafting: turning raw notes into client-ready frameworks, synthesizing findings into executive summaries, generating the first version of slide narratives. Human analyzts then did what only humans could do well, verify claims against proprietary data, apply relationship context, make judgment calls about what a specific client actually needed to hear. The output quality went up. Turnaround time dropped by roughly 40%.

The principle buried in that story isn't about speed. It's about where cognitive effort gets spent. AI doesn't make your team redundant, it forces a clearer conversation about which tasks actually require human judgment and which ones were consuming human energy simply because no better option existed. Every team has a version of this bottleneck. The question Part 1 left you with, what AI does well in language and reasoning, now gets a harder test: what happens when you apply that capability to real professional workflows with real constraints?

Click to enlarge

Pattern Recognition at Scale: What AI Actually Sees

To understand why AI performs so differently across tasks, you need a clearer picture of what it's actually doing when it processes your request. Models like GPT-4 and Claude 3 weren't taught rules. They were trained on enormous volumes of text. GPT-4's training data is estimated to include hundreds of billions of words from books, websites, code repositories, and academic papers, and through that exposure, they internalized statistical patterns about how language, ideas, and structures relate to each other. When you ask Claude to write a project brief, it isn't following a template. It's drawing on patterns from thousands of project briefs it encountered during training, weighting which elements tend to appear together and in what order.

This is why AI performs brilliantly at tasks that are fundamentally about pattern matching and pattern generation. Editing prose for clarity? That's pattern matching against high-quality writing. Generating five marketing headline options? Pattern generation from observed headline structures. Summarizing a 40-page report into three key themes? Pattern extraction from document structure. The moment a task requires information the model wasn't trained on, last week's sales data, your company's internal strategy, a niche regulatory update from last month, the pattern-matching engine has nothing reliable to work from, and quality degrades fast. This is the structural reason AI hallucinates: it keeps generating plausible-sounding patterns even when no accurate pattern exists in its training data.

Knowing this changes how you design prompts. When you give AI rich context, background documents, specific data, clear constraints, you're essentially giving the pattern-matching engine better raw material. A prompt that says 'Write a proposal for our SaaS product targeting mid-market HR teams with a $15K annual contract value' will outperform 'Write a sales proposal' by a wide margin, not because the second prompt is vague, but because the first one activates more specific and relevant patterns. The professionals who get the most out of AI consistently front-load their prompts with context that most people would consider obvious and skip.

Click to enlarge

AI in the Operating Room (Sort Of): Healthcare Administration

The healthcare sector offers one of the sharpest illustrations of where AI's strengths land and where its limits become non-negotiable. Nuance Communications, now part of Microsoft, built an AI tool called DAX Copilot that listens to patient-physician conversations and automatically generates clinical documentation: visit notes, referral letters, after-visit summaries. As of 2024, it's deployed across hundreds of health systems in the US, and physicians using it report saving an average of three hours per day on documentation. That's three hours previously spent on structured writing tasks, describing symptoms, logging observations, drafting follow-up instructions, that are precisely the kind of pattern-rich language generation AI handles well.

Notice what DAX Copilot doesn't do. It doesn't diagnose. It doesn't recommend treatment. It doesn't decide which symptoms are significant and which are incidental. Those judgment calls stay entirely with the physician. The AI's job is to listen to a conversation that already contains the clinical reasoning and translate it into structured documentation format, a task that's high-effort for humans but structurally well-suited to AI. The business case is compelling: physician burnout is significantly driven by administrative load, and documentation is the biggest single contributor. AI attacking that specific slice creates measurable value without touching the clinical decisions that require years of trained human judgment.

What AI Does Well vs. What It Struggles With

The Marketing Team That Stopped Staring at Blank Pages

Zapier's marketing team published a detailed account of how they restructured content production using AI in 2023. Their challenge was classic: a small team, a large content calendar, and the perpetual tyranny of the blank page. Their solution was to use ChatGPT not as a content factory, but as a drafting partner operating at the brief level. A human strategist would define the target keyword, the reader's specific problem, the angle that differentiated their take, and three supporting arguments. That brief went into a structured prompt. ChatGPT produced a full working draft, typically 800-1200 words, in under two minutes. Human editors then spent 30-45 minutes reshaping voice, verifying claims, adding original examples, and adjusting the argument. Total time per piece dropped from an average of 6 hours to under 2.

The quality insight from Zapier's experiment is counterintuitive: the AI draft wasn't published because it was good enough. It was used because starting from a structured draft, even an imperfect one, is cognitively easier than starting from nothing. Writers know this phenomenon. The first draft's job isn't to be right; it's to be something you can react to. AI is extraordinarily good at producing that something. The Zapier team also discovered that AI drafts had a predictable failure mode: they were structurally competent but tonally generic. The human editing pass wasn't about fixing errors, it was about injecting the specific voice, specific examples, and specific opinion that made content worth reading. AI handled the skeleton. Humans added the personality.

