Turning frontier capability into finished work.
Models are getting smarter faster than work is changing. We build the systems in between, starting with the service industries that keep insurance running.
Abstract
Frontier models are improving faster than organizations can absorb them. The result is a widening capability overhang: the distance between what models can do and what actually reaches the people doing the work. We believe the next two years will be decided less by better models than by who redesigns work around them.
The effect will be sharpest in service industries that sell commoditized products, where the product is the same everywhere and service is both the largest cost and the only real differentiator.
Keel builds the harnesses, evaluations and managed systems that turn model capability into finished work: verifiable, proprietary to each firm, and held to a standard of taste. We are starting in insurance.
The capability overhang
Model progress is widely discussed and still underestimated. The length of task that frontier models can complete on their own, measured in the time a skilled person would need, has doubled roughly every seven months.1 The price of a fixed level of capability has fallen by about an order of magnitude a year.2
Adoption has not kept pace. In most service organizations, AI arrives as a subscription: a chat panel beside email and documents, used to summarize a thread or soften a paragraph.
That is a failure of the interface, not of the workers. A chat window asks every person to supply their own context, write their own prompts and check their own outputs, for every task, from scratch. So it gets used where all of that is cheapest, which is where it matters least.
The overhang is not a shortage of intelligence. It is a shortage of systems: the context, tools, checks and delivery that turn a capable model into work someone will put their name on.
We’ve seen this before
The ATM
When ATMs spread in the 1970s and ’80s, everyone expected bank tellers to disappear. For decades, the opposite happened. ATMs made each branch cheaper to run, so banks opened more of them, and tellers moved from counting cash to looking after customers.3
The machine took the transaction. People took the relationship, and there was more service, not less.
The factory
A century earlier, factories electrified by replacing the central steam engine with a large electric motor, and productivity barely moved for decades. The gains came only when factories were redesigned around electricity itself: power at every machine, new floor layouts, new ways of organizing work.4
The motor was necessary. The redesign was the point.
A chat box added to today’s work is the motor swap. The redesign is what’s coming.
In a commodity business, service is the product.
Where it matters most: commoditized services
The redesign will pay off most in service businesses that sell commoditized products.
Insurance distribution is the clearest example. A broker cannot make a carrier’s health plan different from the same plan sold down the street. In property and casualty, rates are filed with state regulators, so the same risk carries much the same price wherever it is placed.
What a brokerage actually sells is judgment, service and trust. That is why these businesses are valued for their client relationships and bought and sold by the hundreds each year.5 It is also why their largest cost is people: at the largest public brokers, compensation and benefits consume roughly half to three-fifths of revenue.6
In a commodity business, service is both the biggest cost and the only differentiator. Cut it and you lose clients. Fund it and you lose margin.
Models resolve that tension, but only if the work is divided correctly. People judge each other on two dimensions, warmth and competence, and they judge warmth first.7 Machines are climbing the competence axis quickly. The warmth that people trust with their health, their money and their business will remain human for a long time.
So the division of labor is clear: machines take the competence work behind the service, and people keep the relationship in front of it.
The same structure appears wherever a commodity is sold through a relationship: mortgage and title, payroll, wealth management, freight and staffing. We start with insurance because its work is the most document-dense and consequential. The thesis is general.
What we expect to happen
We expect the next 18 to 24 months to look like this. Each prediction follows from the curves above, and each can be checked.
- 01
AI moves from a chat box to a file in your inbox
You won’t ask for help drafting the renewal summary. It will be waiting when you get in, with the spreadsheet attached and every number reconciled.
- 02
Checking becomes the job
Anyone will be able to produce a proposal in a minute. The valuable person is the one who can see that the total on page seven is wrong, and prove it.
- 03
Your firm’s know-how becomes the advantage
Every firm will rent the same models. What differs is your past proposals, your client history, and the way your best producer explains a deductible.
- 04
You stop logging in
You forward an email or leave a voice note on the drive home, and the finished work comes back, in your firm’s voice.
- 05
Good enough stops being good enough
When anyone can make a deck, the one that looks and reads like it came from the best firm in the room wins.
- 06
The job changes before anyone announces it
The first sign won’t be layoffs. It will be the colleague who finishes a week of work by Tuesday. The firms that notice, retrain and redesign will keep their best people.
