What is the 10-20-70 rule for AI?

Strip the jargon out of most AI strategy decks and you find one stubborn idea hiding underneath. The 10-20-70 rule. Originally popularised by BCG, it suggests that the value you get from any AI initiative breaks down roughly like this:

• 10% comes from the algorithms and models themselves

• 20% comes from the underlying technology and data

• 70% comes from people, process, and change

  
  

Read it again. The model is the smallest slice. The thing every vendor demo, every keynote, every "we've just integrated GPT-X" press release wants to talk about, the 10%. The unglamorous work of redesigning how people actually do their jobs, retraining them, redrawing process flows, killing legacy reporting, sorting out who owns what when the AI gets a decision wrong, all of that lives in the 70%.

If that sounds suspiciously like every digital transformation lesson of the past thirty years, that's because it is. AI hasn't repealed human nature. The reason CRM projects under deliver, the reason RPA programmes plateau at thirty bots, the reason your last data lake is still half empty, it's the same reason most AI pilots never make it to production. The technology was the easy bit.

Why most companies invest in the inverse

Walk into the average AI steering committee and watch where the money goes. Licenses. Cloud credits. A new vector database. A pilot with a flashy vendor. Maybe an internal "Centre of Excellence" with three data scientists and a Confluence page. Total spend, heavy. Time and attention on the 70%, almost nothing.

There are reasons this happens, and most of them are uncomfortable.

First, the 10% is tangible. A model has a name, a benchmark, a logo. You can put it on a slide. You cannot put "we got the regional ops managers to actually trust the recommendations" on a slide in the same way, even though it matters ten times more.

Second, the people selling AI are mostly selling the 10% and the 20%. Hyperscalers sell compute. Foundation model vendors sell tokens. Big SIs sell implementation. Almost nobody in that supply chain has a commercial incentive to tell you the honest answer, which is that the model is rarely the bottleneck.

Third, boards reward visible movement. A signed deal with a named AI vendor reads as progress. A six month effort to redesign a claims handling process so it can actually absorb AI assistance reads as overhead. One gets a press release. The other gets a budget review.

The bit nobody wants to admit

I spent the last few years running an intelligent automation delivery partner. We saw the same pattern again and again. Clients would arrive convinced they had a technology problem. Three months in, the technology was working fine. The thing that wasn't working was that nobody had decided who owned the exception queue, or how performance reviews would change once sixty percent of the casework was automated, or what to do with the team leader whose entire role had just evaporated.

The fix was never another model. It was someone willing to sit in difficult rooms and make decisions about people and process. That is the 70%. It is also, by the way, the bit that consultancies tied to a particular tech stack find hardest to deliver honestly, because the answer is sometimes "you don't need more technology, you need to use what you've already got better."

What this means if you're a leader staring at an AI plan

Three practical reframes.

One, when a proposal lands on your desk, look at the budget split. If more than 30% of the cost is going on technology and data, you are probably about to recreate someone else's pilot. Push back.

Two, ask who owns the 70% before you sign anything. Not in theory. By name. If the answer is vague, the project is already in trouble.

Three, be suspicious of any partner whose revenue depends on convincing you the 10% is the hard part. They are not lying to you. They are just optimised for a different question than the one you actually need answered.

The 10-20-70 rule isn't really a rule about AI. It's a rule about where value actually sits in any technology programme, restated for an audience that keeps forgetting. The companies that will get serious returns from AI over the next five years are the ones taking the boring 70% seriously now, while everyone else is still arguing about which model to license.