The Ten Twenty Seventy Rule That Reframes Every AI Project

What the Rule Actually Says

The 10/20/70 rule splits the total work of building or adopting an AI system into three rough slices. The first ten percent is the model itself: selecting an approach, fine-tuning or prompting, getting the core logic to produce correct output. The next twenty percent is data: gathering, cleaning, structuring, and keeping it fresh so the model has something reliable to work with. The remaining seventy percent is everything else — integration into real workflows, error handling, user experience, monitoring, security, compliance, and the organizational change that makes people actually use the thing.

I want to be clear that this is not a precise law of nature. I have seen projects where data wrangling ate forty percent of the timeline and others where a bad integration decision turned the seventy into ninety. But as a planning heuristic, it tracks closely with what practitioners report across industries. The pattern holds because the model is increasingly a commodity; the friction lives in making it work inside a messy, human, operational reality.

Why Everyone Overbuilds the Ten Percent

Here is the uncomfortable truth: the model is the most visible, most marketable, most fun piece of the project. It is what you put on the pitch deck. It is what gets the demo working for ninety seconds. So teams pour their energy there — swapping architectures, chasing a slightly better benchmark score, arguing over which approach handles edge cases best. And that work matters, but it is the smallest slice.

I made this mistake in my first year running SEMPITE. I spent weeks obsessing over how our internal visibility scoring handled long-tail queries, tweaking thresholds until I got a headache, while the actual client-facing workflow — pulling data on a schedule, flagging anomalies, writing the summary a small-business owner would actually read — was held together with duct tape and hope. The model was elegant. The system around it was embarrassing. Fixing the seventy took longer than the ten ever did.

The reason this keeps happening is that the ten percent has a clear finish line. You can measure it, benchmark it, show a graph. The seventy percent is diffuse. It lives in Slack threads, in 'wait, what happens when the API times out at 2 a.m.?', in the Tuesday meeting where someone says nobody used the feature last week. There is no elegant chart for that. So it gets deferred, and deferred things do not go away.

What Each Slice Looks Like in Practice

The ten percent, concretely: choosing whether a task needs a generative model, a classification pipeline, or just a well-written rules engine. Prompting or fine-tuning so the output is accurate enough for the use case. Testing against a held-out set of examples you care about. For a small business getting visible in AI search results, this slice looks like making sure the system correctly interprets what a local customer is actually asking and can match it to your content.

The twenty percent: the data that feeds all of that. For my work, it means keeping structured information about a business accurate and current — hours, services, location, reviews, the specific language a customer uses when they search for help. If that data is stale or inconsistent, no model in the world fixes the output. I have seen perfectly capable AI systems produce garbage because the underlying listing was still showing old phone numbers. The model did its job. The data failed it.

The seventy percent: this is where I will be honest and say most of my own work lives. It is the integration layer — connecting a client's website, their review profiles, their local citations into one coherent signal. It is the monitoring: watching whether AI assistants are still recommending the business after an algorithm update. It is the workflow: how does a plumber or a personal brand actually see the alert, understand what changed, and act on it in under five minutes? It is the trust layer: making sure the information surfaced is accurate, attributable, and not accidentally misleading. None of this is glamorous. All of it is necessary.

How This Changes What You Should Spend Time On

If you are a small business owner or a solo professional thinking about 'doing AI,' the 10/20/70 rule tells you to resist the urge to start with the tool. Start with the data. What does an AI assistant need to know about your business, your services, your location, your differentiators in order to recommend you accurately? Can a stranger find that information consistently across every place they might look? If the answer is no, you have not yet reached the model problem. You are still in the twenty.

Then think about the workflow before you think about the technology. Who at your business — if it is more than just you — sees the output? What do they do with it on a Tuesday morning when they are also answering phones and loading inventory? If the answer requires a thirty-minute training session and a printed SOP, the seventy percent is going to eat your launch date. Design for the person, not the platform.

And yes, do the ten percent well. A poorly chosen approach or a prompt that hallucinates specifics will destroy trust faster than an empty page. But budget your time and attention proportionally. I would rather a client have a modest model with a robust pipeline around it than a cutting-edge system that breaks every time someone changes a business hour on their website.

The Part Nobody Puts in the Slide Deck

There is a failure mode I see so often it has its own smell: a team builds an impressive AI feature, demos it to leadership, gets the budget approved, and then quietly abandons it because maintaining the seventy is boring and ongoing. The model does not get tired of running. The monitoring dashboard does not stop generating alerts. The data source changes format on a Friday afternoon in November and nobody notices until a customer calls to complain. The ten percent has an off switch. The seventy does not.

I tell clients this plainly because it argues against my own interest in closing the deal. If your budget is six figures and you plan to spend ninety percent of it on the model and five percent on everything else, I will tell you that is backwards and I would rather lose the project than watch it rot in a corner. The rule exists precisely so you do not fall in love with the smallest piece.

The AI landscape moves fast enough that the specific tools and approaches you pick today will be dated in eighteen months. The operational discipline, the data hygiene, the workflow design — those compound. They are the part that still works when the next generation of models ships. Spend your energy where it accumulates rather than where it gets deprecated.