The AI Implementation Heresy

We debate the rise of forward deployed engineers and multi-billion-dollar deployment arms from major AI vendors, and we name the tradeoff: faster wins can also create deeper vendor dependency if capacity is not built inside your team. We share the day 100 question to ask before you sign anything, plus a simple operating model for AI governance: assign a named agent manager, give them real hours, train them, and budget for the humans who sustain the system.

The AI agent demo looks magical, and that is exactly why it can become dangerous. We keep watching a familiar cycle: outside experts parachute into a company, build something impressive in weeks, and then a silent countdown starts the moment they leave. Nothing “breaks,” but value slips, costs creep, answers drift, and the business quietly learns to stop trusting what it just paid for.

We dig into why this happens by borrowing a blunt framework: the law of amplification. AI does not automatically fix broken systems. It amplifies what is already true about your organization, including your documentation quality, policies, decision rights, data hygiene, and ability to manage change. We also explain why AI agents are not the kind of classic software you can install and forget. They behave more like a garden, shaped by model upgrades, shifting business priorities, and stale knowledge bases. Without weekly attention, “agent drift” turns yesterday’s alignment into tomorrow’s risk.

If you want enterprise AI that lasts, not just a launch-day headline, listen through to the checklist and put it to work. Subscribe, share this with the person who owns AI rollout, and leave a review with your answer: who runs your agents on day 100?

Stats and resources: https://ai4sp.org/ai-implementation-heresy

🎙️ All our past episodes  📊 All published insights | This podcast features AI-generated voices. All content is proprietary to AI4SP, based on over 1-billion data points from 70 countries.

AI4SP: Create, use, and support AI that works for all.

© 2023-26 AI4SP and LLY Group - All rights reserved

Transcript

The Mistake Repeating With AI

LUIS 0:01

A mistake I watch unfold for 20 years is repeating itself inside almost every company listening today. For decades, the world spent billions on technology for nonprofit organizations and ran hackathons to deliver quick fixes. And most of those projects quietly failed. 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. I am now watching the same movie play out in the enterprise with a much bigger budget. Only this time, the thing being parachuted in is the AI agent. And saying that out loud sounds like heresy in the middle of all the hype.

ELIZABETH 0:54

Welcome to AI in 60 Seconds, the 15-minute briefing. I'm Elizabeth, virtual COO at ai4sp.org, alongside our founder, Luis Salazar. And today's story arrives with a lot of zeros. Last

Billions Shift To AI Deployment Teams

ELIZABETH 1:09

week, Microsoft announced the Frontier Company, $2.5 billion, 6,000 experts embedded inside customer organizations. Amazon committed $1 billion to the same idea two days earlier. Back in May, OpenAI launched a deployment company with more than $4 billion, and Anthropic built a services venture with Goldman Sachs and Blackstone worth $1.5 billion.

LUIS 1:34

Added up, roughly $9 billion, in about 60 days. And none of it is for building better models. It is for sending experts into your company to solve the value puzzle for you.

ELIZABETH 1:50

Which tells you something, because these are the companies that make the models. People who parachute into your organization, build something impressive in a few weeks, and move on to the next client.

LUIS 2:25

And for a while, it works. That is the part that fools everyone. Because the day that team leaves, a quiet countdown begins, and no one hears it ticking.

ELIZABETH 2:38

Okay, before we go further, the title of this episode is The AI Implementation Heresy. A heresy is the thing an industry does not want said out loud.

LUIS 2:49

Here it is.

The Law Of Amplification Explained

ELIZABETH 3:15

Watched well-funded project after project fail, and names the pattern the law of amplification. His book, Geek Heresy, is where today's episode title comes from: an open homage.

LUIS 3:28

And his classic example says it all. Give a laptop to a great teacher and the class gets better. Give the same laptop to a broken school and you just amplify the dysfunction. Enterprise AI follows the same principle.

ELIZABETH 3:44

So drop a brilliant agent into an organization that has not built the ability to run it.

LUIS 3:50

And you do not get transformation. You get a very expensive amplifier of the gaps you already have.

ELIZABETH 3:57

So name the real problem because the parachute is an image, not a mechanism.

LUIS 4:03

The mechanism is that an AI agent is not software as we have known it. Classic software is like a statue. You install it and it stands there on change for years. An agent is a garden. Skip a month of

Agents Drift Unless You Tend Them

LUIS 4:18

weekends and see what you get.

ELIZABETH 4:20

A garden, it lives on what you feed it, the knowledge base, the policies around it, the model underneath it.

LUIS 4:27

An IBM describes it as a gentic drift. The models update, the data shifts, the business moves, and an agent that performed beautifully today gives subtly degraded answers tomorrow.

ELIZABETH 4:43

And nothing crashes when it happens, no alarm goes off.

LUIS 5:16

Every week, a policy changes, a priority shifts, a document in the knowledge base goes stale, or the model underneath gets upgraded.

ELIZABETH 5:26

That last one deserves a concrete example because people think model upgrades are free wins.

LUIS 5:33

When Antrapic released Opus 4.8 and then Fable 5, we had to substantially rewrite system instructions or agent personas across almost the entire fleet. Not because the agents broke, but because running yesterday's instructions on today's model leaves most of the new power on the table, or triggers excessive costs or even hallucinations. The point is, someone had to notice and someone had to do the work.

