“Just like every business today has an email address and a website and a social media presence, in the future, every business is going to have an AI agent that lives in the different messaging platforms.”

That’s Mark Zuckerberg at Stripe Sessions 2025, in conversation with John Collison. Verbatim. Not the version that got paraphrased around X this morning — that one swapped “email and social presence” for “website and phone number.” The original is more honest about which channel Meta is trying to own (messaging — WhatsApp, Messenger, Instagram DMs), and generalizable enough to land as a deadline for anyone selling B2B.

Aakash Gupta amplified it this morning with the historical parallel that saves you three strategy meetings: the AI agent is the 2005 website. Not a prediction. A countdown.

I’ve been in computers and business for 36 years. Since I was 15. I’ve watched this kind of filter appear twice before. It didn’t have a press release either time. It didn’t have a cutoff date. It showed up as a quiet filter and closed businesses that didn’t realize they were closing until about six months after the fact.

Wave 1: the website (1995–2010)

The first time I heard the phrase “my customers know where to find me” was around 2001. A lot of business owners said it. In the short term they were right. Their existing customers did know.

What they missed was that new customers stopped using the Yellow Pages around 2005, and word-of-mouth migrated to reviews and search results. The shape of the business didn’t change overnight. The shape of the top of the funnel did. Between 2005 and 2010, I watched 20-year-old businesses fold. Not because a competitor made a better product. Because a competitor showed up in Google, answered its reviews, and could be compared against three alternatives in the 90 seconds a prospect was willing to spend looking.

The uncomfortable detail: in 2003, saying “I don’t need a website, my customers find me” was a defensible position. By 2008 that same sentence was a clinical symptom — the kind of thing an owner says the quarter before they can’t cover payroll. Five years between “reasonable opinion” and “warning sign.” Nobody marked the calendar.

Wave 2: mobile and search (2008–2015)

The second wave was shorter and more brutal. Businesses that survived Wave 1 with a desktop-only site written in Flash ran into a problem in 2012: customers stopped opening their site on a computer. They opened it on a phone while standing on a bus, and that was the new default. The site looked broken or took twelve seconds to load. Google adjusted its algorithm to demote mobile-hostile sites. Page 2 of Google is the best-kept graveyard on the internet.

Businesses that didn’t move to responsive design plus local search disappeared from the literal map — Google Maps. Same pattern as Wave 1. No ceremony. Fewer customers coming in for about eighteen months, until payroll stopped clearing.

Two waves, same shape. Invisible filter. Causes assigned retroactively (“the market shifted,” “the pandemic,” “rising costs”). The real cause was simpler: the business didn’t cross an infrastructure threshold that was optional in year X and mandatory in year X+5.

Wave 3: the agent that answers (2024–)

We’re inside the third wave. This isn’t a thesis; it’s an inventory.

Gartner projected in August 2025, via senior director analyst Anushree Verma, that 40% of enterprise applications will feature task-specific AI agents by 2026, up from less than 5% at the close of 2025 (UC Today coverage). That’s not a roadmap — it’s happening in real time. Salesforce Agentforce, HubSpot Breeze, Microsoft Copilot Studio, Google Agentspace. When your customer’s SaaS stack ships six default agents and yours doesn’t respond, the filter is already running.

Sathish Harry summarized the pattern in Aakash’s thread as three rows:

• 1995–2005: website.
• 2008–2015: mobile + social.
• 2025–2030: AI agent.

Doesn’t have to be right to the year. Has to be right in shape. The barrier looked optional at the start of each wave. Then early adopters quietly closed the door on the late ones, and the failures that followed got written up as market shifts, rising costs, or anything except the threshold the business never crossed.

The Tuesday 11pm scenario

A customer emails your company at 11pm on a Tuesday. Your team sees it at 9am on Wednesday. Ten hours of silence.

Your competitor’s agent answered in two seconds. It attached the contract PDF to a quote and dropped a Thursday sales call on the prospect’s calendar before your inbox had even pinged. By the time your rep opens it on Wednesday morning, the customer is already signed somewhere else.

Blake Heron left a sharp comment under Aakash’s post: “the 11pm email example is the one that’ll convert people. abstract arguments don’t move founders, specific scenarios do.” He’s right. A founder or CTO doesn’t rearrange a roadmap because of a Zuckerberg thesis. They rearrange it when they see a lead they paid $200 for on Google Ads walk away because their agent didn’t exist.

The piece that changes the math: your business scales with payroll, while theirs scales with compute. A GPU loses roughly 40% of its price per year. Payroll doesn’t. Compound that delta across three years and it’s the difference between a business that can cut prices and one that has to raise them.

