Insight · October 15, 2026

The model is
twenty percent.

A useful vertical agent has four parts besides the model. Most of the value sits in the four.

01 · The shift

The model is the part everyone can rent.

A recent post by Greg Isenberg, on what he heard from AI founders in San Francisco, made a point worth repeating. The builders doing well with vertical agents are not winning on the model. They treat it as one input, and they put their effort around it.

Twenty percent is a way of speaking, not a measured split. The point is the direction. The same strong models are available to every competitor on the same day, at similar prices. Whatever you build on top of them has to be something a competitor cannot rent.

02 · Why the model alone falls short

A general model knows the world. It does not know your work.

Ask a strong model to process an insurance claim, review a supplier invoice or check a maintenance record. It will produce something fluent. It will not know your systems of record, your thresholds, your exceptions or what a good result looks like in your trade.

That knowledge lives in four places outside the model. It lives in the data the industry keeps. It lives in the heads of people who have done the job for years. It lives in the rules the business has about who may do what. And it lives in the moment a person looks at the result and says yes or no. Build those four, and the model becomes a part inside a system. Skip them, and the model is a demo.

The one line to keep

“Treat the model as a part you can swap.”

03 · The four parts

Four things around the model.

Connectors to the data

Read access to the systems where the industry work is recorded: files, records, feeds, tickets. The agent is only as informed as what it can reach, and reach has to be clean and current.

Evals from the people who know the work

A set of real cases with the right answers, written and judged by practitioners. Engineers cannot write this alone, because they do not know what right looks like.

Rules for read and change

A clear list of what the agent may read, what it may change and what it must never touch. Enforced in the system, not requested in the prompt.

A review screen for a person

A clean place where a human sees what the agent read, what it wants to do and why, and approves or rejects in seconds.

04 · Evals

Written by people who know the work.

The eval is the part most teams underbuild, and it is the one that carries the expertise. A claims examiner knows that a certain phrase in a policy changes the answer. A buyer knows which supplier names are the same company. An engineer knows which readings are noise.

Those people write the cases and the pass rule. The system runs them on every change, including a change of model. We covered the mechanics in Evals Are the New Unit Test. The short version is that an eval is how you know. Without one, you cannot swap a model safely, because you cannot tell whether the new one is better.

05 · An example

One workflow, different models for different jobs.

Picture an agent that handles supplier invoices. Reading the invoice into fields is a simple, repeated job. Matching it to a purchase order needs access to the finance system and some care. Deciding what to flag as an exception takes the most judgment.

There is no reason to use the most expensive model for all three. Route each job to the model that handles it well at the right cost, and let the evals confirm each choice. When a better or cheaper model ships, change one line and run the suite.

route
  read invoice             small, fast model
  match to purchase order  mid model, connector to finance system
  flag exceptions          strongest model
  every output             eval gate, then review screen
  permissions              read finance data, change nothing

A routing table · the model is a setting, the system is the product

06 · Rules

The agent should not decide what it is allowed to do.

Permissions belong in the system, not in the instructions. An agent told in its prompt not to touch payments can be talked out of it. An agent that has no write access to payments cannot.

Write the rules as a short table. What can it read. What can it change. What needs a person first. What is off limits. Keep the default narrow and widen it only when the evals show the agent earns it.

07 · The review screen

The last screen is where the trust is built.

Someone has to stand behind the result. A good review screen shows the source the agent used, the action it proposes and the reason, on one page, with approve and reject as the obvious moves. A person can clear a queue quickly when everything needed to decide is already in front of them.

This is also the cheapest source of improvement. Every rejection is a labeled example. Feed it back into the eval set and the system gets sharper without anyone writing a new prompt.

Closing

Build the four parts. Rent the model.

List the four parts for the agent you are planning. Mark each one as built, partly built or missing. The missing ones, not the model, are where your next quarter should go.

Written by · October 15, 2026

Donny Smith · ECD, Founder, Bttr.

Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.

A recent post by Greg Isenberg on what he heard from AI founders in San Francisco, paraphrased · Bttr. perspective on agent systems

Share this perspective

More insights

Adjacent perspectives.

8 min read

How to Evaluate a RAG System

Evaluating a RAG system means scoring its two stages on their own, because a retrieval augmented generation answer is only as good as the passages behind it, and a wrong answer has two possible causes a single number cannot tell apart. Retrieval is measured before the model writes anything. The LlamaIndex evaluation guide scores a retriever with hit rate, the share of questions where the right passage lands in the top results, and mean reciprocal rank, which rewards ranking it high. Generation is measured given what was retrieved. Ragas, the open evaluation toolkit now maintained by VibrantLabs, splits the answer into claims and scores faithfulness as the share the retrieved context supports, then scores answer relevance by generating questions from the answer and comparing them to the real one. Read as two pairs, the four metrics locate the fault: high recall with low faithfulness is a generation problem, low recall with high faithfulness is a retrieval one. Anthropic showed the payoff in September 2024, cutting a retrieval failure rate measured by recall at 20 from 5.7 percent to 1.9 percent by adding context, keyword search, and reranking. And it is not a one time check. Models drift, documents go stale, so the teams that trust their systems run these scores on live traffic the way they run unit tests.

8 min read

What Is a Reasoning Model

A reasoning model is a large language model trained to work through a problem in steps before it answers, rather than reply in one pass. It spends extra computation at the moment you ask, called test-time compute, to write out a chain of thought and check its own work. Anthropic describes its extended thinking as letting the very same model give itself more time and effort rather than switching to a separate model, and reported that Claude 3.7 Sonnet reached 84.8% on the hard GPQA science set, with a 96.5% physics subscore, using parallel sampling at a 64K token thinking budget. The habit is trained, mostly by reward. DeepSeek showed in January 2025 that its R1 Zero model learned to reason purely through reinforcement learning with no supervised fine-tuning first, developing self checking and long chains of thought on its own, and shipped DeepSeek R1 under the MIT license at performance comparable to OpenAI o1. The cost is real, so the newest models add a dial. Claude Opus 4.5 hit its best score on SWE-bench Verified using 76% fewer output tokens at a medium effort. The skill is matching how hard the model thinks to how hard the task is.

8 min read

What Is Fine-Tuning

Fine-tuning is training a pre-trained model further on your own labeled examples so a behavior is set into its weights. It changes how a model acts, its tone, its format, its grip on a narrow task, rather than what facts it knows. Google frames the move as taking a general-purpose model and adapting it to your domain on a curated dataset of your own examples. The gains are measurable. When Anthropic opened fine-tuning for Claude 3 Haiku in Amazon Bedrock, a team fine-tuned it to moderate comments and lifted classification accuracy from 81.5% to 99.6% while cutting tokens per query by 85%. Lighter methods such as LoRA keep the base model frozen and train a small set of new weights beside it. The order that saves money is a plain prompt first, retrieval when the model lacks facts, fine-tuning only when it lacks the right behavior. A fact that changes is a retrieval problem. A behavior that drifts is a training problem.

View all insights →

Bttr. Field Brief

The brief Bttr. writes for senior buyers.

Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.

Build the system around the model

We build the connectors, evals, rules and review screens that make an agent useful in your industry. The model is the swappable part.

See how we make companies agent ready →

The model is
twenty percent.

The model is the part everyone can rent.

A general model knows the world. It does not know your work.