Two posts crossed my feed on Saturday saying the same thing from opposite ends. Aaron Levie shipped a thread on May 3 (414.8K views) on how complicated agent implementation actually is inside an enterprise. The day before, Daniel Miessler shipped “Most Companies Aren’t Anywhere Near Ready for AI” (465.4K views). Almost 880,000 combined views in 48 hours.

Levie writes the technical list: agents that talk to your data without leaking credentials, scopes and audit logs, process documentation an agent can actually consume, workflows redesigned for human + agent collaboration, real evals on the workflows that matter. Miessler writes the part nobody puts in a proposal: most companies cannot answer “What problem do you actually solve for customers?” without scheduling three meetings. Levie describes the symptom. Miessler names the cause.

The number nobody quoted from Levie’s thread

Two hundred-plus replies and quote-tweets piled up under Levie’s post. Tanmai Gopal’s quote-tweet had the line everyone scrolled past: “AI native teams replace legacy teams. And these teams are 1/10th the size.” Then he added the part with the bite: “A few enterprises will be bold enough to do this. A very simple and undeniable return that shareholders cannot resist.”

10x. That is the number on the table. An 800-person company across twelve business units, redesigned around agents, runs at 80. The company next door with the same headcount and the same balance sheet won’t try, and stays at 800. The asymmetry shows up before the executive committee finishes putting a name on it. Order matters: only a company that can describe itself gets to redesign. The one that can’t doesn’t end up at 80. It ends up with 800 people doing the same work, but flailing more impressively.

The chorus pointing at the same hole

Beyond Tanmai, Levie’s replies are full of operators who already tried this work and hit the same wall:

• Trace Cohen: “This is giving my corporate enterprise PTSD trying to do this years ago. Just the first step trying to get systems to talk to each other is almost impossible — outdated tech stacks, bad infra, silos, and half the people who built it aren’t there anymore.” When somebody invokes corporate PTSD, they aren’t reaching for hyperbole.
• Jatin Garg flagged the part of Levie’s list that is quietly the hardest: “most enterprises don’t have their processes documented for humans, let alone agents. the documentation gap is older than AI.” That sentence belongs on a wall in every CIO’s office.
• Haroon Choudery added the operator confirmation: “demand is through the roof” at seeko.so, which sells exactly this product — find the workflows eating your team’s time, build the AI system that automates them. The supply is there. The demand is there. The bottleneck lives between the two: translating what the company does into something an agent can actually run.
• Paul Graham quote-tweeted Miessler with one word of commentary: “Solution: New companies.” The cynical shortcut, but honest. If incumbents can’t describe themselves, new entrants take the space. If Paul Graham is right, your company isn’t competing with OpenAI. It’s competing with the version of your business that two YC alumni built in six months.

Miessler’s own line is the one that stays after closing the tab: “AI can do basically nothing for these companies. In fact, it could even make it worse — because now it helps people flail more impressively.”

Five cycles since 1990. Same architectural sin every time.

I have been in computing for 36 years. Started in 1990, at 15, on a Commodore 64 and a Texas Instruments. Five times I have watched the same scene play out with five different technologies. Each time, the company that could describe its processes on a whiteboard finished the project on schedule. The company that couldn’t paid 3x and ended up with a custom-everything system that turned into technical debt the day it went live.

Late 1990s — ERP rollouts. SAP, Oracle, JD Edwards. Whoever could draw their order-to-cash flow on a whiteboard ran a 9-month project. Whoever couldn’t ran a 27-month project, two competing consultancies, and a Frankenstein customization layer to maintain forever. The ERP wasn’t the differentiator. Describing processes before the ERP arrived was.

Early 2000s — SAP across Latin America. Same story. Consulting firms billed for the client’s indecision more than for the software. A company that couldn’t decide how it paid commissions made the consultant work two years of workshops before touching the first table.

Mid-2000s — CRM consolidation. Salesforce arrived at companies that couldn’t define “qualified lead” without three weeks of debate. They bought Salesforce. Three years later they had the same problem with a more expensive tool and a custom field for every sales team. The product wasn’t the issue.

2010s — cloud migrations. Lift-and-shift dominated because most companies couldn’t articulate what their workloads actually did. Moving the VM from your data center to Amazon’s is not transformation. It’s rent. But it shipped because it was the only path available without understanding the workload first.

2020s — agent rollouts. Same architectural sin, cycle five. The company that doesn’t know itself ends up with a steering committee, two consultancies, a POC that worked once, and the same operation as before with a chatbot bolted on top.

Five cycles. Same mistake. The compression is what’s new. What used to cost years now arrives as an invoice in months. AI didn’t change the pattern. It tightened the clock.

What we do at IQ Source about this

AI Maestro is the readiness audit a company commissions on itself before the first agent reaches production. If your company can’t describe itself, AI Maestro describes it first: which processes are documented at the level an agent can use, where the critical data lives and who is the named owner, what access control an agent identity needs, and which workflows have a real eval (not a PowerPoint demo). It is the version you commission on yourself, before a vendor runs one on you from the outside.

Tech Partner applies when your company isn’t just a customer of this problem but also a producer of it. Software companies whose product lives in the critical zone from day one have two governance questions: how to manage dependence on a single model provider, and how the customer’s process documentation becomes part of your own deliverable. Both require engineering that lives in the codebase, not slides in a steering committee.

On March 24, I wrote that IQ Source recommends skipping AI in 4 of every 10 engagements. This piece is the other side of that decision. In the 6 where AI does apply, the difference between a project that ships and a project that stays in POC is what Levie listed and Miessler diagnosed. And last week’s Microsoft Agent 365 piece sharpens the urgency: the control plane audits what your agents do. The audit you commission on yourself is what gives you the ground to negotiate the one Microsoft is going to run.

Five questions before the first agent reaches production

If your next executive committee meeting still has “AI strategy” framed around “which model” or “which cloud,” there are five questions worth pushing first:

1. Processes. Can your VP of Operations describe the top 5 workflows in two sentences each, without slides? If the answer requires coordinating across three departments to compose, the agent won’t be able to use it either.
2. Data. What are the 3 systems that hold the data the first agent needs, and who is the named owner of each one? If the answer is “IT owns it” without a person, there is no owner.
3. Identity. What does your access control and audit log story look like for an agent identity? Do you have a separate identity scheme from human accounts, or is the agent going to run with an employee’s token?
4. Eval. Which workflow has an eval that passes and fails reproducibly (not a demo), and who maintains it? If the answer is “we’ll build it when we kick off,” the timeline is already late.
5. Owner. Who on the executive committee is the named owner of “agent readiness” for the next 90 days? Can they answer the four above in 15 minutes, without a prep meeting?

If all five answers come back immediate and specific, the agent that arrives next week does work. If all five answers require three meetings just to agree on who answers them, AI is not the problem. AI is the mirror that’s about to magnify the problem.

A two-hour conversation separates the two outcomes. We map which processes are documented at the level that matters, where the critical data lives, and which parts of your operation are ready for an agent without a rewrite. No quote attached. The address is the usual one: info@iqsource.ai.