What people will tell you is an overreaction

A good contrarian in the same thread is Eshan (@eshanbuilds): “every business will use AI the way every business uses electricity. not as a differentiator but as infrastructure.” He’s right on the part that matters. You won’t win customers by “having AI” any more than anyone won customers in 2010 by “having a website.” You’ll lose customers by not having one. The upside is floor, not ceiling.

Another sharp pushback is Kobe Theodore (@Alphainvest20): “AI agents have payroll too.” True, up to a point. A Claude, GPT or Gemini token costs money. Inference at scale costs money. But the correct comparison isn’t “agent vs. free.” It’s “agent at $0.10–$1.00 per resolved conversation” vs. “human at $15–$25 fully loaded per hour, with sick days, turnover and shift coverage.” A well-designed agent runs 24/7 without asking for a raise. Agent payroll is real. It’s not the same order of magnitude.

The argument that isn’t valid: “my business is too specialized for AI.” That was the exact line in 2003 against websites. And in 2012 against mobile. Every single time, it turned out to be the most reliable signal that the owner didn’t know which part of the business was the actual bottleneck.

The moat moves to legibility

This is where the conversation stops being “get an agent” and starts being a question about how your operation is organized.

Dharmesh Shah is HubSpot’s CTO and co-founder. For months he’s been writing about what he calls “agent-readable infrastructure” and AUX (agent user experience). The thesis is straightforward. If your operation isn’t consumable by an agent (documented APIs, structured data, granular permissions, written-down processes), putting an agent on top of it just automates the chaos. It doesn’t reduce it.

TheDiaw, in the thread, quotes Jack Dorsey and Alex Hormozi: “The next moat will be an AI trained on a company’s unique data.” The agent code itself becomes a commodity in six months. Your operational data is the slow-cooking half. Knowing which specific contract clause your top customer always pushes back on is a data point that took six years of sales calls to accumulate, and you don’t buy it in a marketplace. The same applies to your effective discount curve, to the recurring onboarding breakpoint that only reveals itself on customer number 40, and to every other operational pattern that only becomes visible after you’ve operated.

The moat isn’t “we have an agent.” It’s “our operation is written in a way an agent can actually run it.” That ties to what we published yesterday: Benioff admitting the API is the UI. Clean APIs and documented workflows are the 2026 equivalent of a responsive site in 2013.

What we do about this at IQ Source

We stopped selling “implement this CRM module” and started building agents that answer on behalf of the client. That’s not a cosmetic pivot. It’s a direct application of one idea: if the business owner doesn’t have an internal team to stand up an agent, they need a partner to build the agent-readable layer first — and operate the agent on top of it.

Our AI Operations service starts with an audit of which of your workflows are already agent-ready. With that baseline, we build out the API and data layer needed to make the rest consumable. Then we run the agent as an ongoing service, not a handoff. An agent needs maintenance the same way an employee does. Nobody sells “permanent employee” as a fixed-price project.

For software companies that already ship SaaS and want to add an agent-native layer on top of their own product, we work as a technology partner. We don’t charge per seat. We charge per unit of outcome operated, as we wrote in the Headless 360 post.

What we don’t do: sell an agent as a shrink-wrapped product. The shrink-wrappable part is the least valuable. The valuable part is spending the first 90 days learning your industry’s workflow, writing it down, and only then building the agent. That order matters.

Your invisible deadline

There’s no announced date. There never is with this kind of filter. But there are signals that tell you whether you’re already on the wrong side of the one that’s running.

The easiest place to look is the inbox, because the data is already there. When average first-response time is lengthening instead of shortening, a competitor has probably already deployed an agent, and your human team is now carrying the same volume plus the friction of fighting new tooling at the same time. A tired human doesn’t compete with a rested agent in the same league, regardless of how good the human is.

Onboarding tells you a second, quieter version of the same story. When a new hire takes three months to learn the workflow because the knowledge lives in the senior human’s head instead of a readable file, an agent will take the same three months for the same reason. The AI is ready. The documentation isn’t. That’s the real barrier to automating, and it’s solved with a text editor, not a model vendor.

Pricing tends to react last because it hits revenue directly, which makes it the hardest signal to act on. If your billing is still anchored to seats, hours, or resolved cases, you’re pricing against a unit the market just left behind. It’s the same per-seat model Benioff already admitted was obsolete when he exposed Salesforce as headless.

If any one of those applies to your company, the question isn’t “Do I need to adopt AI?” It’s “Who helps me close the door before the filter starts affecting my funnel?” You already have the public language from Zuckerberg; the historical parallel is courtesy of Aakash, and Gartner already published the calendar. What’s still missing is the decision on who executes it with you.

If you’re in a business where at least 30% of new leads come in outside working hours, reach out. We won’t try to sell you an agent. We’ll help you read your own inbound numbers and decide whether the filter is already hitting your P&L or whether you still have six months of lead time. 36 years of watching these filters form makes that read quick.