TL;DR
- AI no longer needs to be told how to do a job, only how to tell whether the job was done well.
- That moves the value from whoever knows the steps to whoever holds the record of outcomes.
- Moative helps companies find that record and turn it into the thing AI gets graded against.
The value in AI has moved from knowing how work is done to knowing whether it was done well. Three consequences follow.
- Building AI gets cheap.
- Vertical AI startups that sold workflows get expensive to keep alive.
- And, the dullest data in your company, the record of what went right and wrong, becomes the most valuable thing you own.
Building got cheap
Until recently, building AI for a business process meant writing the process down: forty steps, twelve exceptions, one engineer quietly updating his LinkedIn that the is now a forward deployed engineer.
Now you write two things.
A goal: get the claim paid.
A grader: was it paid?
The model tries, gets scored, keeps what worked, and tries again. It’s called hillclimbing. It’s how evolution works, except evolution took four billion years and produced the platypus. This takes shorter time than the time between two tweets from Musk.
It works best when your score doesn’t depend on anyone else’s. A paid claim is paid no matter what your competitor does. It works worst when everyone competes for the same handful of slots. When every company climbs the same hill with the same model, they all reach the top together, which is less a summit than a crowded elevator.
If your goal is to reduce cost for an internal process, hillclimbing makes it work. If your job is to rank your website on google and answer engines, hillclimbing makes it a treadmill instead. Everyone climbs and the summit becomes the Subway station after it rains.
Workflow startups got expensive
A great many vertical AI companies are a frontier model, a documented workflow, and a website with a photo of a nurse who looks as happy as Ryan Hamilon. Look him up.
The workflow was the product. It took AI vendors six months to map. A model with a decent grader finds it before lunch and doesn’t need a new statement of work to be signed. That sequence was rarely as intricate as the price tag suggested, a dynamic we detailed regarding medical claims processing in AI for Complicated is not AI for Complex.
They’re pressed from above by the labs, which move up the stack every quarter, and from the side by data companies whose entire business is defining what good work looks like across every profession. Either one could go direct the moment it gets bored.
Death is inevitable for shallow vertical AI. Being the last one to raise a bridge round is optional.
The dull data got valuable
If the grader decides everything, the grader’s raw material is the asset: years of outcomes. Which claims were denied and why. Which loans defaulted. Which shipments came back and whose fault it was.
Clicks won’t do. Usage data tells you what people did. Outcome data tells you whether it worked, and only the second can grade a machine.
Most companies have not documented this data or it is maintained by someone who has never once been invited to a strategy offsite.
What enterprises should do
- Count your outcomes before you count your vendors. Denials, defaults, returns, rework, write-offs. That inventory is your AI strategy. Everything after it is purchasing.
- Ask every vendor how they measure good, and whose data they measure it with. If the answer is a workflow diagram, you’re buying last year’s product at this year’s price.
- Keep your data AND the labels. Your outcome data trains whatever model it touches. Contract for it the way you’d contract for your customer list, because to a vendor that’s exactly what it is. This is where I will pitch Moative to you but I will take the high road.
- Turn your experts into graders. Your best people stop doing the work and start judging it, which is what most senior people believed they were doing anyway.
- Audit your evaluation criteria. Evaluators mirror the metrics assigned to them, a failure mode analyzed alongside target gaming in Goodhart’s Law & Wrong Measures for Agent Outputs, making impartial verification necessary prior to deployment.
At Moative, we find the outcome data a company already has, turn it into a grader, and point frontier models at the work that moves money: denials, underwriting, operations.. If your most valuable asset is a spreadsheet nobody wants to own, we’d like to meet it. Take this AI readiness assessment to get started.
Frequently asked questions
What is hill-climbing in AI?
Hill-climbing is a way of building AI where you give the model a goal and a grader instead of step-by-step instructions. The model tries the work, gets scored, keeps what worked and tries again. For a claim, the goal is to get it paid and the grader checks whether it was paid.
What is outcome data, and how is it different from usage data?
Outcome data is the record of whether work actually succeeded: which claims were denied and why, which loans defaulted, which shipments came back and whose fault it was. Usage data, like clicks, only shows what people did. Only outcome data tells you whether the work was done well, so only outcome data can grade an AI model.
When does hill-climbing work, and when does it fail?
It works best when your score does not depend on anyone else’s. A paid claim is paid no matter what a competitor does, so cutting the cost of an internal process is a strong fit. It works worst when everyone competes for the same few slots, such as rankings on Google and answer engines, because everyone climbs the same hill with the same model and reaches the top together.
Why are vertical AI startups at risk?
Many vertical AI companies are a frontier model plus a documented workflow, and the workflow is the product. A model with a decent grader can find that workflow quickly, without a new statement of work. These companies are also squeezed by AI labs moving up the stack and by data companies whose business is defining what good work looks like across professions.
What should I ask an AI vendor before signing?
Ask how they measure good work and whose data they measure it with. If the answer is a workflow diagram, you are buying last year’s product at this year’s price. Contract for your outcome data and its labels the way you would contract for your customer list, because that data trains whatever model it touches.
Where should a company start with its AI strategy?
Start by counting your outcomes: denials, defaults, returns, rework and write-offs. That inventory is your AI strategy, and choosing vendors comes after it. Then keep your data and the labels, and turn your best experts into graders who judge the work instead of doing it.