- 07
Better and cheaper, at the same time
Clients get faster answers and clearer materials, while the firm spends less to produce them. For the first time, service quality and cost stop pulling in opposite directions.
Better service and lower cost stop being a trade-off.
What we deliver
We measure everything we build against three outcomes.
Better experiences for the people being served
Faster answers, clearer materials and advice that fits their situation, in their language, at the hour they need it. The customer should feel the difference first.
Less busy work for the people doing the serving
No more rekeying the same data into three systems or rebuilding the same spreadsheet every renewal. The repetitive work goes to the system; the judgment and relationships stay with the person.
Lower costs, passed on as savings
When the cost to serve falls, a firm can serve more clients well, take on smaller clients it once turned away, and price more competitively.
The cycle
Lower costs mean better prices and more attention for every client. Better service keeps clients longer and brings in new ones. Longer relationships give the system more context, so it does better work, and the costs fall again.
In a commodity business, where service is the only real differentiator, this cycle compounds. The firms that start it first will be very hard to catch.
What we build
We do not train frontier models. We build what turns them into dependable work.
Harnesses
They give a model the right task, the firm’s own context, the right tools and clear limits.
Evaluations
They measure outputs against professional standards before anyone relies on them, and continuously afterwards.
Managed systems
We operate them and answer for them, measured on outcomes rather than demos.
Every output is proprietary, built from the firm’s own documents and standards, and verifiable: every figure traces to its source and every total reconciles.
Where the system is unsure, it says so and hands the question to a person.
Taste
In a world of abundant output, the scarce thing is knowing what is good.
to judge the visual appeal of a page.8
is how easier-to-read information is judged.9
of people could answer four basic questions about their own health insurance.11
This is not a matter of decoration. Interfaces that look better are even perceived as easier to use.10 In insurance, clarity is not a finishing touch. It is how people come to understand, trust and use what they bought.
So we treat taste as an engineering problem. To us, taste is the ability to predict what a discerning expert will prefer, and to know when to stop. We collect judgments from experienced designers and producers, turn them into evaluations our systems must pass, and encode restraint as a rule, because the best work usually has less on the page, not more.
Taste is also local. A firm’s voice, brand and standards are part of its proprietary context, and the work should sound like that firm at its best.
The human part
The best part of a service job was never the screen. It was the client who finally understands their coverage, the family that picks the right plan, the relationship that lasts twenty years. That is the work machines cannot do, and the work people actually like.
Imagine never logging in to do your job. You send an email or speak into a box, and the work comes out the other side: analyzed, reconciled, beautiful, and sounding like the best version of you. Your time goes to the person across the table.
Work with us
Builders with taste
Most of what we build has never been built before, so the work starts in the field: watching the work, talking to the people who do it, finding the pain they have stopped noticing. Then we model the finished outcome and work backwards until the problem is resolved, not just shipped. Part product, part design, part engineering. Models are the material; taste is the job.
Brokerages
Design partners, not customers. Bring your hardest workflows and your best people. We build the systems that give them their time back.
Notes
- METR (2025), “Measuring AI Ability to Complete Long Tasks.”
- Andreessen Horowitz, “Welcome to LLMflation” (2024): about 10× cheaper each year at equivalent performance. Epoch AI (2025) finds 9× to 900× a year, depending on the task.
- James Bessen, Learning by Doing: The Real Connection between Innovation, Wages, and Wealth (Yale University Press, 2015).
- Paul A. David (1990), “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review.
- OPTIS Partners counted 833 announced insurance agency and broker acquisitions in North America in 2023, and 750 in 2024.
- Compensation and benefits as a share of 2024 revenue: Marsh McLennan 57%, Aon 53%, Brown & Brown 50% (company filings).
- Fiske, Cuddy & Glick (2007), “Universal dimensions of social cognition: warmth and competence,” Trends in Cognitive Sciences.
- Lindgaard et al. (2006), “Attention web designers: You have 50 milliseconds to make a good first impression!” Behaviour & Information Technology.
- Reber & Schwarz (1999), “Effects of perceptual fluency on judgments of truth,” Consciousness and Cognition.
- Kurosu & Kashimura (1995), “Apparent usability vs. inherent usability,” CHI ’95.
- Loewenstein et al. (2013), “Consumers’ misunderstanding of health insurance,” Journal of Health Economics.
That’s what we’re building.