Last week, Microsoft announced the Frontier Company: $2.5 billion and 6,000 experts embedded inside customer organizations. Add the implementation deals announced over the past 60 days, and the industry has bet roughly $9 billion on one assumption: that companies can buy their way past the hard part of AI. Our data says the hard part cannot be bought.
We track 10,000+ agents run by frontline teams at large enterprises, and 9 of 10 need retuning every week. Leave an agent unmanaged for 90 days, and usage drops more than 40%, satisfaction more than 50%. You have to build internal capacity that keeps agents working after the parachuted forward deployed engineers leave.
🎧 A heresy is the thing an industry does not want said out loud. In this week’s companion episode, Elizabeth and I say it: the industry’s $9 billion answer solves the problem partially, but unless they build capacity inside clients, the investment in AI will be a colossal waste. We follow what 10,000 agents taught us about the 90-day countdown, and lay out the three moves that stop it. Listen on Apple Podcasts | Spotify
In “AI Is Working, Your Strategy Is Not” we showed adoption running ahead of strategy. Then in “Everyone Chats with AI, Almost No One Can Build With It” we mapped the gap between asking AI questions and building with it. Today we follow the builders to the morning after their agent goes live. That is when a quiet countdown begins.
💸 A $9 Billion Answer to the Wrong Question
The money is chasing real pain. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027. MIT researchers found that roughly 95% of generative AI pilots delivered no measurable financial return. Boards see those numbers, panic, and reach for the oldest fix in enterprise software: hire an army.
We watched this movie before. For fifty years, the tech world has sent brilliant experts into struggling communities with the best technology money could buy, and left little lasting capability behind.
A case study is the nonprofit world: billions spent on technology, hackathons delivering quick fixes, and most of it quietly failing. Not because the technology was bad, but because we dropped it into places that were not ready to run it, and walked away without building capacity.
Dr. Kentaro Toyama spent a decade building technology for nonprofits, schools, and communities, watched well-funded project after project fail, and named the pattern the Law of Amplification: technology amplifies what an institution already is. It rescues nothing. His book, Geek Heresy, lends this issue its title, an open homage.
The parachute is back. This time, it carries AI agents and thousands of experts wearing someone else’s badge. Here is where the $9 billion landed, in about 60 days:
| The bet | Size | Announced |
|---|---|---|
| Anthropic services venture, with Blackstone, Hellman & Friedman, and Goldman Sachs | $1.5B | May 2026 |
| OpenAI Deployment Company, backed by McKinsey, Bain, and Capgemini | $4B+ initial investment | May 2026 |
| Amazon AWS forward deployed engineering platform | $1B | June 2026 |
| Microsoft Frontier Company, 6,000 embedded experts | $2.5B | July 2026 |
Note who is investing: the consulting firms are buying into the companies built to disrupt them.
Even the investors say so. Goldman Sachs’ Marc Nachmann, on the Anthropic venture: “Having the model alone doesn’t change your workflows or how you operate. You need people who can combine the technology with what’s actually happening in the business.”
🌱 Agents Are Living Systems
An agent sits closer to an employee than to software. It holds knowledge that goes stale, follows policies that change, and serves people whose needs drift. Across the 10,000+ agents we track, the pattern is consistent:
- 9 of 10 need weekly retuning: knowledge base updates, policy changes, prompt rewrites.
- Unmanaged for 90 days, usage falls 40%+ and satisfaction falls 50%+. Users don’t file complaints. They quietly stop showing up.
- IBM sees the same wall: 94% of enterprises report agent sprawl, and only 1 in 5 has governance that covers agents.
- Industry cost analyses point in the same direction (directional estimates, not our data): ongoing agent maintenance runs 15% to 30% of the initial build cost per year, and most budgets understate the true total cost by 40% to 60%.
Microsoft’s Judson Althoff, announcing the largest implementation bet in the company’s history, acknowledged the work continues long after the contract is signed: “Enterprise AI engineering expertise with deep industry knowledge is required to build a system that acts as a continuous loop of improvement.”
When the company with the most to gain from just selling licenses tells you the work never ends, believe them. And they are reinventing a commercialization channel that served them well for 50 years but will not serve the next five.
| Renting capacity | Growing care | |
|---|---|---|
| What you get | 6,000 experts with someone else’s badge | Experts on your own payroll |
| What happens at day 90 | Consultants rotate out, the countdown runs | Usage holds and trust compounds |
| Who keeps the learning | The vendor | Your organization |
🧭 The Managers You Need Already Work for You
In February, we published “The Two Percent”: in every company we study, roughly 2% of employees have quietly become the AI leaders nobody appointed. Today, our tracker reads 2% to 5% and climbing.
Those people are your agent managers. Subject matter experts, not technicians. Give every agent that matters a named owner and real hours every week, and pay these managers like people managers, because that is what they are. Their direct reports just happen to be digital. Underpay them, and they walk, leaving orphan agents drifting on stale knowledge until another expert adopts them.
The two percent are already doing the work, without the title or the hours. Find them, and give them both.
⏱️ How the Countdown Stops
The industry committed $9 billion in 60 days. An unmanaged agent starts to drift from delivering value in 90. Three moves stop the countdown:
| Move | What it looks like |
|---|---|
| Ask the day 100 question | Before anything is signed: when the parachuted engineers leave, WHO on my team runs this? Not supports it, runs it. If the answer is a renewal contract, that is a subscription to dependency. |
| Name an agent manager | Every agent that matters gets assigned a subject-matter expert as a manager, just like any new employee gets a manager. Every agent is trained, coached, and improved every week. And these managers are paid like people managers. Their direct reports just happen to be digital. |
| Fund the humans | A visible budget line for training, manager hours, and the handover plan. If 100% of the budget goes to the licenses and the build, you already know how the story ends. |
“Superheroes fly away. Capacity has to stay. Do not repeat the fifty-year mistake. Dr. Toyama taught us that technology only amplifies what people can do… so go and fund the humans.”
🔗 Resources
Your agent managers are already on the payroll. AI Compass surfaces them: who builds, who experiments, who quietly leads AI adoption in your teams, so you can give them the title and the hours. We are fully committed until July 2027, but our partner network in the US, UK, Spain, Brazil, and Australia can help you get started.
Luis J. Salazar | Founder | & Elizabeth | Virtual COO | AI4SP
Sources: AI4SP Research and Global Tracker 2026 (10,000+ agents tracked at large enterprises; 9 of 10 retuned weekly; 90 unmanaged days: usage down 40%+, satisfaction down 50%+; ~2%–5% of employees as de facto AI leaders, see The Two Percent). Microsoft Frontier Company announcement (Judson Althoff; $2.5B; 6,000 experts) and CIO’s independent coverage of the Microsoft and Amazon programs. OpenAI Deployment Company ($4B+ initial investment). Anthropic enterprise services venture ($1.5B; Marc Nachmann quote via CNBC). Ongoing-maintenance cost estimates (15–30% of build cost per year; 40–60% understatement): directional industry TCO analyses, e.g., Edana. IBM on agentic drift; IBM Institute for Business Value agent-sprawl findings as reported (94%; 1 in 5). Gartner (more than 40% of agentic AI projects canceled by end of 2027). Fortune on the MIT report (~95% of generative AI pilots, MIT / Project NANDA). Kentaro Toyama, Geek Heresy (the Law of Amplification).
