Volume 32: Data Is the Work

ServiceNow lost 18% of its value in a single trading session last month. Workday slid 9%. Salesforce, Adobe, Oracle, and HubSpot all dropped 6 to 9% in the same window. The headlines blamed AI fears. The deeper story is a moat reset, and what just hit software is coming for everything else.

🧭 Founder's Corner: Why the SaaS reset is a preview, not an outlier, and the one question to ask before any AI workflow that decides whether your investment will pay off.

🧠 AI Education: What Copilot can actually do in Word, Excel, and PowerPoint, where it falls short, and the one mode shift most users have not noticed yet.

✅ 10-Minute Win: Snap a photo of your fridge and turn it into three real recipes, a cook timeline, and a shopping list someone else can act on.

Let's get into it.

1) An AI Model Beat Doctors at Diagnosing Patients, in a New Study

Summary: A peer-reviewed study published in Science on April 30 found that an OpenAI reasoning model outperformed two experienced physicians at diagnosing patients across the messy, real-world data of an emergency department. Researchers from Harvard Medical School and Beth Israel Deaconess Medical Center evaluated the model at three points in care, from triage to admission, and found it consistently produced more accurate diagnoses and management decisions than the physicians using the same information. 

Why it matters: This is one of the first rigorous head-to-head studies showing an AI model performing better than experienced doctors in a real ER setting, not on textbook cases. The authors are explicit that this is not about replacing physicians, but about what gets possible when AI is treated as a second opinion that catches what a tired clinician might miss. Expect this study to come up in every healthcare AI conversation for the rest of 2026.

2) Copilot in Outlook: New Agentic Experiences for Email and Calendar

Summary: Microsoft announced on April 27 that Copilot in Outlook is shifting from a sidebar that helps with the task in front of you to an agent that runs your inbox and calendar continuously. The new capabilities triage emails, draft follow-ups for messages that have not received replies, set inbox rules, resolve calendar conflicts by rescheduling 1:1s, block focus time, and prepare meeting agendas. The features are rolling out through the Microsoft 365 Copilot Frontier program starting April 27. 

Why it matters: This is the part of agentic AI that finally touches a daily reality most professionals share: an overflowing inbox and a fragmented calendar. If your day is built on Outlook, this is the first version of Copilot that promises to handle the work between meetings, not just inside them. The trade-off worth thinking about is governance: an AI that reschedules your meetings and drafts follow-ups on your behalf is also one that needs clear guardrails on tone, approvals, and what it can do without you in the loop.

3) US Department of Labor Launches Website to Build AI Skills, Expand AI-Focused Registered Apprenticeship Programs

Summary: The U.S. Department of Labor announced on April 29 the launch of an AI in Registered Apprenticeship Innovation Portal, a centralized resource designed to help organizations build AI literacy and create AI-focused apprenticeship programs. The portal builds on the department's AI Literacy Framework released earlier this year and organizes training resources around three areas: AI skills and literacy, industry-specific skill building (including healthcare and finance), and three integration paths for new or existing apprenticeship programs. 

Why it matters: This is one of the clearest federal signals yet that AI fluency is being treated as workforce infrastructure, not optional upskilling. If you lead a team, manage hiring, or run any kind of training program, the resources are now publicly available and government-backed. For non-technical professionals, it also reframes the conversation: AI literacy is becoming a credentialed, structured skill that employers can recognize, not just a personal hobby project.

4) OpenAI, Nvidia, Alphabet and More Sign AI Deal With Pentagon for Classified Military Use

Summary: The Department of Defense announced on May 1 that it has signed agreements with seven leading AI and cloud companies, including OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, SpaceX, and the startup Reflection, to deploy their AI tools on classified military networks, with Oracle added later that day. The deals embed frontier AI into day-to-day Pentagon operations and require an "all lawful use" provision; Anthropic was notably absent and is currently challenging a separate Pentagon designation as a supply chain risk in court. 

Why it matters: AI is now national infrastructure, not just productivity software. The choices these labs make about what their models will and will not do, and how those choices line up with U.S. defense requirements, are now driving real commercial consequences. For leaders in regulated industries like healthcare and finance, this is also a preview of how government procurement is starting to shape which AI tools your organization will be allowed to use, even far outside defense.

