Your dashboard is getting very good at its job

An AI dashboard can tell you who’s opening a tool. An employee survey can tell you who’s feeling supported. Neither can tell you everything you want to know about useful work. This issue, we keep finding the same temptation: take the easiest signal to collect, put it in a chart, and let it stand in for the harder question.

A daily prompt can keep the dashboard happy

Business Insider reports that KPMG tracks AI usage against targets, while two employees say those measures are easy to manipulate; KPMG describes broader training and innovation incentives too.

Our take: Reward visible activity and people will find efficient ways to produce it. We would give a useful improvement more attention than a perfect streak. The person who eliminates a recurring task might have an excellent reason to open the chatbot less often.

A promise of results changes when the money arrives

The Financial Times reports that McKinsey plans to retain more partner remuneration as equity as fees tied to client outcomes make cash receipts less predictable.

Our take: A contract that pays for results makes the definition of “results” expensive to get wrong. Agree on the baseline, whose contribution counts, and when the outcome can reasonably be observed before celebrating the pricing model. The cash may arrive years after the presentation.

Even a very large survey has a boundary

Microsoft’s Work Trend Index finds organizational conditions strongly associated with reported AI benefits in a survey of 20,000 people who already use AI at work.

Our take: This is a useful challenge to the “employees just need better prompts” theory of adoption. But a survey of AI users can’t speak for everyone, and reported benefits don’t establish that culture caused better performance. Use the findings to examine what managers reward, then look at the actual work before declaring victory.

A promising signal still has to deliver

Microsoft’s Katy George argues in TIME that AI value can emerge through better decisions and customer engagement before it appears in labor savings, and recommends connecting early indicators to business outcomes.

Our take: We like the wider definition of value. We are less convinced by the essay’s assurance that AI is delivering and organizations are simply measuring it badly. More customer time is a plausible step toward stronger sales, not a receipt for them. Name the expected connection and revisit it when the outcome should be visible; otherwise “the value comes later” can become an indefinitely renewable excuse.

A benchmark label isn’t a validated capability

In a March essay, Ben Wong argues that DeepMind’s proposed cognitive framework for AGI evaluation needs stronger engagement with psychometrics to establish what its scores actually permit us to infer.

Our take: A score labeled “reasoning” can acquire a lot of authority before anyone demonstrates what it predicts. When choosing a model for work, ask what the evaluation reveals about the task you need done. An impressive score on familiar questions gives us little reason to trust a system’s judgment in an unfamiliar handoff.

Cheap tools don’t automatically produce builders

Daniel Williams argues that cheaper AI tools expand who can build, while organizations may still reward presentations and coordination over working results.

Our take: A rough prototype can make an argument a slide deck never will. We also think the cry of “everyone should build” needs a follow-through clause: someone must maintain the thing, explain it, and help colleagues use it. Five impressive tools nobody owns can become five new chores.

“Do you like AI?” hides most of the useful detail

Ethan Mollick suggests that broad AI sentiment questions blur together very different reactions to particular uses.

Our take: Someone can love an editing assistant and dislike an automated performance rating without being confused. Ask which application they mean and what it does to their work. A single enthusiasm score is convenient for a presentation and pretty unhelpful for deciding what to improve.

One useful move

Have a weekly update full of activity and a little light on accomplishments? Ask your AI tool:

Status update: {{status_update}} Rewrite this update so the useful result comes first. For each item, distinguish what I did from what changed for someone else. Use only what’s here. Where I’ve described activity without a result, leave a short question for me to answer.

Answer the questions you can, and keep unfinished work clearly unfinished. Your reader gets a better update; you get a clearer account of what mattered.

Something fun

Design a terrible office leaderboard:

Invent five office metrics that would get worse the more successfully employees optimized them. Give each a trophy name and a ridiculous winning strategy.

Our entry: the Golden Reply-All, awarded for keeping everyone informed until nobody can find the original question.

Built in the past two weeks

We reworked a draft resource page in the Mastermind’s assistant-building curriculum around what a learner needs next: core lessons, custom-assistant walkthroughs, preparing reference files, and guides for other platforms. The page now explains when each resource is useful and makes the optional reading easier to distinguish from the core path.