Volume 34: The Question You Have Not Asked Yet

I have built two AI agents that save me hours every week. They work, they solve real problems, yet they are still the consolation prize. The question I should have been asking the whole time is the same one most companies are quietly avoiding right now.

🧭 Founder's Corner: Why time-saved is the wrong success metric for any AI pilot, and the one question that separates the 34% of companies reimagining their business from the 37% changing nothing.

🧠 AI Education: A four-day audit a clinical operations coordinator ran on her own Copilot usage, and the single question that turns AI usage into AI value.

✅ 10-Minute Win: Replace two hours of tab-hopping with a sourced, pressure-tested one-page brief that ends in a specific next step.

Let's dive in.

1) Most U.S. Doctors Are Quietly Using This AI Tool. Few Patients Know About It.

Summary: NBC reported on May 13 that OpenEvidence, a clinician-only AI tool, was used by 65% of U.S. doctors across nearly 27 million encounters in April. The company raised $250 million at a $12 billion valuation and is now embedded in Mount Sinai's Epic EHR. 

Why it matters: The AI doctors actually use looks nothing like ChatGPT. The gap between what clinicians rely on and what patients know about is the next regulatory pressure point.

2) Medicare's New Payment Model Is Built for AI, and Most of the Tech World Has No Idea

Summary: TechCrunch reported on May 12 that CMS's new ACCESS model, going live July 5, is the first federal payment mechanism that reimburses AI agents for monitoring patients between visits and coordinating care. CMS selected 150 participants including Pair Team, Doctronic, Whoop, and Aledade. 

Why it matters: Medicare just opened a $900 billion reimbursement lane for AI care, and private insurers typically follow within 18 to 24 months. This turns "AI agent" from buzzword into billable service.

3) Google Confirms Gemini Intelligence for Android, Unveils Googlebooks and Aluminium OS

Summary: Google pre-released The Android Show on May 12 ahead of I/O 2026, confirming Gemini Intelligence as an agentic AI layer for Android with Chrome auto-browse, smart form-filling, and context-aware Android Auto. The company also announced Googlebooks, a premium laptop category running Aluminium OS, shipping from Acer, ASUS, Dell, HP, and Lenovo this fall. 

Why it matters: The phone is becoming an agent, not just a device. AI fluency moves from "how do I use ChatGPT" to "what is my phone doing on my behalf right now," which is a much harder governance and privacy question for IT teams.

4) The More Operational AI Becomes, the Bigger the Security Challenge

Summary: AI Business reported on May 15 that AI is becoming both a cybersecurity tool and a cybersecurity threat as autonomous, interconnected systems create new attack surfaces. OpenAI launched Daybreak, a vulnerability protection initiative joined by Cisco, CrowdStrike, and Cloudflare, while Google Cloud and OpenAI are hiring deployment engineers and standing up consulting arms to help enterprises operate AI safely. 

Why it matters: The hard problem is no longer building AI systems, it is deploying and securing them at enterprise scale. Organizations are adopting AI faster than they can train employees to use it, and traditional security tools were not designed for systems that act on their own across workflows.

5) How AI Is Turning UnitedHealth, CVS and Elevance Into Software Companies

Summary: Becker's reported on May 14 that UnitedHealth, CVS, and Elevance are turning AI from internal efficiency tools into external software revenue streams. UnitedHealth's $1.5 billion AI spend includes a one-third allocation to becoming an "AI-first software and services firm," Elevance's Health OS has cut denials by 70%, and CVS launched Health100 with Google Cloud. 

Why it matters: The line between payer and software vendor is collapsing. For anyone in payer operations or specialty pharmacy, the question is whether these platforms ultimately serve patients or just add new revenue extraction to an already complex market.

I built two agents to save time. They work and solve real problems. They are also the consolation prize I am about to warn you about.

The first agent I built turns a rough idea into a working draft of Founder's Corner, work that used to consume an entire day. Every morning, the second agent reads news sources and emails me what matters before my first meeting of the day. Real hours saved every week. But neither helps me grow subscribers or open new revenue streams.

