This week the data caught up with the hype, and it drew a sharp line between people who produce AI results and people who only forecast them.
The Snapshot
A Federal Reserve Bank of St. Louis study analyzing roughly 490,000 corporate earnings calls found that AI has not yet produced a measurable aggregate productivity boost. About 95% of the AI-related productivity comments from executives described expected future gains rather than realized ones. That gap between what leaders promise and what the economy can actually measure is now the defining feature of the two-track labor market.
A. The New Performance Standard
For most of the last two years, talking about AI was enough. Mentioning a pilot, a copilot rollout, or a "transformation roadmap" signaled that you were forward-looking. The Fed study quietly ends that era. When 95% of executive AI productivity comments are about future gains, it means the market has been rewarding intention, not output. That cannot last, because eventually someone asks where the numbers are.
Here is what this means for you as a knowledge worker or small team. The performance standard is shifting from "Are you using AI?" to "What did AI actually change in your work?" The first question is easy to answer and impossible to verify. The second is harder, and that is exactly why it will start to matter. If you can point to a task that used to take a day and now takes an hour, with the same or better quality, you are on the track that pulls ahead. If your AI story is still a promise about next quarter, you are indistinguishable from the 95%.
Practically, start keeping a realized-gains log. Not projections, not vibes. Write down the specific thing you shipped, how AI changed the process, and what the before-and-after looked like. You are building the evidence the aggregate data is still missing, and in a market flooded with promises, evidence is the scarce asset.
B. Proof and Stewardship
There was no dedicated governance signal this week, but the macro data carries a governance lesson we should not skip. The gap between AI promises and realized performance is, at its core, a trust problem. When almost all the claims point to the future, the burden of proof lands on whoever is closest to the actual work.
That is you. In your organization, on your team, in your own portfolio, you are becoming the auditor of your own AI claims. Stewardship here means refusing to inflate. It is tempting to describe an AI experiment as a win because everyone else is doing the same. But the Fed study shows what happens at scale when everyone rounds up: the aggregate numbers refuse to move, and eventually credibility corrects.
The stewardship move is to separate three things clearly whenever you report on AI: what you expect, what you have tried, and what you have verified. Most people collapse all three into one confident sentence. If you keep them distinct, you become the person whose AI claims can be trusted, and in a two-track market, trust is what routes the good work, the budget, and the authority toward you.
C. AI Skills and Workflow Design
This week's clearest skills lesson comes from a small refactor with a big principle behind it. Greg rebuilt the EAID content automation worker so that it retrieves genuine grounding material before it generates any LinkedIn drafts. Previously, the main defense against fabrication was an instruction: essentially telling the model "don't hallucinate." That instruction is now a backstop, not the primary strategy.
The new design calls existing tools, synthesize_sessions and search_knowledge, through a delegation pattern to pull real session history and knowledge base excerpts, then feeds that material into the generation prompt. The model is no longer asked to be honest through willpower. It is given something real to be honest about.
This is the workflow principle worth internalizing, and it maps directly onto the macro signal. You cannot instruct your way to real output, whether you are prompting a model or reporting to a board. Stricter language about accuracy does not create accuracy. Grounding in real material does.
Apply it to your own AI use. When your AI output feels impressive but hollow, the fix is almost never a cleverer prompt telling it to be more rigorous. The fix is upstream: give it your actual data, your real documents, your genuine history, before you ask it to produce anything. The two tracks in the labor market increasingly divide along this exact line. One track prompts harder and hopes. The other track designs retrieval first, so the model works from reality and the output survives scrutiny.
The skill to build this week is grounding design. Ask, for every recurring AI task you run: what real source should this be drawing from before it generates? If the answer is "nothing, it just writes from the prompt," you have found the reason the output is generic, and you have found the thing that separates a demo from a result.
This Week's Moves
- Start a realized-gains log. Record only tasks where AI actually changed your process, with a clear before-and-after, so you are not one of the 95% describing future promises.
- Separate your AI claims into three buckets whenever you report: expected, tried, verified. Never let them blur into a single confident sentence.
- Audit one recurring AI task and ask what real source it should retrieve from before generating. If the answer is nothing, redesign it to pull genuine material first.
- Replace one "be accurate" style instruction in your prompts with an actual grounding step. Give the model your data instead of your demands.
- Pick one AI output you were about to circulate and check it against real source material before you send it. Make verification a habit, not an afterthought.
If you want help turning these moves into a repeatable grounding workflow for your team, reply and I will point you to the EAID piece that operationalizes exactly this pattern.
Next step