The AI Moat Pyramid

Series: AI Moats, Post 1 of 4 - How to turn a “cool demo” into a fortress your competitors can’t breach

Five years and US$62 million later, MD Anderson Cancer Center moth-balled its IBM Watson oncology pilot without ever putting it into production.¹
AI isn’t a silver bullet, especially inside legacy enterprises groaning under decades of tech-debt, brittle processes, and risk committees.

So when an executive says, “Let’s build an AI moat,” the correct response is: What are we defending and how?
Below is the framework to answer that question. Picture Maslow’s hierarchy for AI products: climb one layer at a time or tumble back to earth.

Layer 1 — Custom Models & Algorithms

Data → insight is table stakes; advantage appears only when your model outperforms the best open-source baseline and can be retrained in hours, not weeks. Deloitte pegs the cost of training an enterprise-grade LLM between US$1 million and US$100 million.²

3-Point Smell Test

Money – Exactly how many dollars does that ≥ 5 pp lift translate to this quarter?
Motion – If production data changes tomorrow, how soon will the new model be live, hours or sprints?
Muscle memory – Can ops point to the API call in their system, or do they “check the dashboard at 4 pm”?

Pass all three? Move up. Miss one? Fix it before bragging about “custom AI.”

Layer 2 — Proprietary Data

Historic data isn’t a moat if it’s trapped in PDFs or lawsuit-bait. Cigna’s PxDx algorithm auto-denied roughly 300,000 claims in two months, triggering congressional scrutiny.³

Gut-check:

• Unique – Could a rival legally gather 80 % of this data in six months?

• Usable – What share of high-value fields are fully labeled and version-controlled (target ≥ 80 %)?

• Ubiquitous – How many revenue-impacting teams can self-serve the data without opening a Jira ticket?

Layer 3 — Workflow Integration

Moats start forming when predictions drive or automate a decision path in under 30 seconds, not when they languish on a dashboard.

Friction: No copy-paste between model and action.
Feedback: Every automated decision is traced to its outcome.
Fallback: Ops can auto-rollback in < 5 minutes if the model misbehaves.

Layer 4 — Domain Expertise

In regulated, life-or-death arenas, domain nuance is non-negotiable. The Watson/MD Anderson pilot stalled partly because clinical protocols inside the model lagged behind current oncology guidelines.

Compliance | Credibility | Continuity

1. Which regulation (HIPAA, SOX, FAA, etc.) could shut you down tomorrow—and how do you prove compliance today?

2. Can an SME explain a live decision in < 60 s and sign their name to it?

3. When rules change, does the model update inside the same quarter?

Layer 5 — Network Effects

AI flywheels appear only when you capture user telemetry and retrain fast. Zillow’s iBuyer unit couldn’t keep up with market drift, leading to a US$500 million write-down and the division’s shutdown.⁴

Telemetry | Tempo | Trajectory

• Log at the interaction level, not pageviews.

• Target ≤ 14 days from data arrival to model live.

• Look for a ≥ 20 % lift during the first six months of live feedback.

Layer 6 — Strategic Moats

Rare air: advantages outsiders can’t swipe with a bigger budget, exclusive data ecosystems, regulatory barriers, or brand-embedded workflows. Think Bloomberg Terminal or GE’s industrial IoT stack.

Scarcity | Switching Cost | Scale

• What critical asset can only you deliver, and for how many years is it locked?

• If a customer tried to leave tomorrow, how many systems and filings would they have to re-engineer?

• Do margins expand or compress as you add the next ten enterprise clients? Show the math.

Why Most Teams Stall on the Lower Rungs

Gartner estimates that 85 % of AI projects fail to deliver business value.⁵
Most initiatives die before they reach network effects, let alone achieve strategic moats.

Climb Your Next Layer

1. Score yourself 0–3 on each layer’s gut-check. Any zero is a red light.

2. Fix the lowest weak layer first. Moats are built from the bottom up, never from the top down.

3. Instrument everything. If you can’t measure lift, you can’t prove a moat.

Let’s build moats; dashboards can’t hide.

1. https://www2.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2025/past-tmt-predictions-hits-and-misses.html

2. https://www.propublica.org/article/cigna-pxdx-medical-health-insurance-rejection-claims

3. https://www.forbes.com/sites/matthewherper/2017/02/19/md-anderson-benches-ibm-watson-in-setback-for-artificial-intelligence-in-medicine/

4. https://www.forbes.com/sites/matthewherper/2017/02/19/md-anderson-benches-ibm-watson-in-setback-for-artificial-intelligence-in-medicine/

5. https://www.forbes.com/councils/forbestechcouncil/2024/11/15/why-85-of-your-ai-models-may-fail/