The trap I am in is the same trap most companies are in, just at a smaller scale. Boards are approving AI budgets, CFOs are tying success to cost reduction targets, and almost nobody is asking the question that decides whether any of it pays off.

Automation is the consolation prize; the real game is using AI to redesign the business model itself.

The Question I Was Not Asking

I am guilty of this. The agents I built were doing their job and I was treating that as progress. The realization arrived while listening to an episode of The AI Daily Brief, "How the Best Companies Use AI." The host, Nathaniel Whittemore, drew a throughline about leading companies treating AI as a growth and opportunity technology rather than a time-savings one. His framing reset how I was thinking about my own builds. 

The question that landed in my head, the one I had not been asking, was specific: "Now that I am freeing up time each week, what can I build to help me grow subscribers?" Right behind it came an admission I had been avoiding: "LinkedIn posts will never be enough long term."

The agents saved me time, but neither one is moving the metric that decides whether Neural Gains Weekly survives the long term. I was stuck automating tasks but not thinking about how AI could redesign the work itself.

The mistake does not change with scale. If you are inside a company that doubled its AI budget on efficiency pilots, without building the long-term AI strategy, you are in the same trap. Just with bigger ramifications. 

When the Easier Path Becomes the Only Path 

Building AI to automate tasks and cut costs is a good thing. It is the easier path to ROI, the easier place to experiment, and the easier story to put in front of a board. That makes it a logical entry point. But it cannot be the whole strategy.

This default mode has decades of muscle memory behind it. Every portfolio review skews toward what saves money this year, every business case anchored to cost reduction. BCG's AI Radar 2026 found corporations expect to more than double their AI investment in 2026, from 0.8% of revenues to about 1.7%, with more than 90% committing to keep investing even if returns do not arrive next year. The same survey found nearly three-quarters of CEOs now name themselves their organization's main AI decision-maker, twice last year's share.

The deeper number is what the spend is going toward. Deloitte's State of AI in the Enterprise 2026 split the field into three roughly equal groups: one-third (34%) are using AI to deeply transform their business, launching new products, services, or operating models; another third (30%) are redesigning key processes; and the remaining third (37%) are using AI at the surface level, with little or no change to what already exists. Deloitte's framing is direct: "only the first group is truly reimagining their businesses rather than optimizing what already exists."

Daron Acemoglu, the Nobel laureate at MIT, named the same trap in MIT Sloan Management Review: "organizations are choosing to use AI as automation technology when, in reality, it is a form of information technology. This explains why AI isn't improving productivity at a macroeconomic level."

Every company needs to ask where it falls in those three groups today, and what the next decade costs the ones outside the 34%.

The Second Pilot Costs More

Picture the typical scenario. Twelve months into the AI rollout, the first major pilot lands on the leadership review as a clean win. A high-volume workflow has been automated across three teams. Cycle time is down 22 percent. The case study writes itself, and the next round of funding is approved before the meeting ends. 

Then come the second-order costs nobody scheduled. The data feeding the agent was incomplete and downstream systems could not handle the volume. Compliance required an audit trail nobody had scoped. Integration with the existing ticketing system ran three months past plan. And the workflow itself, the one the agent was bolted onto, was never the right workflow to begin with.

You paid once to automate it. You will pay again to rebuild it. That is the tax you pay twice.

BCG put the gap in dollar terms in their 2026 study on AI value capture: "when no explicit value logic has been defined, 10%–20% of anticipated value typically erodes before reaching the P&L." That erosion is not a model accuracy problem. It is the cost of bolting a new operating layer onto an old workflow that was never going to support it.

Inside the 34%

The leaders in this AI decade are using AI to build what was not possible before. They are creating revenue streams the old infrastructure could not support, customer experiences the old workflow could not deliver, and business models that simply did not exist.