What This Means for How You Work

The three examples above, consulting research, clinical documentation, content production, share a structural pattern worth naming explicitly. In each case, AI performed best when it was inserted into an existing workflow at a specific, bounded stage rather than handed an entire job end-to-end. Huron's analyzts didn't ask AI to do research; they asked it to handle information gathering so humans could focus on interpretation. Nuance's DAX didn't attempt diagnosis; it handled transcription and structuring of a conversation that had already happened. Zapier's editors didn't ask AI to produce final content; they used it to eliminate the blank-page problem so editors could focus on voice and accuracy. The pattern is consistent: narrow the AI's role to the stage it handles best.

This has a practical implication for how you evaluate AI tools for your own work. The right question isn't 'Can AI do my job?', it's 'Which specific stages of my job consume time and energy disproportionate to the judgment they require?' Documentation, first drafts, summaries, research triage, formatting, template population, brainstorming lists, these are the stages where AI creates the most immediate value with the least risk. Strategic decisions, stakeholder relationship management, ethical judgment calls, creative direction, and anything requiring current proprietary data, these stay human. The professionals who benefit most from AI aren't those who hand it the most work. They're those who are most precise about which work to hand it.

There's also a capability curve to account for. The AI tools available today. GPT-4, Claude 3 Opus, Gemini 1.5 Pro, are meaningfully more capable than their predecessors from 18 months ago, and the trajectory hasn't flattened. Tasks that AI handles moderately well today (nuanced analyzis, multi-step reasoning, longer document coherence) are likely to move into the 'handles well' column within a product cycle or two. Building your AI intuition now, understanding which tasks to assign, how to prompt effectively, how to verify outputs, gives you a durable skill even as the specific capabilities shift. The mental model matters more than any particular tool.

Click to enlarge

Principles Extracted From These Stories

1. AI creates the most value when inserted at a specific workflow stage, not handed an entire job, the tighter the scope, the higher the quality.
2. Pattern-rich tasks (drafting, summarizing, structuring, reformatting) are where current AI models consistently outperform time-cost expectations.
3. The 70% rule applies broadly: AI produces a strong first pass, but the remaining 30%, accuracy verification, voice, judgment, requires human attention every time.
4. Prompt quality is the primary variable under your control, context-rich prompts consistently outperform vague ones because they activate more relevant training patterns.
5. AI's weaknesses are structural, not random: real-time information gaps, numerical unreliability, and hallucination under uncertainty are predictable failure modes you can design around.
6. The cognitive value of AI drafts isn't always their quality, it's that reacting to something is easier than creating from nothing, which changes how creative and analytical work flows.
7. Building AI intuition now, knowing what to assign, how to prompt, how to verify, is a durable professional skill regardless of which specific tools dominate next year.

Key Takeaways From This Section

• AI models like GPT-4 and Claude work through statistical pattern matching, understanding this explains both their strengths and their failure modes.
• Real enterprise deployments (Huron, Nuance, Zapier) show AI performing best at bounded, high-volume, pattern-rich stages of professional workflows.
• The comparison table gives you a practical reference for which task types to trust AI with, which to use cautiously, and which to keep fully human.
• Context in your prompt isn't optional, it's the primary mechanism for improving AI output quality.
• The right framing isn't 'what can AI do' but 'which stages of my specific work consume effort disproportionate to the judgment they require.'

When AI Meets the Messy Real World

In 2023, Klarna deployed an AI assistant built on OpenAI's technology to handle customer service conversations. Within a month, it was managing two-thirds of all customer chats, work that previously required 700 human agents. Average resolution time dropped from 11 minutes to under 2 minutes. Klarna's CEO publicly credited the system with contributing to a $40 million annual profit improvement. On paper, this looks like a straightforward AI success story. But the details matter: the AI handled repetitive, well-defined queries about refunds, order status, and payment plans. It did not handle fraud disputes, complex complaints, or emotionally charged situations. Those still went to humans.

That boundary is the whole lesson. Klarna did not replace customer service, it replaced the boring, predictable 70% of customer service, freeing human agents for work requiring judgment, empathy, and authority. The AI excelled because the task had clear inputs, known rules, and measurable right answers. A customer either got a refund or didn't. An order either shipped or it didn't. The moment a query drifted into ambiguity, "I've been a customer for eight years and this is unacceptable", a human took over. Understanding that line is what separates professionals who deploy AI effectively from those who either over-trust or under-use it.

Pattern Recognition at Scale: What That Actually Looks Like

PathAI, a Boston-based medical technology company, trained deep learning models to analyze pathology slides, the tissue samples doctors examine to diagnose cancer. Their system processes thousands of cellular patterns per slide in seconds, flagging regions a pathologist should examine closely. In clinical studies, the AI-assisted pathologists caught more early-stage cancers than unassisted ones. The AI's strength is raw pattern recognition at a scale no human can match: it has effectively "seen" millions of slides and learned which cellular arrangements correlate with malignancy. It does not diagnose. A licensed pathologist still makes every clinical decision. The AI is a tireless, fast first-pass filter.

