Stop Waiting for Better AI Models—Build Better Systems

The Model Stays the Same. Your Moat Doesn't.

When Salesforce's research team took an AI agent from 43.5% task completion to 93% without changing the model at all, they answered the question nobody wants to hear: the competitive advantage isn't in the AI you rent. It's in the infrastructure you already own but probably haven't instrumented.

I keep hearing the same question in every strategy session: "When does the next model come out?" As if GPT-5 or Claude Opus 4 will be the thing that finally makes AI work for regulated work. As if the frontier model is the missing piece.

Salesforce just proved it's the wrong question entirely.

The 50-Point Jump That Changes Everything

Their research team ran a simple experiment. They gave an AI agent real browser-based tasks — the click-through-forms, navigate-the-portal, pull-the-data work your operations team does every day. First pass: 43.5% completion rate. Unusable. You can't delegate anything to a system that fails more than half the time.

Then they got it to 93%.

Here's what they didn't do: they didn't wait for a smarter model. They didn't fine-tune weights or increase parameters. The model stayed frozen. Same engine, same intelligence, same API every other firm has access to.

What changed was everything around it. The prompts. The tools it could call. The workflow structure. The validation checks that catch it before it makes a consequential error. The operational scaffolding that turns a language model into a system you can actually trust with work that matters.

The model never got smarter. The system around it did.

The Part You Rent vs. The Part You Own

Sit with that result for a second, because the implication cuts against every "wait for the next model" roadmap I've reviewed this quarter.

The frontier model is a commodity you rent. You type an API key. Your competitor types the same API key. You both get Claude 3.5 or GPT-4o. There is no moat in the model itself — everyone is standing in the same river.

The workflows, the quality definitions, the permission structures, the evaluation frameworks that define "correct" for your specific domain — that's the part you already own. Nobody rents it out from under you. It's not on OpenAI's roadmap or Anthropic's feature list. It's the boring operational infrastructure you've built over decades to onboard people, catch errors, and maintain standards.

That infrastructure is suddenly the whole game.

Railroads, Switching Yards, and Ghost Towns

None of this is new. The railroads ran the exact same play 150 years ago.

When rail expanded across the United States, the track itself became a commodity fast. Same steel, same gauge, running to every town that wanted access. The track was cheap to lay and impossible to differentiate.

The towns that built switching yards, freight terminals, and maintenance depots won. The ones that just happened to be near the rails but didn't build infrastructure around the track? They emptied out. They thought proximity to the technology was enough. It never was.

Value never lived in the track. It lived in what you built on top of it — the operational infrastructure that made the commodity useful for actual work.

AI is following the same pattern. The model is the track. Your evaluation frameworks, your quality definitions, your workflow orchestration — that's the switching yard. And most firms are still acting like proximity to the rail line is a strategy.

The Work Nobody Wants to Do

Here's the uncomfortable part, and I'm watching clients hit this wall in real time.

Salesforce's agent improved because they had clean scoring. They could definitively say "this task succeeded" or "this task failed." Browser automation has that property: either the form submitted correctly or it didn't. Either the data extracted matches or it doesn't. Pass/fail is unambiguous.

Your work doesn't come with an answer key.

When does a financial reconciliation count as "correct"? When is a compliance review "complete"? What does "approved" actually mean when the rules are written in policy documents, oral tradition, and twelve years of email precedent?

The job isn't deploying AI. The job is turning your real tickets, your real approval workflows, your real error corrections into a scorecard the system can learn from. It's tedious infrastructure work. It's also the entire competitive advantage.

I was on a call last week with a CFO who wanted to automate month-end close. Great use case — repetitive, high-volume, painful. I asked him: "If I gave you ten completed closes and told you one was wrong, could you write down the test that catches it?"

Long pause.

"We'd know it when we see it."

That's the problem. "We'd know it when we see it" is not an instruction set. It's not a quality gate. It's not something you can compound on. And until you translate institutional knowledge into executable evaluation, the model plateau is guaranteed — regardless of what OpenAI ships next quarter.

Evaluation as Moat

The firms treating evaluation infrastructure as strategic are compounding. Every task the AI completes becomes training data for the next one. Every correction sharpens the quality definition. Every workflow instrumented makes the next workflow cheaper to automate.

The firms waiting for GPT-next are standing still. They're renting the same commodity everyone else rents, watching the leaderboard, hoping the next model will be smart enough that they don't have to do the hard boring work of defining "correct."

But what do I know — I've only watched this movie four times. Client-server, internet, mobile, cloud. The technology commoditizes. The infrastructure compounds. And the firms that mistake access for advantage end up as case studies in someone else's disruption keynote.

What to Do Monday Morning

Here's the test. Pick one workflow in your operation — not a moonshot, just something repetitive and painful. Now answer this:

Could you write down the criteria that define a successful completion? Not "we'd recognize it," not "our team knows," but actual testable criteria. If you handed me ten examples and told me one was wrong, could I find it without asking you?

If the answer is no, that's not the model's problem. That's yours. And it's the work that determines whether you're building a switching yard or just standing near the rails.

The model isn't getting 10x smarter next year. But the gap between firms that instrumented their quality definitions and firms that didn't? That gap is about to get very, very wide.

What's the one workflow in your shop where you couldn't cleanly define "correct" right now? Start there. Not because it's easy. Because it's the moat.