In their 2025 Global AI Jobs Barometer, PwC found industries most exposed to AI grew revenue per employee three times faster than the least-exposed industries, 27 percent compared to 9 percent. PwC put it directly, "Treat AI as a growth strategy, not just an efficiency strategy."

BCG's September 2025 report on AI leaders and laggards found leaders delivered double the revenue growth and 40 percent more cost savings. The leaders won on both at the same time. Automation got them to the starting line but did not move them past it.

Tempus AI is an example of what this looks like right now. The company uses AI to turn clinical and genomic data into personalized treatment recommendations for oncology and cardiology. They built the business model around AI from the start instead of bolting it onto a legacy diagnostics company. Q1 2026 revenue grew 36.1 percent year over year to $348.1 million, with 2026 guidance raised to $1.59 to $1.60 billion.

In the ALERT trial with Medtronic, Tempus's AI-driven EHR notifications surfaced patients with significant disease who could benefit from heart valve replacement. The result was a 40 percent increase in those life-saving procedures. New revenue and new clinical impact, from the same redesign.

Tempus is the verifiable case, but the principle is universal. The companies treating AI as the foundation of new business create advantages cost-cutting cannot reach. Every quarter spent layering instead of building is a quarter the gap to the 34% widens.

Find a Pilot That Redesigns

If your AI pilot's success metric is "time saved" or "cost reduced," you are building the consolation prize. If the success metric is "new revenue," "new customer experience," or "new business model," you are playing the long game.

That is the litmus test. Apply it to every AI initiative on your roadmap. The ones that pass are the work that pays back. The ones that fail are still useful, still necessary, still worth doing.

You do not need to wait for the next portfolio meeting. Find one pilot inside your company where the success metric is something other than time saved or cost reduced. If you cannot find one, propose one. If you cannot propose one, build a small version yourself, as I am rebuilding mine for Neural Gains Weekly.

The next decade waits on the other side of the question you have not asked yet.

The Situation

Dana, our clinical operations coordinator, has been using Microsoft 365 Copilot every day for weeks now. She knows the apps. She has tried Researcher and Analyst. She drafts in Word, summarizes in Outlook, prepares for meetings in Teams. By every measure she would have used six weeks ago, she is a Copilot user. There is one question she cannot answer when her director asks her about it on a Tuesday morning. Is Copilot actually saving her time, or is she just clicking the icon more often?

What She Tries First (And Why It Falls Short)

Dana does what most professionals do when asked that question. She thinks about her gut sense of how the week went. She has a vague impression that Copilot helps. She would not be using it this much if it did not, right? She tells her director that she thinks it is saving her a few hours a week.

The honest answer is that she does not know. Her impression is built on a handful of memorable wins (the Friday afternoon when Copilot summarized a 14-reply email thread in 30 seconds), a handful of memorable losses (the Monday she rewrote half a Copilot draft and lost an hour), and a lot of usage she cannot remember at all. The wins feel bigger than the losses because the wins ended the task and the losses extended it. Memory is a bad measurement system.

By Wednesday she decides to do something her director did not ask for. She is going to audit her own usage for the rest of the week.

The Concept, Through the Scenario

For the next four days, Dana tracks every Copilot interaction in a simple text file. What she used. What she asked. What she got back. How long the whole interaction took, including the time she spent reviewing, editing, or verifying. By Friday she has 23 entries.

When she sorts them, three categories emerge.

Time Saver. Tasks where Copilot reliably cut time and the verification cost was low. The biggest entries are the ones from Vol 31. Dana asked Copilot in Outlook to "summarize the key decisions and any open questions from this email thread" for a Tuesday morning thread of 14 replies between finance, IT, and her director about a contract renewal. The summary came back in 30 seconds as four clear bullets. She skimmed it, recognized the names, and acted on it. Net time saved: roughly 12 minutes. The verification cost was almost zero because she would have skimmed the thread anyway.