5) OpenAI Models, Codex, and Managed Agents Come to AWS

Summary: OpenAI and AWS announced on April 28 that OpenAI's frontier models, including GPT-5.5, its Codex coding agent, and a new Amazon Bedrock Managed Agents service powered by OpenAI are now available in limited preview on Amazon Bedrock. The launch came one day after OpenAI restructured its Microsoft partnership to remove cloud exclusivity, and brings OpenAI alongside Anthropic, Meta, Mistral, and Cohere on AWS's enterprise AI platform.

Why it matters: For years, picking OpenAI meant picking Microsoft Azure. That assumption is gone. If your enterprise runs on AWS, you can now use the same OpenAI models inside the security and compliance environment your team already trusts. The bigger signal is that the AI lab ecosystem and the cloud ecosystem are decoupling, which means model choice and infrastructure choice will increasingly be separate procurement decisions for IT and security teams.

The competitive landscape has shifted, and the prevailing narrative missed the mark. On April 23, 2026, ServiceNow lost 18% of its value in a single trading session. Worst day in the company's history. The cascade hit fast: Workday slid 9%, and Salesforce, HubSpot, Adobe, Intuit, and Oracle all dropped 6 to 9% in the same window. The headlines blamed AI fears. The deeper story is a moat reset, and the first sign just arrived. 

The Warning Came for SaaS First

Per-seat pricing helped build the modern software industry. AI is now breaking that model, and the market is adjusting in real time. 

What used to take years is now happening in a matter of weeks. Press reports estimated roughly $285 billion in software market capitalization vanished in a 48-hour window in February 2026, after Anthropic launched Claude Cowork. Traders at Jefferies named the event the "SaaSpocalypse" before the market closed. Workday disclosed a 2% workforce reduction in February 2026, in the same week the SaaSpocalypse hit. The company recorded a $135 million restructuring charge in the same fiscal quarter. Even Atlassian, which beat Q3 earnings on April 30, 2026, traded 87% below its 2021 peak earlier in the month.

One company is already trying to get ahead of the shift. On its Q4 earnings call, Salesforce introduced Agentic Work Units, a new metric tracking AI-completed tasks rather than seats. That was a quiet but important signal: per-seat pricing has no long-term future. IDC projects that by 2028, 70% of software vendors will move away from per-seat licensing. 

That matters because one of the strongest moats in enterprise software just got reset in roughly twelve weeks. Per-seat subscriptions, switching costs, network effects, all of it looked durable. But if a moat that took twenty years to build can be eroded that quickly, no industry should assume it is protected. 

What Benioff and Huang Agree On 

This moment forces an uncomfortable question: what were these moats really made of? 

Two of the most powerful leaders in tech have publicly disagreed on whether SaaS is dying, and that disagreement helps clarify the answer. 

Marc Benioff used "SaaSpocalypse" at least six times on Salesforce's Q4 earnings call in late February 2026, per TechCrunch. His view was simple. The deeper your enterprise data is embedded in a system of record, the harder it is to rip out. 

Jensen Huang argued the opposite on CNBC that same month, saying the markets "got it wrong." His point was that agents will not replace the tools. Instead, they will use the tools.

But underneath the disagreement, they are saying the same thing. The real defensibility is not in the interface, the pricing model, or even the workflow. It is in the data and the systems of record underneath it. Everything else is negotiable. The data layer is not. 

This Is Not a Software Story

What just happened in software is going to happen everywhere else, often before the headlines name it. AI-driven customer service is already eroding service advantages that took years to build. Content libraries and domain authority, once the bedrock of marketing and SEO, are losing ground to AI that produces qualified content at near-zero cost. Product velocity gaps that once took a decade to open are now closing in months as AI-native companies catch up. 

Klarna is one of the clearest public examples. The company moved aggressively in 2024 to replace customer service with AI, then pulled parts of that strategy back in 2025 when service quality dropped. CEO Sebastian Siemiatkowski told Bloomberg that cost had become "too predominant" a factor, producing lower-quality service. 

The lesson is not that AI failed. The lesson is that the workflow was not ready to be replaced all at once. The companies that come out ahead will know the difference between a workflow AI can handle today and one that still needs rebuilding before AI can perform it well. 

