The AI Transformation Challenge

Make a prediction before you read on. Suppose an organization that is genuinely good at AI has 100 units of money, time and attention to spend on it. How many of those units go into the algorithms themselves, the models that the headlines are about? Many people guess well above ten.

In 2024 Boston Consulting Group surveyed 1,000 senior executives in 59 countries and described the rule that its AI leaders follow: about 10 percent of their resources go into algorithms, 20 percent into technology and data, and 70 percent into people and processes¹. BCG reads the obstacles the same way: about 70 percent of AI challenges stem from people and processes, 20 percent from technology and only 10 percent from the algorithms. The same study found that only about a quarter of companies, 26 percent, had built the capabilities to move beyond proofs of concept and produce tangible value. The other 74 percent had not.

The ratio is a rule of thumb, not a law of nature, but it points the same way as the other evidence in this chapter. AI Is Changing Everything described the gap between using AI and getting value from it. McKinsey’s 2026 global survey, with 1,719 respondents, shows how stubborn that gap is. Eighty percent of respondents who use AI in their roles say it has improved their individual productivity. Yet only 37 percent of all respondents attribute any EBIT impact at all to their organization’s use of AI, about the same share as a year earlier, and the high performers who attribute at least 5 percent of EBIT to AI, and call its impact significant, remain about 6 percent².

Put the two findings together and the shape of the problem is clear. People already have the technology, and it already helps them. What is missing is the rest of the organization. That is the AI transformation challenge.

Six components that multiply

An AI capability creates business value when six components hold at the same time. Technology is the model, the platform, the integration and the monitoring. Data is the information the system needs, available every day with an agreed meaning. People are the named users whose roles, rewards and skills change. Process is the workflow redesigned around the new capability, exceptions included. Governance sets what the system may see and do and who checks it. Leadership decides which outcome matters and holds the authority to change the other five.

The important word is together. In 1990 the economists Paul Milgrom and John Roberts explained, with a formal model, why modern manufacturing changes arrived as a cluster rather than one at a time. Flexible machines, smaller batches, faster product changes and new ways of organizing work were complements: each one raised the return on the others, so adopting one in isolation often disappointed³. AI investments appear to behave the same way, though no study has measured a formula for them. A forecast is worth little if the data behind it is unreliable, worth little if planners are still rewarded for the old plan, and worth nothing if no process acts on it. Read the components as multiplying rather than adding, a way of reasoning rather than a measured law: a zero in any one of them takes the product to zero, however strong the others are.

This explains why the money goes where BCG found it. The complements to a general-purpose technology are mostly intangible: redesigned processes, new skills, new decision rules. Erik Brynjolfsson, Daniel Rock and Chad Syverson point out that these investments are large and poorly measured, so they rarely appear on a balance sheet even though they decide the return⁴. Marco Iansiti and Karim Lakhani make the strategic version of the argument: the firms that pull ahead rebuild their operating model around data and algorithms instead of attaching AI to the one they already have⁵.

Technology still matters. A model that is wrong too often, cannot be integrated or cannot be monitored will stop everything else. But it is the component executives feel most confident about, because it is the most visible and the easiest to buy, and that confidence is the trap. If technology is the only component with a budget and an owner, the organization has made a purchase, not a transformation. The other five are where most initiatives stall, so the rest of this chapter takes them in turn.

Data: demos run on clean samples

AI needs the right records, at the right time, with the right meaning and under the right access rules. A demonstration rarely tests any of this. Someone selects a tidy sample, fixes the obvious errors by hand and leaves out the awkward cases. The demo works, and the organization learns nothing about the data it actually runs on.

AI often exposes problems that were there all along. If two divisions disagree on what counts as an active customer or an open order, people have been reconciling the difference from memory for years. A model cannot do that, and it will not tell you that it is guessing. The useful question for any demonstration is blunt: who cleaned this data, and will they still be doing it every Monday morning? If the honest answer is “an analyst, by hand, for the pilot”, the data component does not yet exist. How to build it is the subject of Data Strategy for AI in Module 04.

People: training does not change the work

The people component is not the same as the talent question in The AI Talent Gap. That chapter asked whether the organization has the right capabilities. This one asks whether the everyday work of the people who must use the system actually changes. Roles shift: the planner who used to build a forecast by hand now judges one that the system prepared, and the supervisor who used to count volume now coaches on exceptions.

Training is necessary, and it is the part programs most readily fund. It is seldom the only part that has to change. People who are still measured on the old targets will return to the old way of working, politely and quickly, because that is what their managers reward. Two questions expose the gap. Who, by name, will use this system on a normal Tuesday? And what will their manager praise them for at the end of that Tuesday? “Everyone, eventually” is not an answer to the first, and “the same as before” is the wrong answer to the second.

Process: a faster broken workflow is still broken

Most functions run some version of the same flow. A request arrives, someone gathers information, someone decides, someone acts and the result is recorded. The tempting move is to drop AI into one step and leave the rest alone. The step gets faster; the queue, the handoff and the approval that caused the delay stay exactly where they were.

The test is the one that AI Leaders vs AI Followers applied to leading organizations: if we designed this process today, with AI available from the first day, would we design it this way? A process you would not design today is not one to automate as it stands. Redesign is hard because it touches handoffs, approvals, job descriptions and habits that belong to several managers at once. That is why teams skip it and hope the tool will absorb the mess. It will not. Someone still has to decide who handles the exceptions, what “done” means, and which step the new capability makes unnecessary.

