Working across teams — engineering, product, operations, leadership — I keep noticing the same thing. Small habits around how AI sessions are run become second nature for some people and stay unknown to others. Not because anyone's doing anything wrong; these things are just rarely taught.
For one person, on one day, the difference is small. For an organization running thousands of AI sessions a week, it's the same story I wrote about in AI isn't expensive. Unmanaged AI is. — the headline cost is the vendor invoice; the quieter cost is every session that could have been better. The fixes are simple. Not technical. Not expensive. The kind of thing a 15-minute conversation can change.
Five things I keep seeing
A handful of patterns come up again and again. Each one has a simple fix.
Sessions become dumping grounds.
I see people run a single AI session across wildly different topics — a product strategy question, a quick code review, a sales email draft, a bug — all in one long thread. They've been told the AI "remembers," so they assume more context is better. It isn't. Every message carries the entire conversation with it, and the model starts juggling unrelated mental models. Quality drops. Hallucinations creep in.
The fix: one session, one topic. When the subject changes, close the session and start fresh. A 30-second habit that many people simply don't know they should have.
Nobody closes a session cleanly.
When a long, valuable session ends, I watch people just close the tab. The context — the reasoning, the decisions, the trade-offs — evaporates. Next time, they either reopen the old thread (slow and expensive) or start from scratch (losing everything they'd built up).
The fix: sixty seconds before ending, ask for a summary. Ten bullets: what was decided, what's still open, what to bring into next time. Save it somewhere — a project doc, a CLAUDE.md or AI-context.md file in the repo, a Notion page. That summary makes the next session twice as good, in a fraction of the time.
Before we end — give me a 10-bullet summary: what we decided, what's still open, key files or decisions, and what I should bring into the next session.
People lead with the question, not the context.
Most sessions open with "help me with X" and no setup. Five turns in, the AI still doesn't know who the audience is, what format is expected, or what tone is appropriate — and the user is doing the work of correcting each of these one at a time.
The fix: thirty seconds of context up front. Who is this for. What's the goal. What's the format. What's the tone. It doesn't take more time overall — it shifts the time from reactive correction to proactive framing. The first draft becomes something you can actually use.
Related work gets fragmented across sessions.
Three related deliverables — a PRD review, release notes, a customer announcement — get opened in three separate sessions. Each one starts with a fresh context that's nearly identical to the others. The duplication is invisible but real.
The fix: batch related work into one session. Set up the context once, then work through the items in sequence. You'll also catch cross-item inconsistencies that three separate sessions would miss.
Nobody knows when to walk away.
This one is costly when it happens. A session goes sideways — the AI keeps making the same mistake, the user keeps trying different phrasings, and the conversation starts feeling like an argument. Twenty messages in, no progress. The session is stuck. More attempts won't fix it.
The fix: close it. Open a fresh session. Include what the failed attempt taught you. That lesson is usually worth more than all the previous back-and-forth combined.
What to do about it
A 15-minute internal guide, a five-minute habit demo in a team meeting, a shared doc on "how we run AI sessions" — none of it is expensive. But the focus in most organizations is on tools and models, because those have clear price tags and procurement paths. Education happens on its own, or not at all.
If you lead a team or an organization and want to close this gap, three simple steps:
First, find who's already doing this well. Every team has a few people whose AI habits are worth learning from. Ask around — they're not hard to identify. The way they work is the starting point for what to share.
Second, make it easy to share. A short internal doc. A 20-minute recorded demo. A monthly "what I learned using AI this month" slot on a team call. The format doesn't matter. What matters is that the habits get named, written down, and passed on.
Third, measure something. Not everything. Just enough to see if it's working. Number of sessions per person. Average session length. Common failure modes people report. You don't need a dashboard. You need a pulse.
This isn't about the tools. It's about the ordinary, practical work of teaching people how to use them.
The organizations that will do best with AI over the next five years won't be the ones that bought the most expensive stack. They'll be the ones that built the simplest organizational muscle around it — starting with the small, everyday habits that are easy to share once someone notices they're worth sharing.