The hottest idea in GTM software right now is nine years old. And we're still getting the important part wrong.
The idea: software is moving "from system of record to system of intelligence." The CRM stops being the prize. The reasoning layer on top of it, the thing that pulls your context, decides, and acts, becomes where the value lives. The team at a16z made the case sharply, and they're right. I've been arguing a version of it to anyone who'll sit still long enough.
But Jerry Chen already wrote this. The New Moats, 2017: the durable moat is the intelligence layer on top of the system of record, not the database underneath. Same firm, even. So before we treat it as a 2026 epiphany, sit with the fact that the insight is old. What's new is that it's finally buildable. And what's still broken is the part nobody wants to talk about.
Everyone is describing the system of intelligence. Almost nobody is describing what it needs to actually work.
Here's the uncomfortable version. Most "systems of intelligence" being built right now are gorgeous at orchestration and have the memory of a goldfish. They wire up a dozen tools, fire off an agent, produce something that demos beautifully, and forget all of it by morning. That isn't intelligence. That's automated forgetting at scale.
Why it's finally buildable (and why that's the trap)
The plumbing exists now. The Model Context Protocol went from roughly 100,000 to 97 million installs in 16 months and got handed to the Linux Foundation with OpenAI, Google, and Microsoft all lining up behind it. For the first time, the orchestration layer is buildable by mortals.
Which is exactly the problem. When the hard part gets easy, everyone rushes to do the easy part. Orchestration is becoming table stakes. a16z calls it "the new gravity well." Close. But gravity doesn't come from connecting things. It comes from what compounds.
Let me be specific about what a real system of intelligence needs, because "reasoning layer" is doing a lot of unpaid labor in these essays.
1. Memory, or you've just automated forgetting
The system of record has exactly one job: remember things. It is quietly terrible at it. 76% of teams say less than half their CRM data is accurate and complete, and 45% say it isn't ready for AI at all. So the foundation we're stacking our brilliant reasoning layer on is already rotting.
Now make it worse. Most agents have no memory of their own. Each one starts from a blank window, does its thing, and remembers none of it. The context your best rep spent three years building? It walks out the door with them. Gallup pegged voluntary turnover at more than $1 trillion a year, and that number doesn't even count the institutional knowledge that leaves with the person.
An agent that orchestrates beautifully and remembers nothing is a very expensive goldfish with API access. Memory isn't a feature of the system of intelligence. It's the precondition for one.
2. Governance is the product, not the speed bump
Here's where the demos go quiet. An intelligence layer is only worth anything if it can act. Write to the CRM. Send the email. Move the forecast. And the second it can act, it can act wrong. At scale. Unsupervised.
This is why Gartner expects 40% of agentic AI projects to be abandoned by 2027, largely on cost and governance. Not because the agents were dumb. Because nobody could trust them to touch anything that mattered.
We treat governance like friction. The approval queue, the audit trail, the permissions. The boring stuff that slows the demo down. That's backwards. Governance is the thing that turns a clever agent into a system you'll actually let near your pipeline. The boring stuff is the product.
3. Compounding is the only moat that ages well
This is the one everyone skips, and it's the whole game.
A real system of intelligence runs a loop. It reads from shared context. It runs a governed action. It writes the result back. Then it learns from what actually happened. Read, run, write, learn. Skip that last step and you don't have a system of intelligence. You have a very confident intern who reintroduces himself every single morning.
The moat was never the database. It was never even the orchestration. The moat is the loop. Every approval, every correction, every deal you win or lose makes the next action a little sharper. Miss the loop and you aren't building an asset. You're renting intelligence by the day and paying full freight every morning.
So what do you actually do
If you're buying or building in this space, stop grading the demo. The demo is orchestration, and orchestration is solved. Ask the three questions that separate a system of intelligence from a parlor trick:
What does it remember tomorrow that it learned today?
Who approved the last thing it changed, and can you prove it?
Does it get smarter from outcomes, or just faster at the same mistakes?
If the answers are "nothing," "no idea," and "faster at the mistakes," you're not looking at a system of intelligence. You're looking at a system of record wearing a smarter hat.
That gap (memory, governance, the loop) is the entire reason we started Wysdym. We're building the harness, not another agent. The layer underneath that every agent reads from, runs on, and learns through. We're early. Pre-revenue, building it with a small group of design partners who'd rather shape the foundation than buy the fortieth tool that demos well and dies in production.
a16z is right that the value is moving up the stack. But up the stack is also where it gets hard. The system of record was where your data went to sit. A real system of intelligence is where everything your team learns finally stays.
Anything less is just a database with better manners.