Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think
Thursday, September 17, 2026
The hype around large language models and viral demos is drowning out a clearer opportunity: applying AI where real business processes run every day.
By Jordan Reed · Published · 8 min read
Introduction
Many organizations chase the latest model and miss the work that actually creates value—cleaning data, automating repetitive tasks, and embedding models into workflows.
Andrew Ng reframes the game: scale narrow, reliable AI across operations rather than building a single, general-purpose marvel.
This post will show what Ng highlights as the highest-leverage areas, how to evaluate projects, and a practical checklist to start capturing those gains.
Where does Andrew Ng say the biggest opportunities in AI are?
Ng says the biggest opportunities are in automating specific, repeatable tasks inside established industries, not in building general-purpose consumer-facing models.
According to the Silicon Valley Girl video summarizing Andrew Ng’s talk, Ng emphasizes industrialization: applying ML to real workflows, investing in data plumbing, and deploying small models where they reliably replace manual effort. This approach shifts effort from chasing model novelty to operational reliability and measurable impact.
Comparison Table: Typical AI Focus vs. Andrew Ng’s Recommended Focus
| Area | Common Industry Focus | Ng’s Recommended Focus |
|---|---|---|
| Model Choice | Build the biggest, newest model | Choose compact, robust models tailored to the task |
| Value Capture | Consumer attention and demos | Internal efficiency, cost reduction, process speed |
| Data Work | Experiment with public datasets | Invest in domain data pipelines and labeling |
| Deployment | Pilot prototypes that rarely scale | Repeatable deployment with monitoring and maintenance |
How can organizations spot the highest-leverage AI projects?
Pick tasks that are frequent, well-defined, and currently manual—but where decisions are regular enough to be learned from data.
Ng recommends prioritizing tasks where automation replaces repetitive human effort with clear KPIs (time saved, error reduction, throughput). These projects scale because they produce short feedback loops and measurable improvements.
Why does Ng prefer domain-specific models and pipelines?
Domain-specific solutions reduce brittleness and lower integration costs compared with general models that need heavy adaptation.
Smaller models trained on relevant data often perform better on specific tasks, require less compute, and are simpler to monitor and maintain—making long-term production use far more realistic.
What mistakes do teams commonly make when pursuing AI?
Teams chase novelty, under-invest in data infrastructure, and treat deployment as an afterthought instead of the core problem.
The video highlights that turning prototypes into reliable systems requires continuous data ops, monitoring, and a plan for incremental improvement—neglecting these steps kills most projects.
Real-World Example: Applying Narrow AI to an Operational Process
Company: Example omitted; the video discusses industry-focused automation broadly rather than naming a single company.
Approach: The example described involves identifying a high-volume manual task, building a focused model, integrating it into the workflow, and measuring operational KPIs.
Results in short term
- The video reports qualitative outcomes: faster decision cycles, fewer manual errors, and management confidence in scaling the automation.
- No specific percentages or company names were provided in the source video; results are described as consistently positive when the deployment and data pipeline are robust.
AI Opportunity Checklist
Planning
- Identify repeatable tasks with clear inputs and outputs.
- Estimate frequency and current human time cost for the task.
Execution
- Build a small, domain-tuned model using in-house or labeled domain data.
- Design the integration so humans can override and provide feedback.
Track + Optimize
- Instrument KPIs: time saved, error rates, throughput.
- Set up monitoring, retraining pipeline, and incident playbooks.
Glossary
- Domain-specific modelA model trained or adapted for a narrow use-case within a particular industry.
- Data pipelineThe systems and processes that collect, clean, label, and deliver data for training and inference.
- DeploymentThe process of moving a model from prototype to production with monitoring and maintenance.
- KPIsKey performance indicators used to measure impact, such as time saved or error reduction.
- Automation choke pointA recurring process step that limits throughput and is a prime candidate for automation.
External Resources
- Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think — Silicon Valley Girl (YouTube)Primary source summarizing Ng’s argument.
- deeplearning.aiAndrew Ng’s organization, useful for courses and practical guidance on applied ML.
- Stanford AI LabA resource for foundational research and papers relevant to applied AI strategies.
FAQ
Should companies stop investing in large language models?
No. Companies should balance investments in foundational models with practical, domain-specific deployments.
Large models add capability, but Ng argues they deliver most value when paired with domain data, integration, and repeatable workflows.
Is building pipelines more important than model accuracy?
Often, yes — robust pipelines are the multiplier that turns marginal model improvements into real production gains.
If models can't be fed clean data or retrained efficiently, higher accuracy on paper rarely translates into improved outcomes.
Can a small company realistically follow Ng’s advice?
Yes. Small teams gain the most from targeting high-frequency manual tasks where automation quickly reduces cost and error.
Start with one process, measure impact, and expand if the pipeline and monitoring produce repeatable benefits.
How long before an internal AI project shows value?
It depends on data readiness and integration complexity; many targeted pilots show operational improvements within weeks to months once data is sufficient.
Short, well-scoped pilots with clear KPIs shorten the time to measurable impact.
Does Ng recommend open-source or proprietary models?
Ng emphasizes pragmatism: choose the model type that fits the cost, latency, and data constraints of the task.
Open-source models can be faster to adapt and cheaper to run; proprietary models may offer turnkey capabilities—match choice to operational needs.
Final Summary
Quick Recap:
- Focus AI effort on automating high-volume, repeatable tasks within existing workflows.
- Invest in data pipelines, monitoring, and deployment, not just bigger models.
- Measure real KPIs and scale the solutions that produce consistent operational wins.
Written by
Jordan Reed — Cybersecurity & Technology Writer
Jordan Reed is a cybersecurity and technology writer focused on digital security, data privacy, and emerging security challenges. With 7+ years of experience in cybersecurity research and technology content, Jordan explains complex security concepts in practical, accessible language. His work covers cyber threats, data protection, privacy, AI security, and best practices for staying secure in an increasingly connected world.
Credentials: Cybersecurity Research & Technology Content Specialist
Expertise: Cybersecurity, Data Privacy, Digital Security, AI Security, Risk Management
7+ years of experience