Here's a number that surprises almost every non-technical stakeholder the first time they hear it: 80% of the time spent on an AI project goes to data. Not to model design, not to training, not to deployment. Data. Collecting it, cleaning it, annotating it, verifying it, and iterating on it when the model reveals that something is wrong.
The 20% the part everyone budgets for is the model itself.
This isn't a new insight. Andrew Ng has been making this argument for years under the banner of "data-centric AI." But despite being well-documented, it remains one of the most persistent planning failures in AI development. Teams chronically over-invest in model sophistication and chronically under-invest in data quality.
The 80/20 data reality isn't a problem to be optimized away. It's a structural truth about how AI development works and the teams that plan around it ship better products than the teams that don't.
Why the estimate is always wrong
"The data will be ready in two weeks." This sentence has derailed more AI projects than any technical challenge. Here's why it's almost never true.
Data collection is slower than expected because the data you need often doesn't exist in a clean, accessible form. It's scattered across systems, in incompatible formats, behind legal or privacy barriers, or it simply hasn't been collected at all and needs to be generated from scratch.
Cleaning takes longer than expected because the problems you find while cleaning reveal more problems you didn't know existed. Every batch of data is a new adventure in inconsistency.
Annotation takes longer than expected because annotation is fundamentally a labor-intensive process. There is no shortcut. At five seconds per example fast, for anything non-trivial annotating one million examples takes 1,400 person-hours. Annotation also has error rates that require review, which adds more time. Review finds more errors, which requires rework. The cycle repeats.
And then the model trains, and the error analysis points back to data gaps, which means more annotation, which starts the cycle again.
The rework multiplier
Here's what makes under-investment in data quality especially costly: data problems discovered late are far more expensive to fix than data problems discovered early.
If you find an annotation inconsistency during QA before training, you fix it in a few hours. If you find it after the model is deployed and failing in production, you're looking at re-annotating data, retraining the model, re-running evaluation, re-deploying, and potentially explaining to customers why your AI behaved unexpectedly.
The cost differential between early and late discovery can be 10x or more.
This is why treating data preparation as a fast precursor to the "real work" of model training is such an expensive mistake. The money you save by cutting corners on data preparation, you spend back multiple times when the model fails.
What budgeting for the 80% actually looks like
Budgeting for data means staffing an annotation team or working with a specialized annotation partner. It means investing in annotation tooling that enforces consistency at scale.
It means building QA processes into the data pipeline from the start. It means allocating calendar time for the multiple rounds of data improvement that follow the initial model evaluation.
It means treating your dataset as a living product that will continue to evolve throughout the AI product's lifecycle.
None of this is glamorous. All of it is necessary.
Real-World Relevance
For product managers and engineering leaders building AI products, the 80/20 rule is a resourcing and planning principle. If your AI project timeline treats data preparation as a short precursor to model development, it's wrong.
If your budget allocates the majority of resources to model infrastructure and a small fraction to data, it's probably inverted. The teams that ship reliable AI are the ones who internalized this reality early and built their plans around it.
The best AI teams don't fight the 80/20 rule. They build around it.
They treat data as the primary engineering challenge, invest in it accordingly, and discover that the "harder problem" the model becomes significantly easier when the data is actually good.
Here's a number that surprises almost every non-technical stakeholder the first time they hear it: 80% of the time spent on an AI project goes to data. Not to model design, not to training, not to deployment. Data. Collecting it, cleaning it, annotating it, verifying it, and iterating on it when the model reveals that something is wrong.
The 20% the part everyone budgets for is the model itself.
This isn't a new insight. Andrew Ng has been making this argument for years under the banner of "data-centric AI." But despite being well-documented, it remains one of the most persistent planning failures in AI development. Teams chronically over-invest in model sophistication and chronically under-invest in data quality.
The 80/20 data reality isn't a problem to be optimized away. It's a structural truth about how AI development works and the teams that plan around it ship better products than the teams that don't.
Why the estimate is always wrong
"The data will be ready in two weeks." This sentence has derailed more AI projects than any technical challenge. Here's why it's almost never true.
Data collection is slower than expected because the data you need often doesn't exist in a clean, accessible form. It's scattered across systems, in incompatible formats, behind legal or privacy barriers, or it simply hasn't been collected at all and needs to be generated from scratch.
Cleaning takes longer than expected because the problems you find while cleaning reveal more problems you didn't know existed. Every batch of data is a new adventure in inconsistency.
Annotation takes longer than expected because annotation is fundamentally a labor-intensive process. There is no shortcut. At five seconds per example fast, for anything non-trivial annotating one million examples takes 1,400 person-hours. Annotation also has error rates that require review, which adds more time. Review finds more errors, which requires rework. The cycle repeats.
And then the model trains, and the error analysis points back to data gaps, which means more annotation, which starts the cycle again.
The rework multiplier
Here's what makes under-investment in data quality especially costly: data problems discovered late are far more expensive to fix than data problems discovered early.
If you find an annotation inconsistency during QA before training, you fix it in a few hours. If you find it after the model is deployed and failing in production, you're looking at re-annotating data, retraining the model, re-running evaluation, re-deploying, and potentially explaining to customers why your AI behaved unexpectedly.
The cost differential between early and late discovery can be 10x or more.
This is why treating data preparation as a fast precursor to the "real work" of model training is such an expensive mistake. The money you save by cutting corners on data preparation, you spend back multiple times when the model fails.
What budgeting for the 80% actually looks like
Budgeting for data means staffing an annotation team or working with a specialized annotation partner. It means investing in annotation tooling that enforces consistency at scale.
It means building QA processes into the data pipeline from the start. It means allocating calendar time for the multiple rounds of data improvement that follow the initial model evaluation.
It means treating your dataset as a living product that will continue to evolve throughout the AI product's lifecycle.
None of this is glamorous. All of it is necessary.
Real-World Relevance
For product managers and engineering leaders building AI products, the 80/20 rule is a resourcing and planning principle. If your AI project timeline treats data preparation as a short precursor to model development, it's wrong.
If your budget allocates the majority of resources to model infrastructure and a small fraction to data, it's probably inverted. The teams that ship reliable AI are the ones who internalized this reality early and built their plans around it.
The best AI teams don't fight the 80/20 rule. They build around it.
They treat data as the primary engineering challenge, invest in it accordingly, and discover that the "harder problem" the model becomes significantly easier when the data is actually good.