McKinsey Global Institute reported in November 2025 that 57% of U.S. work hours could, in theory, be automated using AI and robotics technology that already exists today, with AI agents alone capable of performing tasks that occupy 44% of work hours (McKinsey Global Institute). Most founders read that number and start shopping for tools. That is the wrong move. The number is a ceiling on what is technically possible, not a plan for what to hand off this month. The plan requires an audit, and most people run the audit wrong twice before they get it right.
Run Drucker's audit, not a vendor demo
Peter Drucker's time-audit method is still the template: record every activity and its duration as it happens, for three to four weeks at a stretch, twice a year, then ask of each recorded activity (Psychology Today):
What would happen if this were not done at all?
That question still works. What has changed is the second question you need to ask once an activity survives the elimination test: does this task require judgment I hold, or does it require execution I merely perform?
That second question matters because most real AI use is not a clean handoff. Anthropic's Economic Index, built from millions of anonymized Claude.ai conversations, found usage leaned 57% toward augmentation (AI collaborating on a task a person still owns) versus 43% toward automation (AI performing the task outright) (Anthropic). If you audit your week looking for tasks to fully delete, you will find few. If you audit for tasks where you can shed the execution and keep the judgment, you will find many more, and you will actually use them.
Where automatable work actually hides
The same Anthropic data found automatable work is not spread evenly across a job, it clusters. The top 10 Claude.ai task categories account for 24% of all sampled conversations, roughly 36% of U.S. occupations have AI touching at least a quarter of their associated tasks, and only about 4% of occupations have AI touching three-quarters or more (Anthropic). Translate that to your week: do not audit task by task. Audit by category. Find the recurring bucket, first-draft writing, calendar triage, meeting notes, competitive scans, and test the whole bucket at once. One bucket cleared is worth more than ten scattered tasks nibbled at.
Shopify already turned this into policy. CEO Tobi Lutke told staff, in a memo later posted publicly, that reflexive AI usage is now a baseline expectation at Shopify, and that employees must show they cannot get what they want done using AI before requesting more headcount (CNBC). That is the audit turned into a gate: before you add a person, produce the answer to Drucker's question and the augmentation question, in writing.
The subtraction problem
Here is where most audits fail even when the categorization is right. A Harvard Business Review study that followed 200 employees at a U.S. tech company for eight months found 83% said AI increased their workload rather than reducing it, driven by task expansion, blurred work and non-work boundaries, and more multitasking (Harvard Business Review). Handing a task to AI does not shrink a week if the freed hour immediately fills with a new task. The audit has to end with subtraction: name the hour AI frees, and remove it from the plate, not just from the to-do list.
The subtraction step also has to check quality, not just hours. A Stanford and BetterUp study of 1,150 U.S. full-time employees found 40% had received workslop, AI-generated work that looks finished but lacks substance, from a colleague, costing nearly two hours to fix per instance and an estimated $186 per employee per month in lost productivity (Harvard Business Review). A task that moves to AI but still needs a human to catch and repair the output was never actually automated. It was relocated, and it got more expensive.
This week, run the three-part version:
- Log every task for five days, not four weeks. That is enough to see the recurring buckets.
- For each bucket, ask Drucker's question first, then ask whether you need the judgment or just the execution. Move the execution-only buckets to AI, and delete the corresponding block from your calendar rather than filling it.
- Check back in seven days on whether anyone downstream is quietly fixing what the AI produced. If they are, the task was not automated, it was outsourced to a person you cannot see.