AI at the Core vs AI at the Edge
Decide whether to sprinkle AI onto existing code or rebuild the workflow with the LLM as the core
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 88%
A diagnostic for how a company adopts AI. 'AI at the edge' inserts LLMs at intersection points while the existing codebase and software stay intact. 'AI at the core' fundamentally re-examines the problem space and workflows and uses AI as the central capability — often the codebase shrinks and the LLM becomes the primary way the problem is solved. Companies that only attach AI to old code rarely revolutionize their industry; those whose codebase shrinks around the LLM tend to do it right.
Origin
Ezinne and Oji Udezue's framing from their consulting work with companies adopting AI, discussed in the context of their book 'Building Rocket Ships.'
Core principles
- 01AI is a core set of multiple capabilities, not a magic layer you slather onto a product
- 02The problems are still the problems and the customer is still the customer — AI doesn't change the underlying sharp problem
- 03AI at the edge = same software with LLMs inserted at connection points; AI at the core = rebuild the workflow with the LLM central
- 04Specificity beats breadth: specialize first, then add a connective-tissue intelligence layer across specialized pieces
- 05Incumbents with revenue tied to old codebases must use AI on the old thing while navigating to the new thing 'before someone eats your lunch'
How to run it
- 1
Reject the 'slather it on' reflex
Recognize AI as a set of capabilities, not a magic coating. Reaffirm that the customer's sharp problem still exists and must still be solved.
Watch out Expecting an LLM to be 'as broad as the problem sets you want to solve for' leads to a bloated, mediocre solution.
- 2
Diagnose edge vs core
Assess your current build: are you keeping the existing bits and inserting LLMs at UI/connection points (edge), or fundamentally looking at the problem space and workflows and using AI to solve it (core)?
Pro tip A blank-slate test: 'if we built this today with the LLM as the core capability, how would it look?' often reveals acceleration and new adjacencies.
Watch out Companies whose codebase merely gains attached AI usually won't revolutionize their industry.
- 3
Specialize, then add a connective intelligence layer
Instead of one massive do-everything LLM, build specific specialized solution sets, then tie them together with a broader multi-model connective layer that offers intelligence across them.
Watch out Trying to build one broad solution set or single LLM that 'can do it all' is the failure pattern they repeatedly observed.
- 4
For incumbents, navigate from old to new
If you have revenue tied to legacy codebases, use LLMs on the old product while deliberately building the AI-core version before a competitor displaces you.
Pro tip Clay is cited as a six-year-old company that reimagined its product once LLMs arrived and stayed successful.
Watch out This is complicated and hard, but avoiding it risks 'someone eats your lunch.'
In the wild
Oji, David Okuniev and a small group built Formless, a new type of Typeform with the LLM at the core. Because the AI capability was so strong, the product went beyond forms — it became a sales-lead agent that did pre-sales qualification that normally happens only after a human salesperson gets involved.
→ The AI-at-the-core rebuild took on an adjacency (pre-sales) that the edge version never could.
Oji notes Clay, a six-year-old company, reimagined its product once LLMs arrived rather than merely attaching AI to old code.
→ It stayed 'super successful' by moving toward the AI-core model.
Common mistakes
Slathering AI on as a magic layer
Treating AI as a magic thing you sprinkle at UI intersections leaves the old codebase and workflow intact, which rarely revolutionizes the product or industry.
Expecting one broad LLM to cover everything
Companies try to build a massive solution set or single LLM as broad as all their problems; the successful pattern is to specialize into specific solution sets first and connect them with a multi-model intelligence layer.
Is it for you?
Best for
Product leaders and founders deciding how deeply to rearchitect a product around AI, including incumbents with legacy revenue
Not ideal for
Teams whose problem is genuinely well-served by a thin AI assist and who would over-engineer by rebuilding the core
From the transcript
“we have this phrase called AI at the core and AI at the edge”
“AI at the core means fundamentally looking at the problem space and the workflows and using AI to solve the problem. Not just sprinkling it…”
“Companies for whom their code base perhaps shrinks and the LLM becomes a core part of what it is that they use to solve the…”
“to do that you actually will need to have specialized first and then create a layer of a connective tissue that can offer intelligence”
“navigate to the new thing before someone eats your lunch”
From the episode
How AI is reshaping the product role
Oji and Ezinne Udezue