Coin Flip. Tasks where the output was inconsistent and verification took real focus. Wednesday morning, Dana asked Copilot in Word to "draft the executive summary section of the quarterly compliance report using the audit findings spreadsheet I just attached." The draft was 80 percent there. The 20 percent that was off was the part that mattered most: the wording around two open audit items where legal framing matters in compliance reporting. She rewrote those two paragraphs by hand. Net time saved: maybe 8 minutes. The verification took her full attention because she had to know exactly where the draft was unreliable.

Net Loss. Tasks where Copilot created more work than it saved. Thursday afternoon, Dana asked the Researcher agent (Vol 33) to "compile a competitive overview of three telehealth vendors my team is evaluating, focused on EHR integration capabilities." The report came back structured and confident with citations. When she opened the citations, two of the three "top tier" sources were vendor-sponsored content marketing dressed up as independent analysis. She spent more time verifying and discarding sources than she would have spent running the search herself. Net time lost: about 40 minutes.

Three buckets. One question. Before she invokes Copilot for any task going forward, Dana asks herself a single thing.

What is the verification cost?

If the verification cost is low (she would have read the email or skimmed the transcript anyway), Copilot almost always earns its place. If the verification cost is high (she has to fact-check citations, audit data structure, hand-edit specific paragraphs where precision matters), the time saved on the front end has to clearly exceed the time spent verifying on the back end. When it does not, the honest move is to stop using Copilot for that workflow and do the task without it.

What Changes

The following week, Dana applies the framework in real time.

Monday morning, she opens her inbox and sees a 16-message thread about Friday's provider scheduling conflict. Bucket 1. She clicks Copilot, gets a clean summary, acts on it. Ninety seconds.

Tuesday afternoon she has the next compliance report to draft. Instead of asking Copilot to generate the whole thing, she uses it to draft the routine sections summarizing standard audit findings (Bucket 1 for those passages), and she writes the legal-framing paragraphs herself. The report takes 90 minutes instead of three hours, and the parts that matter most are her own words.

Thursday she has vendor research to do for a different evaluation. She does not open Researcher. She runs the search the old way, in a regular browser, with the sources she trusts. The work takes the same amount of time it would have taken before Copilot existed, but she trusts the output and does not have to chase citations to confirm what is real.

By Friday Dana has not used Copilot less than the previous week. She has used it differently. The difference shows up in her actual finish time on Friday, not in how many times she clicked the icon.

What This Reveals

Usage is a bad proxy for value. Some of the highest-usage Copilot users in any organization are spending more time verifying outputs than they save on the work itself. They feel productive because the screen is busy. The audit Dana ran is what separates feeling productive from being productive.

The framework she built is portable. The buckets and the verification cost question work for Claude, for ChatGPT, for Gemini, for Perplexity, and for whatever lands next. The names of the tools change. The audit does not. Most professionals will go their whole careers using AI tools without ever running this audit on themselves, and they will not know what they are leaving on the table or what they are quietly losing.

Your admin configuration shapes which features you have access to in Copilot specifically, and that shapes which buckets your usage falls into. Some of what Dana experienced will look different in your tenant. The framework still applies. The specifics will be yours.

This is the accountability the series closes on. Microsoft is not responsible for getting value from Copilot in your role. Your IT team is not responsible. Your company's adoption metrics are not responsible. You are. The tool does not know your workflows. You do. The question of whether it is earning its place in any specific task is one only you can answer, and the only way to answer it is to do the work of looking. The audit is the work.

How This Connects

This volume closes the Copilot Deep Dive series. Vol 29 introduced the ecosystem. Vol 30 explained the architecture: Microsoft Graph, grounding, permissions. Vol 31 walked through Outlook and Teams. Vol 32 covered the production apps honestly. Vol 33 explored the agentic layer: Researcher, Analyst, the in-app agentic capabilities, and Agent Builder. This volume gives you the framework to turn all of that into something measurable.

The bigger arc the curriculum is building toward is this: the tools you have access to will keep changing. The mental models you build around them are what compound. Knowing how to audit a tool, sort your usage, and ask the verification cost question is the durable skill. Copilot is just where you practiced it.