Governance: early and in proportion

Governance answers a short list of questions. What may the system see? What may it do? Who may use it, and for what? How are its outputs checked, which decisions still need a person, and what happens when something goes wrong? The US National Institute of Standards and Technology treats governance as the function that runs across every other part of managing AI risk, rather than as a gate at the end⁶.

Executives tend to meet governance at one of two extremes. In the first, it is absent while the prototype is built and then appears at the rollout door, where a late review freezes the work for months. In the second, it is so heavy from the first day that nothing can be tried and nothing is learned. The useful position sits between them: enough governance, designed at the start and in proportion to the risk, that a capability can leave the lab without being recalled. Done that way, governance can be what makes speed possible. Governance that arrives only at the production door is not governance. It is a surprise. Risk is the subject of Module 06, and AI Lifecycle Governance in Module 07 shows how to build the system.

Leadership: business leaders own outcomes

Leadership is the sixth component, and it is the one that connects the other five. It decides which problem matters, how much to invest, what risk is acceptable, how success is measured and when to scale or stop. Without it, initiatives fragment. Every function runs a demonstration, no function owns a result, and the organization collects screenshots.

Technology teams are essential. They build, integrate, test and monitor. They cannot own the business result, because they do not control the process, the targets or the staffing that decide it. If nobody with authority over the process is willing to change the process, there is no transformation to lead. AI Leaders vs AI Followers showed that leading organizations put a named senior owner behind their AI work. The component view adds a sharper point: the owner must be the person who can change all six components for one outcome, not a sponsor of the technology alone. Leadership here is not a keynote. It is decision rights. When two managers disagree on whether a forecast is advice or an instruction, the model cannot settle the argument. Only a leader can.

Story: a day in the library

Picture a metropolitan public library system with forty branches and a single delivery service that moves books between them. Readers wait weeks for popular titles at some branches while copies sit unread at others. The library’s small data team has built a model that predicts, a week ahead, which titles will build long waiting lists at which branches. In testing it was usually right, and three months ago it went live. Follow the collections manager through a Tuesday.

At 07:30 the overnight list arrives: 300 titles likely to build long waiting lists next week, with the branches where spare copies sit. The technology is not her problem today. At 08:15 she finds that a third of the titles will not load into the delivery system, because the catalog and the delivery software identify editions differently, and the colleague who keeps the cross-reference is on leave.

At 10:00 she calls the branch managers. They are measured on loans from their own shelves, so sending copies to another branch lowers their numbers. The most experienced of them listens politely and sends nothing. At 12:30 she tries to book the transfers anyway. The delivery process starts only when a reader places a hold; there is no step for a predicted one, and her manual requests are rejected as duplicates.

At 15:00 the privacy lead asks a question nobody had put on the project plan. The model learns from borrowing histories, and libraries treat what people read as confidential. Who approved this use of the data, and on what terms? Until someone answers, the review says, the predictions should not drive decisions. At 17:45 she escalates. The data team owns the model’s accuracy. The branches own their loan targets. Logistics owns the delivery schedule. Nobody owns the outcome the model was built for: shorter waits for readers.

The following week the waiting lists grow exactly where the model said they would. On the project dashboard, the model is green. The model was right. The organization was not ready to act on it, and nobody in the story was careless. Each failure belonged to a different manager, which is why none of them could fix it alone.

The seed and the field

A grower who buys the best seed on the market and sows it in a field nobody has prepared will get a poor harvest. The soil was never tested, the water arrives late and nobody clears the weeds. At harvest, the seed gets the blame. Growers who succeed with a new variety prepare the whole field first: soil, water, timing and care, ready together. AI is the seed, and the organization is the field. The seed is seldom the only thing that fails.

The analogy breaks down in one useful place. Seeds do not improve every quarter, and AI models do. That makes the lesson stronger, not weaker: a better seed sown in the same unprepared field still fails.

What this means for leaders

Four habits follow from the component view. First, when an AI initiative disappoints, assume the model is probably not the main problem until the other five components have been checked. Switching vendors is the most expensive way to discover that the data, the targets or the process were the real constraint. Second, budget for the complements. If almost all of an initiative’s money and named roles sit in technology, it is a purchase plan. BCG’s leaders put most of their effort into people and processes, and the plan should show where that effort goes. Third, design governance with the prototype, not after it, so that the first serious review is not also the first time anyone has asked the questions. Fourth, give every initiative one business owner who has authority over the process it is meant to change, with the AI team as an enabling partner rather than the owner.

Check yourself

1. If the demonstration works, the hard part is done.MythFact
2. AI leaders put most of their resources into people and processes rather than algorithms.MythFact
3. Strengthening one component makes up for a weak one.MythFact
4. The AI team should own the business result.MythFact
5. Governance mainly slows AI down.MythFact
6. In McKinsey’s 2026 survey, most respondents could not attribute any EBIT impact to AI.MythFact

Reflection: find the missing component

What comes next

Six components holding at once can sound like a reason to wait until everything is ready. It is not. Hard does not mean doing everything at once; it means choosing a first move the organization can actually absorb. The next chapter, Where Should We Start With AI?, takes up that choice.