Automate the read, not the call
The useful question is not whether AI can do the work. It is whether the work sits upstream of a decision, or whether it is the decision itself.
Most teams I have worked with were not short on data. They were short on a shared read of what the data meant. Three people could look at the same week and walk away with three technically correct stories, each backed by a different chart. The bottleneck was rarely analysis. It was synthesis: pulling signal from Slack threads, spreadsheets, SQL tables, and half-written docs quickly enough that leaders could still act on it.
That is where AI has been most useful in my experience, not as a substitute for judgment, but as a way to compress messy context into something a team can actually discuss.
Two different jobs
I find it helpful to split the work into two layers.
The first is the read: gathering context, surfacing exceptions, comparing periods, connecting comments to metrics, and writing a plain-language summary of what changed. This work is repetitive, cross-source, and time-consuming. It rewards breadth and speed more than authority.
The second is the call: deciding what matters, what is noise, what tradeoff to make, and who owns the outcome. This work requires accountability. Someone has to live with the consequence, explain it to others, and change course when the facts shift.
AI is often excellent at the read. It is a weak and risky substitute for the call.
Where it has actually helped
The clearest example on my current team is the weekly business review. The meeting only works if everyone arrives with a shared picture of outcomes, exceptions, and open decisions. Building that pre-read used to mean hours of manual archaeology: pulling numbers, scanning channels for context, and trying to remember what changed since last week.
We now use AI to do much of that assembly work. It can pull context across Slack, Google Docs, and business metrics (sheets, SQL tables, and the rest) and produce a draft read of the week. It is not perfect. Often it lands around 80 percent: good enough to start the conversation, especially once it has learned which goals and topics actually move the room.
That 80 percent is valuable because it changes what humans do with their time. Instead of reconstructing the week from scratch, the team can focus on the parts that require judgment. Is this exception real or seasonal, and does it actually change what we should prioritize next? Are we presenting more certainty than the data supports? The automation did not make the decision. It made the decision better informed, and it made the meeting shorter.
The operating instrument for that forum is the weekly business review. This note is about the judgment split underneath it: automate the read; keep the call human.
I have seen the same pattern in other contexts: synthesizing partner feedback before a negotiation, comparing operational metrics across regions to find where availability is quietly thinning, or turning a pile of incident notes into a structured post-mortem draft. In each case, the automation did not make the call. It just got the room further along before anyone sat down.
What good automations optimize for
The automations that seem to hold share a few traits. They are built to improve insight, not to remove humans from the loop.
- They connect sources humans would otherwise stitch together by hand.
- They surface exceptions early, before a confident conclusion has formed.
- They arrive before the decision forum, while there is still time to act.
- They leave room for correction, so a person can still mark what matters and cut the noise.
That last one matters more than it sounds. Skip it, and the tool stops being an instrument and starts being cover for not thinking.
Where it goes wrong
The failure mode I watch for is subtle. A team automates the read, but over time the draft becomes the decision. Someone generates the memo, the forecast, or the recommendation. A busy reviewer gives it a quick pass. The work moves forward because it looks complete.
That is not automation creating leverage. That is approval theater. The organization gets faster output and weaker ownership at the same time. The document sounds confident. The underlying tradeoff never got examined.
This is especially risky with high-stakes calls: prioritization, incentives, partnership terms, performance feedback, anything that sets how other people will behave. Those decisions need a named owner, not a polished draft with a human signature at the bottom.
A simple design test
Before building an AI workflow, I try to ask one question: if this works exactly as intended, what becomes easier for the human afterward?
If it makes the business easier to see and the decision faster to reach, that is worth building. If it just lets someone skip thinking about it, that is a warning sign.
The best uses of AI in leadership work do not remove responsibility. They remove friction from the work that surrounds responsibility: gathering context, finding the anomaly, writing the first pass, comparing this week to last week, giving the room a shared picture instead of separate stories.
Automate the read. Keep the call human. That is usually where the leverage is.