Design with a Frontier Model, Run with an Open-Weight Model

How to turn expert judgment into AI-enabled workflows that are practical to operate, easier to govern, and built to last.

The most powerful AI model does not need to handle every decision.

In many enterprise workflows, the difficult part is figuring out how an experienced professional makes a judgment. Once that knowledge has been captured, tested, and approved, applying it to routine cases may require something much simpler.

That suggests a different way to design enterprise AI: use a frontier model to help experts define the decision process, then use open-weight models and conventional software to carry it out.

The distinction matters. It can reduce unnecessary complexity, make decisions easier to trace, and give organizations more control over how AI operates.

But the real value comes from the work surrounding the models: understanding the process, making expertise explicit, and keeping people responsible for the decisions that matter.

Stage 1: The Knowledge Problem Behind Enterprise AI

Think about the people everyone turns to when a case gets complicated.

A senior property appraiser can spot a finding that deserves closer attention. A compliance specialist recognizes when an otherwise ordinary document contains an unusual risk. An operations manager knows when an exception is harmless and when it signals a larger problem.

Their decisions draw on formal procedures, previous cases, business context, and experience that is difficult to put into words.

Some of that knowledge is documented. Much of it lives in people’s heads.

This creates a familiar operational problem. Junior employees need frequent guidance. Similar cases can receive different treatment. Senior specialists become bottlenecks because too many decisions depend on their availability.

And when those specialists leave, the organization may lose more than a person. It may lose part of its ability to make sound decisions consistently.

AI can help address this problem, but there is a catch.

Giving a large language model access to internal documents does not automatically give it the judgment of an experienced professional. Nor does asking a frontier model to assess every case guarantee consistent results.

For repeatable, high-volume processes, that approach can also become expensive and difficult to audit.

The better question is not simply which model to use. It is how to turn expert knowledge into a decision process that the organization can understand, validate, and maintain.

Two Different Jobs for Two Different Types of AI

The architecture starts by separating knowledge design from everyday execution.

A frontier model is useful when the work requires exploring complex information, asking questions, finding patterns, and helping experts articulate how they think.

An open-weight model may be selected for a narrow, repeatable task that repeats hundreds or thousands of times.

Conventional software remains the right choice for calculations, explicit rules, and other operations that should produce predictable results.

These technologies do not have to compete for the same job.

The frontier model helps shape the decision framework. Open-weight models and software apply it. Human experts approve the framework, handle exceptions, and remain accountable for consequential decisions.

That separation makes the system easier to reason about, both technically and operationally.

First, Capture How Experts Actually Make Decisions

Suppose a senior reviewer is asked to explain how they assess a property inspection report.

They may begin with a straightforward answer: check the condition of the roof, structure, and mechanical systems, then determine whether any findings require further investigation.

But the real decision process is rarely that simple.

Which findings matter most? What happens when two minor issues appear together? When is the available evidence insufficient? What distinguishes a routine observation from something that could materially affect a valuation?

An experienced reviewer often knows the answers without consciously following a written checklist.

This is where a frontier model can help.

Working with subject-matter experts, it can examine procedures, historical reports, relevant standards, and examples of previous decisions. It can ask follow-up questions, compare cases, and help expose criteria that have never been formally documented.

The aim is to develop a decision map: a clear description of what information is needed, how it should be assessed, and when a case needs human attention.

The model’s role is to support that work, not to declare the rules correct.

Experts still need to challenge the proposed logic. They need to test it against ordinary cases, difficult exceptions, incomplete information, and situations where reasonable professionals might disagree.

Only then should the decision framework be approved for implementation.

Once approved, it should be treated as a maintained operational asset. Changes need owners, testing, version control, and a way to reverse updates that do not perform as expected.

That is an important shift. The organization is no longer relying entirely on knowledge that is difficult to access or transfer. It has a framework that people can inspect, improve, and use consistently.

Stage 2: Then, Use Simpler Systems for Everyday Execution

The approved map goes to a small, inexpensive AI agent that does the routine work. It does not improvise. At each step, one of three things happens:

• Where a step is arithmetic, it calculates.
• Where a step needs a judgment call, such as reading a note and deciding whether it describes a serious problem, it uses a small specialized model and reports how confident it is.
• Where confidence is low, or the situation is not covered by the map, it stops and hands the case to a person.

Because the heavy thinking already happened at design time, the agent can run on modest hardware, including a small machine on a factory floor or in a branch office, close to where the work happens.

Where This Fits into Managed AI Services (MAIS™)

Choosing the right AI model is only one part of building a system that works in production.

Before deciding whether a frontier model should capture expert knowledge or an open-weight model should handle routine tasks, an organization needs to understand its existing workflows, the quality of its data, and the decisions it wants to improve.

That is where First Line Software’s Managed AI Services (MAIS™) provides a practical starting point.

The first two phases help organizations establish whether an AI opportunity is viable and determine how the solution should be designed.

Step 1: Discover — Business & Data Audit 

The first step is to understand how the business operates today.

First Line Software’s Business & Data Audit examines existing workflows, data sources, operational dependencies, and risks that could affect AI adoption.

For the frontier-model/open-weight-model approach, this means identifying where expert judgment enters a process and what information supports it.

Consider a property inspection workflow. An audit would examine how reports move between inspectors and reviewers, where manual checks occur, which systems hold the relevant information, and whether the data is sufficiently complete and reliable.

It may also reveal that different teams follow different review practices or that critical decisions depend on information scattered across documents and systems.

These findings matter because AI cannot compensate reliably for an unclear process or poor-quality inputs.

The audit establishes an evidence-based view of the current environment, including data maturity, operational risks, and areas requiring remediation.

