August 21, 2026

AI for the Over 40 – Week 32: The Below-the-Line Work AI Finally Makes Possible

A few weeks ago, an employee asked me a reasonable question: how could they know how many times they had submitted late time over the past year?

We had been publishing weekly late and missing time reports for years. The data existed, dozens of spreadsheets sitting in my inbox as email attachments. I did not have a good answer.

Not because the data was not there. Because getting it out meant downloading each attachment individually, opening each spreadsheet, copying the relevant data, pasting it into a master file, deduplicating, and building a pivot table. Hours of work, maybe a full day, for a question that, while important, was competing with every other demand on my time.

So I did what every leader does with below-the-line work. I did not do it. Not “I thought about workarounds.” Not “I considered alternatives.” I simply was not going to do that analysis. The question would go unanswered, like it always had, because the effort could not be justified.

One hour for work I never would have done

Anthropic had released Claude Cowork a while back, a desktop AI tool that can read and create files on your computer and connect to services like email through integrations. I had seen the announcement but could not think of a use case. Everything I needed from AI, I could accomplish through Claude Projects, upload the files I needed, get what I needed back, done. I could not imagine why I would need AI working directly on my desktop.

The compliance analysis was the first time I thought this might actually be what that tool is for.

First problem: Claude could read my emails through the Microsoft 365 integration but could not extract the attachments. The data was visible but not accessible. Claude could see the emails existed but could not get to the spreadsheets inside them.

So I turned to Power Automate, Microsoft’s automation tool that I have written about throughout this series. I built a flow to extract attachments from emails automatically, and immediately hit a wall. Getting Power Automate to search across all my email folders for years of specific reports was more complex than I expected. I spent about an hour troubleshooting, trying to get the flow to search the right folders, handle the right date ranges, filter for the right message types.

Then Claude made the suggestion that made this article worth writing.

Stop trying to automate the search. Just copy the emails you need into their own folder.

I opened Outlook, searched for the reports manually, and found them in about two minutes. I selected the emails I needed, dragged them into a dedicated folder, then pointed Power Automate at that single folder, a dramatically simpler task, and had it extract all the attachments.

I copied the extracted files into Claude Cowork’s working folder and asked it to read all the spreadsheets, combine the data, and analyze how many times each person had late or missing time over the past year.

One hour, start to finish, for analysis I simply would not have done.

And the exciting part was not just the compliance answer. It was realizing what this unlocked. Every snapshot report buried in my email represents a question I could now ask. Weekly reports, monthly summaries, quarterly reviews, data that cannot be recreated because it captured a moment in time, sitting in inboxes across every organization, never examined in aggregate. The compliance analysis was one question. The pattern applies to dozens.

Below-the-line work

Every organization has an invisible effort threshold. Above the line is work that gets done because the value justifies the effort: monthly financial reporting, client deliverables, board presentations. This work happens because it has to. The consequences of not doing it are immediate and visible.

Below the line is work that would be genuinely valuable but nobody can justify the hours: the historical compliance analysis I needed, the trend analysis buried in years of weekly reports, the pattern recognition that requires consolidating data from dozens of sources that nobody has time to consolidate.

Below-the-line work is not trivial. It is often exactly the analysis that would make better decisions possible, if anyone could get to it. But the effort-to-value ratio falls just short of what any individual can justify, especially when they are already stretched thin by above-the-line demands.below-the-line work with AI

Here is what makes this category different from what I have written about before. Throughout this series, I have documented AI helping me do existing work faster, research synthesis in Week 6, writing in Week 5. I have documented AI helping me build things that did not exist, the Todoist integration in Week 18, the Career Development Partner in Weeks 27 and 28. I have documented AI helping me redesign broken processes, contract routing in Week 10, performance management in Weeks 27 and 28.

But this was different. I was not doing this work slowly with AI making it faster. I was not doing this work badly with AI making it better. I was not doing this work at all. The manual effort was so prohibitive that the analysis simply never happened. The question went unasked. The data sat in my inbox, accumulating week after week, never examined.

AI did not speed up my process. It made the process possible in the first place.

The diagnostic question for your own work: what analysis would you run if the effort were essentially free? What questions have you stopped asking, not because the answers do not matter, but because you cannot justify the hours to produce them?

That is your below-the-line work. And it is probably more valuable than you think, precisely because you have never been able to see what the answers reveal.

The lesson I keep having to learn

Here is the part I almost do not want to admit, because I have written about this pattern before and apparently still needed to learn it again.

My first instinct was to automate everything. Get Power Automate to search all my folders, find the right emails, extract the attachments, and deliver them to Cowork, fully automated, no manual steps. When that did not work easily, I kept pushing. An hour of troubleshooting, trying to make the automation smarter, trying to eliminate every manual step.

And then Claude, the AI I was working with, suggested I just copy the emails to their own folder manually.

The AI told me to stop overcomplicating it.

This is the same lesson from Week 10. The contract routing solution was not more automation; it was recognizing that the real problem was process ownership, and a simple Microsoft Form put information capture where it belonged. The same lesson from Week 28: the “compromise” of separating evidence collection from AI analysis proved a better design than the seamless automation I originally envisioned.

But here is the twist this time: I did not catch it myself. The AI caught it for me.

I was deep in the troubleshooting weeds, an hour into Power Automate filter logic, and Claude essentially said: Why are you making this so complicated? Just move the emails yourself and let me handle the rest.

I have now hit this insight three times across very different contexts, and I think it deserves to be said plainly: the best automation architectures include manual steps. Not as compromises. As design choices.