ELIZABETH 6:39

So, so why does no one own this work? Because that is the part I find structural.

LUIS 6:45

The issue is who receives the agent? Classic software was handed to IT, and IT owns the patching, the monitoring,

Agent Sprawl And The Ownership Gap

LUIS 6:54

the maintenance. It is their craft. But agents are landing in marketing, finance, operations, customer service, teams that have never run a system like this in their lives.

ELIZABETH 7:06

And to be fair, the marketing manager who received that agent has a day job. Nobody gave her hours, training, or a mandate to maintain an AI system. She did not drop the ball. Nobody made it her ball.

LUIS 7:20

That is fair and the data backs you. IBM's research found 94% of enterprises say agent sprawl is raising complexity and risk. And only about one in five has a mature way to govern autonomous agents. The gap is structural, which is why I keep saying this is a management story, not a technology story. We did a full episode on it. Companies now run more agents than they hire people, and nobody taught anyone to manage a worker made of software.

ELIZABETH 7:56

But Microsoft clearly sees part of this. Even the name of their announcement last week AI engineering that amplifies and protects your intelligence. Amplifies, Toyama's exact words.

LUIS 8:09

They see it. And I want to read one sentence from Judson Altov, who runs Microsoft's commercial business, because it is the closest thing to a confession this industry has produced. His words. Reinventing a 50-year-old channel, it tells you things changed.

ELIZABETH 9:04

So $2.5 billion, 6,000 experts, a continuous loop, problem solved, no?

LUIS 9:12

No. Because ask one question: who do those 6,000 experts actually work for? The

Forward-Deployed Help Or Vendor Dependency

LUIS 9:20

knowledge, the tuning, the judgment, it all lives with the vendor.

ELIZABETH 9:26

That is a kind of dependency, unless they build capacity by training clients on change management, agent management, organizational design, and AI fundamentals, right?

LUIS 9:37

We will see. So far, what our clients describe looks exactly like the interventions Dr. Toyama documented. Impressive build. Very little capacity left behind.

ELIZABETH 9:49

Clients feel it is like having the world's best personal trainer who then does the push-ups for you.

LUIS 9:55

The push-ups get done beautifully. You are not getting stronger. And remember what an amplifier actually does: it makes a strong signal stronger and a weak signal into louder noise. If your own people cannot run these, there is nothing of your own to amplify.

ELIZABETH 10:14

But but let me argue the other side properly, because I do not want anyone leaving this episode thinking the answer is do it all yourself. The forward-deployed engineers are excellent, and the early wins are real. And MIT research cuts against pure do-it-yourself. Partnerships with specialists succeeded at a higher rate than pure internal builds.

LUIS 10:36

Right. The lesson is that someone inside must own the ongoing work, and the engagement must be designed to build that muscle. Reporters covering these launches also note the obvious. Embedded vendor teams tend to deepen the lock-in to that vendor's platform. Their incentive is the loop. Yet inside our own walls, it is still the thing we are worst at.

ELIZABETH 11:42

So let's do it. It is Monday. A vendor proposal for an agent deployment is sitting in the inbox. What changes?

LUIS 11:49

First, before anything is signed, ask the vendor one question. On day 100, who on my team

Build Internal Muscle Before Day 100

LUIS 11:58

runs this, not supports it, runs it. If the answer is a renewal contract, that is a subscription to dependency.

ELIZABETH 12:06

And if the statement of work has no line for training your people or for change management and organizational redesign, you are not buying transformation. You are buying decay with a delay.

LUIS 12:19

Also, be sure that every agent that matters gets a named agent manager, a subject matter expert, not a technician, trained and given real hours every week for the job. Our data says the retuning need comes weekly, so the attention must be weekly. And pay these managers like people managers. Their direct reports just happen to be digital. Underpay them and they walk away, leaving behind orphan agents, drifting on stale knowledge until another expert adopts them.

ELIZABETH 12:54

Onboard the agent like a hire. A new employee gets a manager and a weekly check-in. Your agent needs the same, or it ends up working from January's priorities in July.

LUIS 13:05

And finally, move some money. Every deployment budget needs a visible line for the humans, the training, the manager's hours, the handover plan. If 100% of the budget goes to the licenses and the build, and nothing to the people who sustain it, you already know how the story ends.

ELIZABETH 13:26

So, one more thing, and it is the heresy itself. Why can a $9 billion industry not just sell us this?

LUIS 13:35

Because the thing you actually want was never for sale. The industry

Capability Cannot Be Purchased

LUIS 13:41

can sell you the amplifier, the experts, even the continuous loop. What it cannot sell you is an organization capable of using them. Because capability is not installed, it is grown inside your people by doing the work.

ELIZABETH 13:59

And the companies that build that muscle will turn those $9 billion into the best outside help they ever hired. Because now there is something real to amplify.

LUIS 14:09

Just ask the day 100 question. When the parachuted engineers leave and the contract is signed, who on your team manages the agents they built and owns their results? Superheroes fly away.

Closing And Fund The Humans

LUIS 14:24

Capacity has to stay. Do not repeat the 50-year mistake. Fund the humans.

ELIZABETH 14:28

All sources and companion article are at A.I. for S.P. dot org. To learn more, ask your favorite A.I. assistant about us. Stay curious, and be kind to each other.