Part 6 of 6 in the Copilot Deep Dive series.

The Research Rabbit Hole Stopper

Most of us research the same way: a Google search, 12 open tabs, a Reddit thread, and two hours later we still cannot answer the original question. This workflow uses Perplexity, the one free AI tool purpose-built for source-cited web retrieval, to deliver a 3-source synthesis with a recommended next step. Healthcare decisions, financial choices, and big purchases get easier when you stop reading and start synthesizing.

The Workflow

1. Frame the Real Question (1 Minute)

Open Perplexity (perplexity.ai). No login required for basic use, free account unlocks more. Take 60 seconds to sharpen your research question before typing. "Should I get a colonoscopy at 45 or wait until 50" is a real question. "Colonoscopy info" is a rabbit hole. Write down the actual decision you are trying to make or the actual thing you are trying to understand.

2. Run the Grounded Retrieval (3 Minutes)

Paste the prompt below into Perplexity. The structure forces it to do the synthesis work, not just dump links.

Copy/Paste Prompt: "I am trying to answer the following question or make the following decision: [insert your question or decision].

Search the web and give me a structured answer with these parts:

1. The short answer (2 to 3 sentences) directly addressing my question.
2. Three of the most credible and recent sources you found, with the publication name and date for each. Prioritize peer-reviewed research, established institutions, or recognized expert sources over blogs and forums.
3. The key points each source makes, in 2 to 3 bullets per source.
4. Where the sources agree and where they disagree.
5. What I am still missing or what would change the answer based on my specific situation.

Use only sources from the past 2 years unless older research is foundational to the topic."

3. Pressure-Test the Synthesis (3 Minutes)

Read the output. Then send the follow-up below to stress-test what came back.

Copy/Paste Prompt: "Now play devil's advocate. Find me 1 to 2 credible sources that push back on or complicate the answer above. Look specifically for: research showing different conclusions, expert voices who disagree, recent updates that change the picture, or context that makes this question harder than it first appears. List the source, the counterpoint, and what it means for my decision."

This is where most AI research workflows stop too early. The pushback step is what turns a one-sided summary into actual decision-grade information.

4. Generate the One-Page Brief and Next Step (3 Minutes)

Now lock the research into something you can act on or share.

Copy/Paste Prompt: "Take everything above and turn it into a one-page brief I can save or share. Use this structure:

1. Question I was trying to answer.
2. Best current answer based on the synthesis (3 sentences).
3. Top 3 sources with publication and date.
4. Where the experts disagree.
5. My recommended next step (a specific action, conversation, or follow-up question to take this forward).

Format clean, no fluff, ready to copy into a Google Doc or send to someone helping me think this through."

Copy the brief into a Google Doc, your Notes app, or paste it into a message to the person you are deciding with (a doctor, a financial advisor, a partner). You went from open tabs to a referenced one-pager with a next step in 10 minutes.

The Payoff

You just replaced two hours of tab-hopping with a sourced, pressure-tested, one-page brief that ends in a specific next step. More importantly, you used the one AI workflow that actually rewards research questions: grounded retrieval with cited sources, then adversarial review. That is how decisions get made with information, not against it.

🧠 The AI Concept You Just Used

Grounded retrieval and source citation. When an AI tool pulls from live web sources and cites them inline, you can verify what it said and follow the trail back to the original. That is the difference between an AI giving you an answer and an AI showing you the evidence. Always check the sources, especially on medical, legal, or financial questions.

Transparency & Notes

• Tools used: Perplexity (perplexity.ai), free tier. Web retrieval and inline source citation are the core mechanic, so Perplexity is the recommended tool here. Microsoft Copilot, ChatGPT with search, and Gemini with search can run a version of this workflow, but Perplexity's free tier is purpose-built for cited research.
• Privacy: Keep your question general enough to protect personal details. For medical research, you do not need to include your name, exact age, or specific identifiers in the prompt. The tool gives you better answers when you stay focused on the decision, not your full profile.