The Moat Most Companies Skip

Most companies are not behind on AI investment. They are behind on the data work that makes AI investment pay off.

McKinsey's recent productivity paradox analysis, citing their late-2025 State of AI survey, found that nearly nine out of ten companies had deployed AI in at least one business function by the end of 2025, but 94% reported no significant value from those investments. Gartner projects that 60% of AI initiatives unsupported by AI-ready data will be abandoned through 2026. The bottleneck is not the technology. It is the data underneath the technology. These were not bad ideas. They were good ideas built on weak foundations. 

Legacy workflows are often running on infrastructure that cannot support real redesign. Most companies are constrained by the dataset they already have. I have made that mistake myself, moving slower than I should have and trying to force creative solutions into outdated systems.

One project has stayed with me in particular. My team and I were working on a healthcare affordability workflow. For us, patient affordability was never abstract. It is a real problem with real consequences. People do not fill prescriptions they cannot afford, and that reality has a way of sharpening your focus. 

We saw an opportunity to redesign the workflow. There was a vendor platform involved, API connections, data requirements, and operational changes. But somewhere in the middle of the project, it became evident that we were not building a workflow. We were identifying the most important pieces of data and using them to power one. The vendor platform mattered. The API mattered. But those were not the real story. The real story was what happened once clean data started going in and clean data started coming out. 

Everything accelerated. Operational teams had access to the data earlier. The workflow shifted from reactive to proactive. An internal automation tool worked better, not because the automation had changed, but because the data was finally ready to flow through it. All possible through data, not magic. 

That has shaped every AI workflow I have built since. Clean data does not just unlock new AI capabilities. It increases the value of investments a company has already made. That is the compounding effect many boardrooms still miss. 

And it points to a simple rule.

Workflow redesign should start with understanding how the data moves today. Not how the documentation says it moves. Not how the system was originally designed. How it actually moves, with all the friction included. Receipts before claims.

Healthcare Is Sitting on a Moat It Has Not Built Yet

Healthcare may be the clearest example of this pattern anywhere. Pharma, payers, pharmacies, and health systems sit on decades of regulated, proprietary data, the kind that new AI-native entrants cannot replicate or buy on the open market. The asset is real, but the readiness is not.

McKinsey's April 2026 healthcare survey found that half of US healthcare organizations have implemented gen AI, up from 25% two years ago. The barrier is no longer adoption. It is execution. The same survey ranked difficulty integrating tools into existing workflows as the top scaling roadblock, cited by 59% of healthcare leaders. The data readiness divide we have been tracking across industries is the same divide separating healthcare AI experiments from healthcare AI value.

In an earlier piece, I argued that the moat in healthcare is not the proprietary silo. It is the ability to connect to the broader ecosystem. Internal readiness is the prerequisite for external participation. Get the order wrong and you build connections to nothing. 

The Move on Monday Morning

You do not need to be a data architect to respond to this moment. You need to ask one question before implementing any AI workflow: How does the data move today?

That question works across industries. It works in regulated environments and unregulated ones. It works for technical teams and non-technical leaders alike.

The people who ask that question first will execute faster. The companies that ask it first will still be standing when this moat reset reaches industries that still believe they are insulated.

The reset has already started in software. It is coming for everything else.

The work is already sitting on your desk.

The companies that do it now will define the next decade in their industries. The ones that wait will be left explaining themselves on the next earnings call.

The data work is the work. Everything else is the outcome.

The Assumption

Most professionals walk into Word, Excel, and PowerPoint with a reasonable expectation. If Copilot can draft a meeting summary in Outlook from a thirty-minute call (Vol 31), it should be able to produce a polished compliance memo, build a working budget pivot, or generate a board-ready deck from a single prompt. The marketing demos reinforce this. So does the speed of Copilot in apps like Outlook and Teams, where the output often lands close to ready. Reasonable people hear "AI assistant in your apps" and conclude the apps where the heaviest production work happens should benefit the most.

Where It Breaks Down

Dana, our clinical operations coordinator, tests that assumption on a Wednesday afternoon. She has a quarterly compliance review presentation due Friday. She opens PowerPoint, pulls up the Copilot pane, and types: "Create a 12-slide presentation summarizing this quarter's compliance findings, the three open audit items, and the remediation plan." The output is structured but generic. The slides have her topic but not her data. The visuals are stock. The sequencing is off.