This example illustrates AI's deepest strength: finding signal in high-volume, high-complexity data that would overwhelm human attention. The same capability powers fraud detection at Visa (analyzing billions of transactions for anomalous patterns), content moderation at YouTube (flagging millions of videos before human reviewers see them), and predictive maintenance at Siemens (monitoring sensor data from industrial equipment to anticipate failures before they happen). In every case, the AI processes volume and consistency that humans cannot. In every case, a human makes or reviews the consequential decision.

What connects Klarna's chatbot and PathAI's slide analyzer is the nature of the underlying task: both involve matching new inputs against vast libraries of known patterns. ChatGPT matches your prompt against patterns in billions of text documents. Midjourney matches your description against patterns across hundreds of millions of images. GitHub Copilot matches your code context against patterns in public repositories. The technology is the same family, the applications look completely different. Once you see this, you stop being surprised by what AI can do and start predicting it.

The Consultant Who Stopped Dreading Research

A senior strategy consultant at a mid-sized firm described her pre-AI workflow for new client engagements: two to three days of background research, industry reading, competitor mapping, and framework selection before she could write a single slide. She started using Perplexity AI for initial research sweeps and Claude for synthesizing documents and drafting frameworks. That prep time dropped to four hours. She is careful to verify every factual claim Perplexity surfaces. AI hallucination is real, and a wrong statistic in a client deck is a credibility disaster. But the synthesis, the structuring, the first-draft thinking? She now calls that "the part AI owns."

Her experience maps precisely onto AI's documented strengths: summarization, pattern extraction from documents, structured drafting, and reformatting information for different audiences. These tasks share a common trait, they involve reorganizing existing knowledge rather than generating genuinely new insight. AI is extraordinarily good at the former. The novel strategic insight that wins the client? That still comes from her, built on top of the research foundation the AI laid in four hours instead of three days. She didn't get replaced. She got 16 hours back every week.

What This Means for How You Work

The professionals getting the most from AI right now share one habit: they treat AI as a fast, tireless first-draft machine and reserve their own time for judgment, verification, and decisions that carry real consequences. A marketer at HubSpot described it as "AI does the scaffolding, I do the architecture." She uses ChatGPT to generate ten subject line variations, then picks and refines the best one. The AI's ten options take 15 seconds. Her judgment about which one fits the brand, the audience, and the moment, that takes two minutes and is the part that actually matters. The total time is still a fraction of writing from scratch.

There is a failure mode worth naming: over-delegation. Professionals who send AI-generated content straight to clients without review, who trust AI-sourced statistics without checking, or who use AI for tasks requiring genuine human accountability are accumulating invisible risk. Claude and ChatGPT both hallucinate, they generate plausible-sounding falsehoods with full confidence. Perplexity cites sources, which helps, but sources can be misread or misrepresented. The professional who understands this treats AI output the way a good editor treats a first draft: valuable raw material, not a finished product.

The practical implication is a simple mental model: before using AI for any task, ask two questions. First, does this task involve matching patterns in existing information, or does it require genuinely novel judgment? Second, what is the cost of an AI error here, low (a draft email), medium (a client report), or high (a legal filing, a medical decision, a financial commitment)? High-pattern, low-stakes tasks are AI's home territory. Low-pattern, high-stakes tasks belong to humans. The large middle ground, high-pattern, medium-stakes, is where the most professional leverage lives, with human review as the safety valve.

1. AI excels at pattern recognition and synthesis, tasks where the answer lives in existing data, text, or images at a scale humans can't match.
2. The 70% rule holds across industries: AI reliably handles the routine majority; humans handle the edge cases that require judgment, empathy, or accountability.
3. Real deployments (Klarna, PathAI, GitHub Copilot) succeed by defining the boundary between AI and human roles clearly, not by eliminating that boundary.
4. Hallucination is a structural property of current AI, not a bug to be fixed soon. Verification by a human is non-negotiable for any medium-to-high-stakes output.
5. The professionals gaining the most are using AI as a first-draft and research accelerator, then applying their own judgment to refine and approve outputs.
6. Pattern-based + low error cost = AI's home territory. Novel judgment + high error cost = human territory. The profitable middle ground requires both.

• AI's core capability is pattern matching at scale, this explains both what it can do and where it fails.
• ChatGPT, Claude, Perplexity, GitHub Copilot, and Midjourney all run on the same fundamental mechanism applied to different domains.
• Klarna's 700-agent-equivalent AI and PathAI's cancer-detection system both work because the tasks are high-volume, well-defined, and pattern-rich.
• Always ask before delegating to AI: how pattern-based is this task, and what happens if the AI is wrong?
• Your competitive advantage isn't avoiding AI, it's knowing exactly where to apply it and where to stay in the loop yourself.