📌 Quick Navigation
I remember sitting in a boardroom three years ago, watching a team proudly present their new AI model. They'd spent 80% of their budget on the latest GPU cluster and a fancy deep learning framework. The model hit 98% accuracy in tests. But when it went live? A total flop. Data pipelines broke, the ops team had no idea how to maintain it, and business stakeholders didn't trust the output. That's when I first truly understood the 30% rule for AI — a simple but brutally honest guideline that separates successful AI projects from expensive failures.
Where the 30% Rule Came From
I first heard the term from an old mentor who'd been building AI systems since the 1990s. He said: “If you spend more than 30% of your AI budget on the technology itself, you're setting yourself up for disaster.” The rule has since been echoed by industry leaders like Andrew Ng and countless practitioner reports. It's not a hard law, but after watching over a dozen AI initiatives — both my own and others' — I've seen it hold true almost every time.
The idea is simple: AI success is about 30% model and 70% everything else. The “everything else” includes data quality, infrastructure, team training, change management, and continuous monitoring. It's the boring, unsexy work that nobody talks about at conferences.
Breaking Down the 70/30 Split
Let me get specific. I've categorized the typical spend areas based on what I've seen work:
| Category | Percentage | What It Covers |
|---|---|---|
| AI technology (models, tools, compute) | 30% | Cloud GPU, ML frameworks, model development, initial testing |
| Data engineering & management | 25% | Data cleaning, labeling pipelines, storage, versioning |
| Infrastructure & MLOps | 15% | Deployment pipelines, monitoring, CI/CD, drift detection |
| People & training | 20% | Upskilling engineers, hiring MLOps specialists, change management for users |
| Governance & compliance | 10% | Bias auditing, explainability, legal review, documentation |
Notice that the actual model work fits into just the top 30%. The rest is everything needed to make that model useful in the real world. I've found that teams who follow this split rarely end up with a “proof of concept” that never sees production.
A Real-World Example That Changed My Mind
A few years back, I consulted for a mid-sized e-commerce company that wanted to build a recommendation engine. They had a $500k budget. My first suggestion? Put only $150k into the model and data science. The CTO was skeptical — he thought I was padding my hours. But we went ahead.
We spent $100k on cleaning their messy product catalog and building real-time clickstream pipelines. Another $50k went into training the engineering team on how to serve models via REST APIs. $30k on A/B testing infrastructure. The remaining $70k? Cloud compute and a small team of data scientists.
The model itself was a straightforward collaborative filtering algorithm — nothing groundbreaking. But because everything around it was solid, it achieved a 20% lift in click-through rate within two months. The model wasn't the hero; the infrastructure was.
I've seen the opposite happen too. A healthcare startup spent 90% of their $2M budget on a state-of-the-art NLP model. The model was incredible. But they had no way to integrate it with legacy electronic health records, no data governance, and no one to explain the outputs to doctors. The project was shelved after a year. That's the 30% rule in action.
3 Common Mistakes That Kill AI Projects
1. Underestimating Data Work
Data isn't just a resource — it's a liability if it's bad. I've seen teams spend weeks tuning hyperparameters while ignoring that 30% of their training labels are wrong. The 30% rule reminds you to budget for data auditing and cleaning. A good rule of thumb: budget for twice as much data engineering as you think you need.
2. Ignoring the Ops Side
Deploying a model is the beginning, not the end. You need monitoring for concept drift, retraining pipelines, rollback strategies. I've personally woken up at 3 AM because a model started outputting nonsense. Without MLOps, you're flying blind. Allocate at least 15% to infrastructure and monitoring.
3. Forgetting Human Adoption
Users don't trust black boxes. I once built a model that predicted machine failures with 95% accuracy, but the maintenance crew refused to use it because they couldn't understand why. We had to spend three months building a simple explanation interface and training the crew. That's a people/change management cost — part of the 20% people budget.
How to Apply the 30% Rule to Your Next Project
Here's a step-by-step approach I use with clients:
- Step 1: Estimate your total budget. Be realistic — include cloud costs, personnel, and external tools.
- Step 2: Pull out 30% max for the model development and AI technology. This includes salaries of data scientists (if they're only building models).
- Step 3: Allocate the remaining 70% explicitly: data (25%), infrastructure (15%), training (20%), governance (10%). Use the table above as a template.
- Step 4: Check for hidden gaps. For example, if your team has no MLOps engineer, reallocate some of the model budget to hire one.
- Step 5: Create a risk register. What if data quality is worse than expected? If your model fails in production? Budget a contingency of 10% on top of the 30% technology portion.
I always tell teams: the 30% rule isn't about being cheap on technology. It's about being honest about what it really takes to make AI deliver value. You can still use the latest transformer models — just make sure you have the surrounding ecosystem to support them.
FAQ: Quick Answers to Tricky Questions
This article draws from my personal experience leading AI product teams since 2018. I've fact-checked the budget percentages against the 2023 McKinsey report on AI implementation costs.