At this stage, the objective is to understand the foundation. The detailed AI architecture and model selection come later.

Step 2: Design & Select — AI Alignment with Your Business 

Once the organization understands its current state, the next step is deciding where AI belongs and how it should operate.

Through AI Alignment with Your Business, First Line Software connects business priorities with technical feasibility, implementation requirements, and realistic expectations for value and cost.

This is where the frontier-model/open-weight-model approach becomes an architectural choice.

A team can evaluate which parts of a workflow require complex knowledge analysis, which can follow approved decision rules, and which need human oversight.

For example, a frontier model might help senior reviewers document the criteria behind property inspection decisions. Once those criteria have been validated, the team could assess whether open-weight models are suitable for classifying observations while conventional software handles completeness checks and explicit rules.

The selection should reflect the actual requirements of the workflow.

We decide which processes benefit from AI and which model fits each one. We test every candidate process against four questions. Does it repeat? Does it follow rules? Does it need an audit trail? Is expert time the bottleneck? Processes that pass go first. The model choice then follows from the decision map instead of from preference or hype. Frontier models are used where they earn their cost, in capturing and modeling the knowledge. Open-weight models and plain code run the daily work.

These questions guide model selection, deployment architecture, integration planning, and the controls needed for production use.

The outcome is an implementation direction grounded in business needs and technical constraints.

From Architecture to Ongoing Operations

Discovering the opportunity and selecting the architecture are important early steps. The system still needs to be engineered, integrated, evaluated, and maintained.

That includes validating decision maps, testing model performance, monitoring exceptions, and updating approved rules when business requirements change.

The goal is to build an AI-enabled workflow that the organization can operate with confidence, rather than relying on a powerful model to make every decision independently.

A Practical Example: Reviewing Commercial Property Inspections

Imagine a commercial real estate organization handling a large volume of property inspection reports.

Inspectors document the condition of roofs, structural elements, mechanical systems, and other parts of a building. Experienced reviewers assess the findings and decide which require further investigation.

Much of the work follows familiar patterns. Some findings are routine. Others depend on context or combinations of issues that an experienced reviewer recognizes.

The organization could begin by working with senior reviewers and a frontier model to document those distinctions.

What evidence supports a routine assessment? Which observations should trigger additional checks? When should a case go directly to a senior appraiser?

After testing and approving the criteria, the organization could build a workflow in which an open-weight model classifies observations from inspection notes. Conventional software checks the completeness of the report and applies the relevant rules.

Cases that fall outside the approved criteria are escalated.

The goal would be to make preliminary reviews more consistent while reserving specialist attention for cases that genuinely require it.

This is an illustrative example of how the architecture could work. It does not describe a verified deployment of this exact two-stage system.

First Line Software’s related experience with AI-property inspection reporting includes document extraction, standardized reporting, quality checks, and human validation. The architecture discussed here would need its own assessment and technical validation.

What Changes for the Business?

The immediate attraction of this approach may be lower runtime costs. But the more interesting benefits concern how the organization manages decisions.

Expertise becomes easier to share. Instead of depending entirely on informal guidance, teams can work from approved criteria that experienced professionals have helped define.

Routine work becomes more structured. Similar cases can follow the same process, while exceptions receive explicit attention.

Decisions become easier to examine. When the system records its inputs, rules, and approvals, reviewers have a clearer basis for understanding what happened.

Specialists can spend more time on difficult cases. Preliminary checks may be handled by software, provided the automation is accurate enough to reduce rather than increase review effort.

The organization gains more control over change. Decision criteria can be updated, tested, and released deliberately as policies or operating conditions evolve.

These are opportunities to evaluate, not guaranteed outcomes.

A poorly designed decision map can make the wrong decisions more consistent. An inaccurate classifier can create more work for reviewers. And the cost of maintaining the system may outweigh the savings if the process is too small or unpredictable.

The economics and reliability of the complete workflow matter more than the performance of any individual model.

Where the Approach Makes Sense, and Where It Doesn’t

This architecture is a strong candidate for investigation when the work is repetitive, the decision boundaries are reasonably clear, and experts can explain what good judgment looks like.

Document quality checks, inspection triage, preliminary compliance reviews, and operational exception handling are possible examples.

It is a less natural fit for work that depends on open-ended exploration or rapidly changing context.

A research team investigating an unfamiliar market, for example, may need flexible reasoning rather than a fixed decision map. The same may be true of strategic negotiations or creative work.

Risk also changes the equation. Decisions with serious financial, regulatory, or safety consequences may require extensive validation and mandatory human approval, even when much of the workflow can be automated.

The right architecture is the one that fits the decision, the data, and the consequences of getting it wrong.

Enterprise AI Needs More Than a Powerful Model

There is a tendency to judge enterprise AI by the capabilities of the model at its center.

But in day-to-day operations, the surrounding system often matters just as much.

Who defined the decision criteria? How were they tested? What happens when the model is uncertain? Who approves changes? Can the organization explain an outcome six months later?

These questions determine whether AI remains an experiment or becomes a dependable part of the business.

Using a frontier model to help capture expert judgment, then open-weight models and software to apply it, offers one practical way to answer them.

It puts the organization’s knowledge and decision process at the center of the architecture.

Through Managed AI Services, First Line Software helps organizations identify suitable AI opportunities, connect them to business requirements, and develop solutions designed for real operating environments.

The Asset That Lasts

Anyone can buy access to a powerful AI model. What is hard to replicate is the captured knowledge of how your company makes good decisions. That is the part worth building.

If you want to see how this would work on one of your processes, get in touch with our team.

November 2026