Talk to us
$

$ teds read --post frame-method-for-picking-ai-projects

The FRAME Method for Picking AI Projects

A practical framework for deciding which generative AI ideas are worth building before teams waste time on demos that never become products.

AI EngineeringLLM Systems

The FRAME Method for Picking AI Projects

TL;DR

  • The fastest way to waste money on AI is to skip problem framing.
  • FRAME helps teams evaluate AI projects before building: Find, Research, Assess, Model, Evaluate.
  • The best AI projects have real pain, usable data, measurable quality, manageable risk, and a clear workflow.
  • Use the worksheet in this post before starting a prototype.

Abstract

Generative AI makes it easy to build something impressive before you know whether it should exist.

That is the trap.

The model works. The demo feels promising. The team can imagine ten use cases. But the project still fails because the problem was weak, the workflow was unclear, the data was not ready, or nobody knew how to evaluate the output.

FRAME is a simple method for avoiding that failure. It helps teams move from excitement to judgment before they spend weeks building the wrong thing.

Table of Contents

  • Why AI Project Selection Is Hard
  • The FRAME Method
  • Find the Problem Domain
  • Research Existing Solutions
  • Assess Generative AI Fit
  • Model the Solution Framework
  • Evaluate Feasibility and Impact
  • FRAME Worksheet
  • Summary
  • Next Steps

Why AI Project Selection Is Hard

AI project selection is hard because the technology is too flexible.

A good model can summarize, draft, classify, retrieve, transform, reason over documents, write code, describe images, and generate media. That flexibility creates a planning problem. Almost every workflow begins to look like an AI opportunity.

But an AI opportunity is not the same thing as a good project.

Good AI projects have five traits:

  • a painful enough problem
  • a workflow where AI changes something meaningful
  • data or context the system can use
  • a way to evaluate quality
  • a risk level the team can manage

Weak projects usually lack one of those. They begin with a model capability and then hunt for a business case. That order creates attractive prototypes and disappointing products.

FRAME reverses the order.

The FRAME Method

FRAME stands for:

  • Find the problem domain.
  • Research existing solutions.
  • Assess generative AI fit.
  • Model the solution framework.
  • Evaluate feasibility and impact.

It is not a heavyweight strategy process. It is a practical filter. The goal is to decide whether an idea deserves a prototype, more research, or a polite death.

Find The Problem Domain

Start with the domain, not the model.

A domain is a specific area where a problem exists: customer support, clinical documentation, legal review, construction safety, sales operations, developer onboarding, financial analysis, education, content moderation.

The first questions are simple:

  • Who has the pain?
  • How often does it happen?
  • What does it cost?
  • What does the current workaround look like?
  • Why has this not already been solved?

The last question matters. If a workflow has remained painful for years, there is usually a reason. The data may be messy. The decisions may require expertise. The risk may be high. The users may not trust automation. The process may cross too many systems.

Those constraints do not kill the project. They make the real project visible.

This is similar to a developer journey audit. Before hiring more people or shipping more content, you find the friction. The same idea applies to AI products: find the real friction before you add a model. See the developer journey audit for a related way to think about friction.

Research Existing Solutions

Once the domain is clear, study what people already do.

Look at:

  • manual workflows
  • existing software
  • internal scripts
  • spreadsheets
  • rules-based systems
  • traditional machine learning
  • human review processes
  • market competitors
  • open-source tools

The point is not to prove that AI is more exciting. The point is to understand the default.

Every AI product competes with something. Sometimes it competes with another vendor. Sometimes it competes with a spreadsheet. Sometimes it competes with a human expert who is slow but trusted. Sometimes it competes with doing nothing.

If the existing solution is cheap, reliable, and familiar, AI needs a strong reason to enter the workflow.

Research also reveals the gaps:

  • current tools are too slow
  • users cannot search across messy data
  • experts are overloaded
  • outputs need personalization
  • visual inspection does not scale
  • knowledge is trapped in documents

Those gaps become the places where generative AI might help.

Assess Generative AI Fit

Now ask whether generative AI is actually a good fit.

Use these questions:

  • Does the task require generation, interpretation, transformation, or synthesis?
  • Is ambiguity acceptable?
  • Is there usable data or context?
  • Can the output be evaluated?
  • Can a human review uncertain cases?
  • Would rules or traditional automation be enough?
  • What happens if the model is wrong?

AI fit is strongest when the task has language, images, documents, or messy human input at its center. It is weaker when the task is deterministic, high-stakes, and easy to solve with rules.

For example, an LLM may be useful for summarizing support conversations. It may be unnecessary for routing three known ticket types by keyword. A vision-language model may help inspect job-site images. It may be excessive if a barcode scanner solves the problem.

Good AI judgment is often subtractive. You are not only deciding where to use AI. You are deciding where not to.

Model The Solution Framework

If the idea survives the fit check, model the solution.

This does not mean choosing every vendor. It means sketching the system:

  • user input
  • required context
  • model behavior
  • retrieval or data access
  • output format
  • review path
  • feedback capture
  • evaluation method
  • deployment environment

For an internal policy assistant, the framework might include document ingestion, retrieval, citations, answer generation, abstention behavior, user feedback, and a review queue for failed answers.

For a medical image support tool, the framework might include image upload, model inference, region highlighting, specialist review, audit logging, and a strict boundary around what the system can claim.

The framework should make the product less magical. That is good. Magic is hard to operate.

Evaluate Feasibility And Impact

The final step is deciding whether the project is worth doing.

Evaluate:

  • technical feasibility
  • data readiness
  • cost
  • latency
  • risk
  • regulatory requirements
  • maintenance burden
  • expected ROI
  • user adoption
  • organizational ownership

A project can be technically possible and still not worth building.

It may be too expensive to run. It may require data the team does not control. It may create liability the organization cannot manage. It may solve a problem users do not care about enough to change behavior.

This is where many AI strategies should get smaller. A smaller workflow with a clear user, clear data, and clear evaluation is better than a broad platform idea nobody can measure.

FRAME Worksheet

Use this before starting the prototype.

Problem

  • Who is the user?
  • What is the painful workflow?
  • How often does it happen?
  • What does it cost today?

Existing Solution

  • How is the problem solved now?
  • What tools or workarounds already exist?
  • What is broken about the current approach?

AI Fit

  • What does generative AI improve?
  • Is generation, retrieval, synthesis, or multimodal understanding required?
  • Would simpler automation work?
  • Can a human review the output?

Solution Model

  • What data or context is needed?
  • What model capability is required?
  • What does the user see?
  • How is feedback captured?

Feasibility and Impact

  • Data readiness: 1-5
  • AI fit: 1-5
  • Evaluation clarity: 1-5
  • Workflow fit: 1-5
  • Risk level: low / medium / high
  • Expected impact: low / medium / high
  • Recommendation: reject / research / prototype / productize

Want us to pressure-test your AI project before you build? We help teams turn vague AI ideas into clear build, break, and explain plans.

Summary

Good AI strategy is mostly good problem selection.

FRAME gives teams a simple way to slow down before building:

  • Find the domain.
  • Research what already exists.
  • Assess AI fit.
  • Model the system.
  • Evaluate feasibility and impact.

The goal is not to make every idea sound viable. The goal is to find the few ideas that deserve real work.

Next Steps

Run one candidate project through the worksheet. If the project cannot produce a clear user, workflow, data source, evaluation method, and risk model, it is not ready for a prototype.

Need help turning an AI idea into a working system? We build, break, and explain AI. Then we hand it back working.