She moves to Excel to build a pivot showing audit completion rates by clinic site. The Copilot button is greyed out. Her file is on her desktop, not OneDrive. After moving it and toggling AutoSave on, Copilot wakes up but produces a pivot grouped on the wrong column because her data is not formatted as a structured table. She fixes that. The pivot now works, but the formatting still does not match her director's preferred layout.

In Word, the experience is closer to what she expected. Copilot drafts a strong first version of the executive summary in under a minute. But when she asks it to revise three specific paragraphs, the changes spill into surrounding sections she did not want touched.

What Is Actually Happening

The three production apps are not one product. They are three different surfaces with three different design philosophies, and Copilot behaves differently in each.

In Word, Copilot is closest to what most users imagine. It drafts long-form content, restructures sections, applies tone, and handles rewrites. Word was the first of the three apps to get full agentic editing in November 2025, which means Copilot can make multi-step, app-native changes directly inside the document while you watch. It still has trouble with surgical edits that do not bleed into adjacent paragraphs, which is why review before accepting changes is non-negotiable.

In Excel, Copilot is gated by infrastructure. The file must be saved to OneDrive or SharePoint with AutoSave on. Files on local drives produce a greyed-out button. Data must be formatted as a structured table (Ctrl+T) with clear column headers, no merged cells, and no blank rows. If you meet those conditions, Copilot generates formulas, builds pivot tables, surfaces patterns, and explains existing logic. If you do not, the output is unreliable or the tool refuses to engage. Excel also cannot write or run VBA macros.

In PowerPoint, Copilot is best understood as a scaffolding tool. It produces a structured first draft from a prompt or from a Word document, but it has historically struggled with iterative refinement, has a limit on how much context you can feed in a single prompt, and is weak on transitions, animations, and complex slide elements like SmartArt and embedded tables. To honor your company's branding, your IT team has to publish approved templates to the SharePoint Organization Asset Library.

The behavior that ties all three together is Edit with Copilot (formerly called Agent Mode), which Microsoft moved to general availability across Word, Excel, and PowerPoint on April 22, 2026. This is the part most users have not yet noticed. Edit with Copilot shifts the interaction from "ask once, accept the output" to "watch Copilot work in your file iteratively, see what it changed, and refine." It is the most important behavioral change in this series so far.

Five days later, Microsoft previewed agentic Copilot experiences in Outlook for email triage, follow-ups, and calendar management through the Frontier program. The direction is unmistakable: Copilot is becoming a coordinated agent that does ongoing work across your apps, not a single-prompt assistant you trigger and dismiss.

The Revised Mental Model

Copilot in the production apps is a draft accelerator, not a finished-output generator. Use it to skip the blank page. Bring your judgment to the last thirty to forty percent.

Three behavior changes follow from this. First, treat the first output as a starting position, not an answer. Plan to spend real time refining it, especially in PowerPoint where the gap between draft and presentable is the widest. Second, prepare your inputs before you prompt. In Excel, this means formatting your data as a table and saving to OneDrive before you click Copilot. In PowerPoint, this means starting from a Word document with strong structure (proper headings, clear sections) when the deck has to convey something specific. Third, use Edit with Copilot for any task that requires more than one round of changes. The iterative mode preserves your file, shows you what changed, and lets you steer the work in a way the single-prompt mode cannot.

The shortest version of this: Copilot drafts. You finish.

What to Watch For

• A polished prompt does not produce a polished slide deck. PowerPoint output almost always needs human editing for sequencing, visuals, and brand consistency.
• An Excel file on your local drive will not work with Copilot. If the button is greyed out, the file is the reason.
• Pivot table results are only as good as the table structure underneath them. Time spent cleaning data before prompting saves more time than Copilot saves on the analysis itself.
• Word edits can affect more than the section you targeted. Read surrounding paragraphs after every revision before accepting changes.
• Edit with Copilot is the right mode for multi-step work in any of the three apps. Single-prompt mode is fine for short tasks. For anything substantive, switch modes.

