
Interesting data about Marketing teams and the tools they are using compared to other functions in organizations from our CMO @denislarridin
✓Measure, optimize, and maximize the value of AI across your organization.

Interesting data about Marketing teams and the tools they are using compared to other functions in organizations from our CMO @denislarridin

What’s a Monday morning post without referencing productivity, right? 😁 We looked at 12 engineering orgs with both AI spend and output data. In 10, the cost per unit of measured engineering output was higher in October than in their first observed month. This chart shows an exception. About 70 engineers. Their AI coding bill was $35,000 in July and $34,000 in September. Almost unchanged. Cost per unit of Engineering Output fell from $14 to $7. Roughly twice the measured output for similar spend. The bill alone would have missed that change. It also wouldn't have told you that most orgs in this small sample had a higher unit cost at the end of their observed period. Looking at the exception alone, you could write a reassuring story about AI investment coming good. Looking only at the 10 of 12, you could write a gloomy one. I'd be careful with both. What I'd want to understand is what changed inside this team. Did they get better at using the tools? Did the mix of work change? Did they change the workflow? How much of what got finished needed doing again? These are the kinds of questions we’re working on @Larridin. Those questions matter if you're deciding where the next dollar should go. More tool access might help one team. Another might need training, better tests, or time to fix a workflow that's making everyone's job harder. A bigger bill doesn't tell you which problem you have. Engineering Output is a complexity-weighted measure of merged PRs, not customer value. We added Codex billing in August, so spend coverage expanded during this comparison. These numbers don't establish that spending caused the improvement. I wouldn't use this org to argue that everyone's AI budget needs to go up. But I wouldn't cut it because the bill looked high in July, either. I'd want to understand the work before making the budget decision.

A really interesting comparison by our head of product @mistermedici comparing the notion of trying to find cheaper token options to the fuel hedging that took place with airlines in the early 2000s. "My concern is that a cheaper unit can make an inefficient process look like progress." Just chasing the cheaper unit does not mean the outcomes will be strong or the cost savings even real.

The ‘one engineer running ten agents' demo is real. It's just rare. At @Larridin we help customers measure concurrent coding agent sessions across engineering teams. Here’s one approach: a session counts as live if it got at least two prompts in the same 15 minutes. Subagents don't count. Across about 2,000 engineers at 11 companies, over four weeks: - Half ran two at once at least once. - About 1 in 6 got to three. - About 1 in 40 hit five. A normal day is calmer. On about 3 in 4 of the days people were actively using an agent, they ran one at a time. The pattern was visible at almost every company. Most people top out at two or three. My guess: the limit isn't the agent. It's the human reading what comes back. Two agents is pair programming. Five is a standup where you're the only person talking😅


Does this resonate with you? According to @Larridin , Inc. data, about 1 in 6 non-technical employees is AI-fluent: real work, 3+ days a week. Daily users get nearly twice as much time back as occasional ones. Larridin helps you turn “not yet daily” users into “daily” users. We’re turning this into a map of six AI work types. Recognize yourself? What’s missing?

McKinsey’s 2026 State of AI survey found that 80% of respondents say AI has made them personally more productive, but only 37% say it has moved EBIT at all (which btw is the same as last year). And the share of real high performers is still stuck around 6%. KPMG’s AI pulse puts it more bluntly: just 7% of leaders say they’ve established a measurable return on AI. Those numbers aren’t a contradiction. Feeling faster and moving the P&L are different outcomes, and the gap between them is the whole story. Part of it is that ROI means different things to different people. For the CFO it’s the P&L, for a team lead it’s hours back, and for the person doing the work it’s not having to do the boring part anymore. All of those are true, and none of them adds up on its own. What connects them is the work itself, and few companies have a clear picture of how theirs actually happens: which steps, in which apps, by which teams, how often. Without that picture it’s hard to redesign anything, and ROI ends up as a guess with a slide wrapped around it. We start there, @Larridin: map how the work happens, see where AI is actually changing it, and tie that back to what it costs. (For the record I’m firmly in the 80%. Convincing my inner CFO is another matter.)

Clear insights from @nicktabbal for developer productivity. "The least productive devs aren't leaning on AI. They use it least."

AI bills are not as straightforward as you may think. @mistermedici our head of product explains: "Actual tells finance what they owe. Observed tells engineering what's being used. A total that mixes them is a number nobody pays."

"Every AI number should say what kind of number it is" Our product leader @mistermedici sharing detailed explanations of our our four key metrics really measure AI Impact.

One of the most insightful features at @larridin, when it comes to AI agents, is our ability to score how ready codebases are for them. Turns out the answer depends on which repos you look at. Across every repo we've analyzed at the companies we measure, only about 1 in 16 has an AGENTS.md (the file that tells an agent how to build, test and work in that repo). About 2 in 3 repos sit at the lowest of our five 'readiness' levels. Now look only at the repos where engineers actually ran coding agents in the past four weeks. At every company we checked, AGENTS.md was several times more common there, and about half those repos are at 'readiness level' 3 or higher. Agent work is concentrated too. About 10% of those repos get 70% of the agent sessions. So "is our codebase ready for agents?" is the wrong question. Most of it isn't, and most of it doesn't need to be yet. The useful question is whether the few repos where agents do the real work are ready, and which one is next.



We @Larridin are finding that with cross-company AI benchmarks the most interesting part isn't the median, it was how wide the spread is. Over four weeks, the typical engineer's AI tool bill was about $120. The top 10% were at $1,900 or more. Same job, 16x. Everyone else: about $20 typical, $220+ for the top 10%. Time saved has the same shape. The typical employee gets about 1.5 hours a week back. The top 10% get 12+ hours. So "what does AI cost per employee?" and "how much time does it save?" don't have one answer. A small group pulls the average way up. The better question is where your people sit on the curve, and what the top 10% are doing differently.

Some powerful data from our SVP of Insights @nicktabbal on non-technical AI usage in organizations and the wide range of how tools are being successfully used or not.

Congrats on the round. Proving AI actually paid off feels like the question every exec team is getting asked right now.