The two-minute manual search was not a failure of automation. It was the most efficient part of the entire workflow. I am the one who knows which emails matter. I can scan subject lines and dates faster than any search filter I could configure. Trying to automate that judgment was making the entire project harder for zero benefit.

The pattern: figure out which steps require human judgment and which require machine execution. Then stop trying to automate the judgment parts.

My compliance analysis worked because three tools each did what they do best. I served as the intelligent filter, doing manual search, manual curation, and conscious decisions about what to include, in two minutes. Power Automate served as the reliable extractor, running deterministic file operations on a simple, well-defined folder with no judgment required, in minutes. Claude Cowork served as the analyst, reading dozens of spreadsheets, combining data, and producing the analysis I needed, work that would have taken hours manually.

If you have been following this series, you might recognize that pattern. It is the Intake-Engine-Bridge architecture from Week 26, just at a personal scale. I, as the intake layer, am parsing chaos and deciding what matters. Power Automate is the engine, with deterministic and reliable operations. Cowork as the analytical layer, making sense of what would have been impractical to process manually.

The same pattern that works at enterprise scale works for one person solving one problem.

The data excuse

Now let me get to something that has been nagging at me, because I keep hearing this from clients, and I think it is quietly becoming one of the biggest barriers to AI adoption.

“We need to clean up our data first.”

Not as an objection to AI, but as a genuine first step they believe is necessary before they can meaningfully engage. Classify documents, apply sensitivity labels, build a data governance framework, figure out what AI should and should not have access to, all before exploring what AI might do for them.

It sounds responsible. It feels like proper due diligence. And at an organizational level, it has real merit. When you are deploying AI across an enterprise, data governance matters. I wrote about data sovereignty in Week 26 for exactly this reason. Sensitive data needs to stay within compliance boundaries. That is not optional.

But here is what I have realized: “we need to clean up our data first” is the data version of “we are waiting for the right use case.”

It sounds responsible. And it is a perfect reason to never start. Because the scope is paralyzing, years of documents across SharePoint, OneDrive, shared drives, email. The idea of classifying all of that before you can begin using AI is not a first step. That is a multi-year initiative that most organizations will never finish.

What I actually did for my compliance analysis: I searched my email for the reports I needed, copied them to a dedicated folder, extracted the attachments, dropped the files into a working folder, and gave AI access to exactly those files, nothing else.

That is not enterprise data governance. That is a conscious decision about what a specific task requires. I did not need to classify my entire inbox. I needed to decide, for this task, what AI should have access to, and give it exactly that.

The reframe: instead of the impossible top-down approach, classify everything, then decide what AI can touch, use the practical bottom-up approach. For each task, curate what you need. Give AI access to exactly those files. Keep everything else out. Not through enterprise governance infrastructure. Through a folder and a conscious choice.

To be clear, this does not replace enterprise data governance for organizational AI deployment. When you are deploying agents that access systems autonomously or building RAG systems against corporate knowledge bases, you need comprehensive data classification and security. Rule 4 from Week 26 still applies: sensitive data stays in the garden.

But that is not where most leaders need to start. Most leaders need to start by solving one problem with AI. And for that, you do not need a data governance initiative. You need a folder and a few minutes of intentional curation.

The sequence I used to recommend, build literacy, then get your data in order, then apply AI to real problems, had the last two steps backwards. You learn more about what data governance you actually need by doing AI work than by theorizing about it in advance. Start applying AI to real problems, curating task-specific data as you go, and let that experience inform your governance strategy.

Your Week 32 challenge: find your below-the-line work

This week, identify one question you have stopped asking. What analysis would be genuinely valuable for your work but you have never attempted because the manual effort could not be justified? Consolidating data from multiple reports? Analyzing patterns across months of documents? Comparing versions of something that has changed over time?

Map where the data actually lives. Email attachments? SharePoint? Shared drives? You probably know roughly where it is, you just never had a reason to gather it. Spend five minutes thinking about what you would need to collect.

Try the working folder approach. Create a folder. Put only the files relevant to your question in it, nothing else. You will have just done more practical data curation than most enterprise governance initiatives accomplish in their first six months.

Ask yourself what has been too hard. Not too hard for AI, too hard for you, manually. What work has been sitting below your effort threshold that might suddenly be practical? The answer might surprise you.

Notice the compound question. Once you answer the first below-the-line question, notice what new questions it surfaces. My compliance analysis did not just give me numbers, it gave me the ability to have specific conversations with specific people about their time entry patterns before the new compliance year started. Without the analysis, I would have been sending generic reminders to everyone. With it, I could target the people who actually needed the conversation. Below-the-line work tends to cascade: answering one question reveals three more you can now ask.

The bottom line

For thirty-one weeks I have been documenting what happens when you bring AI into your actual work. Most of those weeks were about doing existing work differently, faster, better, with new tools and new thinking.

This week was about something I had not quite named before: doing work that simply was not happening. Not because it was not valuable. Because the effort threshold was too high for any human to justify.

Dozens of spreadsheets. An hour of work that included building a Power Automate flow I had never built before, discovering that my desktop AI could not do what I assumed it could, and realizing, again, that the best solution included manual steps I had been trying to eliminate.

The answer to my compliance question was always there, sitting in my inbox, accumulating week after week, waiting for the effort threshold to drop far enough that someone would finally do the analysis.

I could have waited for enterprise data governance. I could have waited for AI to seamlessly access every system without friction. I could have waited for the fully automated pipeline that required no manual steps at all.

Instead, I created a folder. And got my answer.

This post is part of my “AI Over 40” series. It first appeared on LinkedIn: AI for the Over 40 [Week 32]: The Work Nobody Does

Read more AI and Copilot blogs.

Trending Posts