Why Your AI Workflow Matters More Than Your Model

There's a misconception in the AI world right now. Companies are racing to upgrade to the latest models, convinced that GPT-4, Claude, or the next big release will solve their AI challenges.

But Andrew Ng, the legendary AI researcher behind Google Brain and Coursera, shared data at Sequoia's AI Ascent that should make everyone pause.

The Surprising Data

On a standard coding benchmark (Human Eval), here's what happened:

• GPT-3.5 (zero-shot): 48% accuracy
• GPT-4 (zero-shot): 67% accuracy
• GPT-3.5 with agentic workflow: 95% accuracy

Read that again. An older, cheaper model with the right workflow crushes the flagship model used naively.

What Is an Agentic Workflow?

Most people use AI the way Ng calls "non-agentic": type a prompt, get an answer. It's like asking someone to write an essay without using the backspace key.

An agentic workflow is different. It's iterative:

1. Write an outline
2. Do any needed research
3. Write a first draft
4. Critique your own draft
5. Revise based on feedback
6. Repeat until satisfied

This mirrors how humans actually work. And it turns out AI works better this way too.

The Four Design Patterns

Ng identified four agentic patterns you can implement today:

1. Reflection (Robust, Use Now)

Have the AI critique its own output before delivering. Simply prompt: "Check this code for correctness, efficiency, and style issues." The same model that wrote it can often spot problems when asked to review.

2. Tool Use (Robust, Use Now)

Let AI use external tools: web search, code execution, calculators. This expands what's possible beyond the model's training data.

3. Planning (Emerging)

Let AI break complex tasks into steps and decide the sequence. It can even recover from failures mid-execution. Ng describes watching demos where something failed and the AI agent rerouted around the problem automatically.

4. Multi-Agent Collaboration (Emerging)

Multiple AI "agents" with different roles (coder, reviewer, product manager) collaborating on tasks. Open source tools like ChatDev let you run this on a laptop.

What This Means for Your Business

If you're investing in AI, stop chasing models and start building workflows.

The ROI opportunity: Companies using GPT-4 naively are leaving 30%+ performance on the table. Those who implement agentic patterns with cheaper models may outperform competitors spending more.

The training opportunity: Your team doesn't just need "prompt engineering." They need to understand iterative workflows, reflection loops, and multi-step reasoning.

The strategic opportunity: This levels the playing field. You don't need the biggest AI budget to get the best results.

Start Here

Pick one workflow your team uses AI for today. Add a simple reflection step: before accepting the output, prompt the AI to review and improve it. Measure the difference.

Then iterate from there.