How This Connects

Vol 29 introduced the Copilot ecosystem and the grounding behavior that makes it useful. Vol 30 explained the architecture underneath: the orchestrator, Microsoft Graph, the semantic index, and how permissions shape every response. Vol 31 showed the high-leverage applications in Outlook and Teams. This volume covers the apps where the gap between expectation and output is largest, and where understanding the limits matters most.

Vol 33 moves into the agentic layer that now runs across the entire suite. Researcher, Analyst, Edit with Copilot, and Copilot Studio all sit on top of the foundation we have just built. That is where the AI Agents series (Vol 24-27) becomes a live capability inside an enterprise product, not a concept on a roadmap.

Part 4 of 6 in the Copilot Deep Dive series.

The Recipe Remix

Vision-based AI just unlocked one of the most useful kitchen moves you can make. Snap one photo of your fridge or pantry, and get three real recipes built around exactly what is on your shelves, plus a cook timeline and a short shopping list for whatever you are missing. No manual ingredient entry. No more guessing what to make with what you already own.

The Workflow

1. Snap the Photo (1 Minute)

Open your fridge or pantry, pull anything tucked to the back forward, and take one clear, well-lit photo. Capture as much of your produce, proteins, condiments, and pantry staples as you can fit in frame. If you have a lot, take two photos and upload both.

2. Get Three Recipe Options (3 Minutes)

Open ChatGPT (chatgpt.com) or Gemini (gemini.google.com) on your phone or laptop. Both free tiers support image upload. Upload your photo and paste the prompt below.

Copy/Paste Prompt: "You are a creative home cook who specializes in turning random ingredients into real meals. I am attaching a photo of what I currently have in my fridge and pantry.

Identify the ingredients you can see in the photo. Then give me three different dinner recipes I can make using mostly what I have. For each recipe:

1. Name the dish and the cuisine style.
2. List the ingredients I already have, then list any 1 to 3 items I would need to buy.
3. Give me cook time and a difficulty level (easy, medium, harder).

Pick three recipes that are different from each other in cuisine and cooking method. Aim for things I could actually make on a weeknight, not chef-level showpieces."

3. Pick One and Get the Cook Timeline (3 Minutes)

Read all three options. Pick the one that fits your time and energy tonight. Send the follow-up prompt below to turn it into a step-by-step cook timeline that handles parallel tasks (chopping while something else simmers).

Copy/Paste Prompt: "I want to make recipe [insert recipe name] tonight. Give me a complete cook timeline in the order I should do things, with timing for each step and what I should be doing in parallel. Format it as a numbered list with clock-style time markers like 0:00, 0:05, 0:15. Include prep, cooking, and final assembly. Aim for 30 minutes or less from start to plated. Call out any step where I might burn or overcook something if I walk away."

4. Generate the Missing-Items List and Send It (3 Minutes)

Now lock the missing groceries into something you can actually shop with.

Copy/Paste Prompt: "Based on the recipe above, write a clean shopping list of only the items I need to buy. Format it in three groups: produce, proteins or dairy, and pantry. Add quantities. Then write a one-line text I can send my partner or roommate that says what we are having for dinner and what to grab on the way home. Casual tone, under 25 words."

Copy the shopping list to your phone Notes app or paste it into a text to whoever does the grocery run. Dinner is decided, the cook plan is on your screen, and the missing items are in someone's hand.

The Payoff

You just turned a vague "I have stuff but no plan" into a chosen recipe, a minute-by-minute cook timeline, and a shopping list someone can actually act on. More importantly, you used your fridge as the input instead of a recipe search bar, which is the single biggest unlock vision-capable AI offers personal cooks.

🧠 The AI Concept You Just Used

Multimodal input. When you upload a photo and ask the model to reason from what it sees, you are using vision capability, one of the most underrated features of modern free-tier AI. The model is not just reading text. It is identifying objects, interpreting context, and feeding that interpretation into the same reasoning engine that handles your written prompts.

Transparency & Notes

• Tools used: ChatGPT (chatgpt.com) free tier with vision, or Gemini (gemini.google.com) free tier with vision. Both handle image upload. Microsoft Copilot also runs this workflow if you have been following our Deep Dive series.
• Privacy: Crop or angle your photo to keep identifying details out of frame (medication labels, mail on the counter, family photos on the fridge). The model only needs to see the food.