Depth-Over-Breadth AI Opportunity Filter
Turn proprietary niche data into a focused, defensible AI product.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 94%
This framework reverses the Web 2 assumption that the largest dataset and broadest platform must win. Begin with a narrow market, a valuable workflow, and data that a small team can access or uniquely understand. Train or adapt the smallest model capable of transforming how customers interact with that information, then compare its usefulness with a general-purpose model. Because focused models require fewer parameters and lower training costs, the team can run faster experiments and improve the product around real customer needs. The resulting advantage does not come from owning the biggest foundational model; it comes from combining specialized data, domain knowledge, rapid iteration, and a deeply useful experience that broad competitors are unlikely to prioritize.
Origin
Extracted from Marketing Against the Grain during Kieran Flanagan and Kipp Bodnar's analysis of Google's leaked memo about small, open-source AI models.
Core principles
- 01Focused data can outperform indiscriminate scale.
- 02Unique data creates opportunities that general-purpose models cannot serve well.
- 03Small models make experimentation cheaper and faster.
- 04Depth of usefulness matters more than breadth of coverage.
- 05Accessible infrastructure lets small teams compete with incumbents.
How to run it
- 1
Choose a narrow market
Define one specific customer group and one information-intensive workflow where a better interface to data would create meaningful value.
Pro tip Prefer a market where your team already understands the language, constraints, and buying behavior.
Watch out Do not begin with a vague ambition to build an AI tool for everyone.
- 2
Inventory differentiated data
List the proprietary, specialized, or difficult-to-aggregate information available to you. Determine whether it is sufficiently relevant and reliable to improve the chosen workflow.
Pro tip Value relevance and quality over raw dataset size.
Watch out Possessing a large dataset does not make it useful or legally safe to train on.
- 3
Build the smallest viable model
Use a compact model, retrieval system, or adaptation that can perform the target task. Avoid paying for unnecessary parameters or broad capabilities.
Pro tip Establish a general-purpose model as the baseline to beat.
Watch out Do not confuse model sophistication with customer value.
- 4
Embed it in the workflow
Create a focused product experience that helps customers interact with the specialized data and complete a real task.
Pro tip Concentrate on the output or decision the customer needs, not the novelty of the model.
Watch out A model without a usable application is not a complete product.
- 5
Iterate for depth
Measure accuracy, speed, cost, and workflow outcomes, then improve the model and experience through short experimental cycles.
Pro tip Use customer corrections and edge cases to prioritize subsequent iterations.
Watch out Do not expand into adjacent markets until the initial use case is demonstrably valuable.
In the wild
A healthcare software company organizes its licensed clinical reference material and builds a focused assistant for one specialty. It compares the assistant with a general chatbot, tests it with qualified users, and improves retrieval around the questions practitioners repeatedly ask.
→ The company produces a more relevant specialist workflow without training a massive general-purpose model.
The hosts cite Bloomberg as an example of a company applying AI to a valuable body of specialized financial data rather than trying to reproduce a model trained on everything.
→ Existing domain data becomes the foundation for a focused AI experience in financial information.
Common mistakes
Chasing maximum parameters
A larger model may increase cost and iteration time without improving the narrow customer outcome that matters.
Starting without unique data
A thin interface over a commodity model is easy to reproduce unless the team contributes distinctive information, expertise, or workflow integration.
Expanding before proving depth
Broadening the product too early sacrifices the focused usefulness that gives a small team an advantage.
Is it for you?
Best for
It is best for teams with domain expertise, distinctive data, and a specific customer workflow they can improve.
Not ideal for
It is not ideal for businesses lacking differentiated data or a clearly defined user problem.
From the transcript
“models trained on very specific data sets are much better than models trained on all of the data.”
“we are moving to an era of depth instead of breadth.”
“if you've got some unique data in a very specific market, you have the ability to build a model, a large language model pretty cheaply…”
From the episode
Leaked Google Memo Reveals A Huge Opportunity For Entrepreneurs (#117)