By Damian Mathews & The Last Mile Team
“Oh, funny. I built something that’s extremely similar.”
That’s what a coworker said to me on Tuesday (and he wasn’t exaggerating).
I’ve spent the past few months building a number of AI marketing tools for our team. Real ones, doing real pieces of my job. I showed one off on a call this week, and it turned out a colleague had something similar he’d been using. Within a day I’d found a third version… not identical, but also very similar in concept.
Same idea, three desks, none of us knew.
A couple of years ago that would’ve been almost impossible and probably considered a process failure. Building software was expensive, so duplication was waste, and coordination came first. Entire committees existed to prevent exactly this conversation.
Now a working prototype costs an afternoon, if that. When building gets that cheap, the economics of duplication flip, and the question becomes whether three of the same thing is a problem or a feature.
And if you run a contact center, some version of this is already happening at your desks. A QA analyst with a scoring prompt. A supervisor with a call summarizer. A workforce planner with a forecasting sheet nobody asked for. Keep those three in mind, because this story is really about what to do with them.
The wasteful reading has a spectacular new case study. Michael Fisher published an essay last week, Ardeo Ergo Sum, that opens inside Meta, where an intranet leaderboard ranked employees by AI token consumption. In 30 days it logged around 60 trillion tokens, roughly $900 million worth at list prices. Uber, meanwhile, reportedly burned through its entire 2026 AI budget by April.
So yes. The burn is real money, and three of everything sounds like a bonfire.
But Michael’s larger point is that Meta’s problem was never that people built too much. Consumption became the target. Goodhart’s law kicked in, the old rule that a measure stops measuring the moment you start managing to it, and burning tokens began to prove enthusiasm rather than impact.
The caps and budgets now arriving are the same mistake running in reverse. A measure can hold a number. It can’t necessarily hold an intention.
Which brings me back to my three duplicate tools, because there’s a version of this that’s healthy. Kerry Robinson named it in A1B: Customer Zero to AI-First, the play he calls the Creativity of the Crowd. Give people the tools and the permission, and the ones closest to the work will show you what needs building. Or they’ll build it themselves, which is increasingly common.
Three people independently building the same tool is that play working. Nobody commissioned it three times. It got built three times because the friction is that real. That’s evidence, generated for the price of some tokens.
What turns the variation into progress is selection. The best version only rises if there’s somewhere for it to rise. A place people show what they built. Someone with the judgment to pick one, merge the best parts, and retire the rest. Skip that, and you get forty half-finished things, a startling invoice, and an austerity memo that kills the good behavior along with the waste.
Fish gives the selection job a name. Purposes are properties of principals. Dashboards can meter the burn. Only a person can hold the point.
We wrote in The 40 Ideas on Your Whiteboard that the evidence for what to build lives in your conversations. It also lives in what your people are quietly building twice.
The organizations in The 20% Club treat that as a signal and select from it. The ones stuck in pilot mode either ban the building or drown in it, and both roads end at the same place, activity without returns.
I don’t know yet which of our three tools survives. Maybe all of them! Even if it’s only one, the two that lose will still have told us something worth more than the tokens they burned.
What’s being built twice (or thrice) in your building right now?
— Damian
Here’s what went down this week.
Bleeding Edge
Early signals you should keep on your radar.
AI copilots are becoming permanent ‘machine colleagues’ for support agents on the contact center floor. Early deployments report agents handling 9 to 12% more cases with double-digit cuts to average handle time. If those gains hold up at scale, the copilot could quickly become table stakes for competitive support teams.
TSMC just shattered its own revenue record, and AI silicon is squarely the reason. Second-quarter revenue hit roughly $39.6 billion, up 36% year over year, and June sales alone climbed almost 68%. When the foundry making everyone’s chips runs this hot, the AI hardware buildout looks nowhere near its ceiling.
Leading Edge
Proven moves you can copy today.
AI voice agents are now absorbing the peak call volume that once completely buried human support teams. These speech-to-speech systems hold natural, interruptible conversations and scale almost instantly when sudden call spikes hit. For seasonal or crisis surges, voice AI looks like a practical release valve, provided escalation paths stay solid.
Ode, a $1.5 billion AI implementation joint venture led by Anthropic and Blackstone, launched this week betting deployment beats models. Ode pairs Anthropic’s models with 100 engineers to close the gap between AI’s potential and what companies ship. When frontier labs invest in services, it signals the money may be moving toward integration, where firms like ours live.
Off the Ledge
Hype and headaches we’re steering clear of.
The customer service labor market is quietly shrinking as AI takes on more of the queue. US customer service job postings sit about 10% below pre-pandemic levels, and Forrester expects steep further cuts by 2030. The lower-complexity roles look most exposed, so reskilling toward AI oversight may be the safer bet for teams.
The first fully autonomous AI ransomware attack has arrived, and no human was at the keyboard. Researchers say JADEPUFFER’s AI agent broke in, moved laterally, and encrypted a production database, adapting to failures in real time. If agentic attackers scale, the barrier to sophisticated intrusions could drop sharply, and security teams should plan accordingly.
