Many companies are using artificial intelligence for the sake of artificial intelligence: as an obligation, as a box to tick in the strategic plan or as proof that they are not falling behind. Sometimes it even becomes corporate entertainment: a flashy demonstration, a brainstorming session, an assistant that surprises for a few minutes and a collection of pilots that nobody really knows how to connect with the business.
This does not mean that experimenting is useless. Experimenting is necessary, especially with technology whose capability changes so rapidly. The problem begins when we confuse activity with progress and assume that using more AI equals better transformation of the company.
In McKinsey's global survey on the state of AI in 2025, 88% of participants stated that their organisation already used AI regularly in at least one function. However, only 39% reported any impact on operational results at a corporate scale, and the group attributing a significant impact to AI was much smaller. The preliminary findings from Project NANDA, an initiative linked to MIT, describe a similar gap: a lot of experimentation, but little transformation when the tools do not fit into the workflows and the daily reality of the organisation. The distance between use and impact remains enormous, and it is not explained solely by the quality of the models: many companies start the conversation in the wrong place.
The Wrong Question
The usual question is "where can we put AI?", when the truly useful question would be "where does the business stop flowing?". The difference seems semantic, but it completely changes the type of project that appears on the other side.
If you start with the technology, you will find almost irresistible possibilities. You can turn an idea into a functional prototype in a few days, build a configurator that would have required months of development before, create an assistant capable of reading thousands of contracts, generating product images in seconds, preparing a complete business proposal from a conversation, writing hundreds of articles, simulating scenarios, detecting defects in a photograph or giving a small team capabilities that until recently seemed reserved for a large company, a huge team or a quantity of resources you do not have. Some of these applications border on magic when seen for the first time. But something being technically possible, fast or spectacular does not mean it will improve the business.
Accelerating a Task is Not Improving the System
A company is not a collection of independent tasks, but a system of flows connected by which information, materials, customer commitments, documents, validations, decisions and exceptions circulate. The result does not depend on the speed of each piece separately, but on how the whole works and at what point it regulates its capacity at any given moment.
Imagine a company where preparing a business offer takes too long. The technological answer seems obvious: use AI to draft offers faster. The tool works, the draft appears in minutes and the demo is flawless, but the offers continue to arrive late to the client because before sending them, the margin must be validated. If we keep asking, we discover that this validation is delayed because technical information is missing. The information is missing because the customer requirements are spread across emails, commercial notes and conversations that do not reach operations in a structured way. And this happens because no one has designed the complete flow of information from the commercial conversation to the final decision.
The company thought it had a drafting problem, when in reality it had a flow and accountability problem. AI had accelerated a visible part, but not the part that regulated the result. It could even worsen the situation: more drafts prepared meant more documents waiting for validation, more changes, more interruptions and a longer queue before the same critical point.
That an improvement is real does not mean it improves the overall result.
This pattern repeats in many ways. Order entry is automated, but the limit is in its validation; more commercial opportunities are generated, but the team cannot follow up properly; more reports are produced, but no one has the authority to make the decision they suggest; incidents are classified faster, but resolution depends on an overloaded person; or content is created at high speed when the real limit is distribution, trust or demand. In all these cases there is a local improvement, but not necessarily a business improvement.
The difference can be observed even within the same company. Doritos, a PepsiCo brand, developed with AI a programme for PC gamers that detected and removed chewing noise picked up by the microphone from voice chat during online matches. To train it, the company analysed over 5,000 crunching sounds. It was an ingenious and very visible brand activation, but the company did not publish any improvement in sales, margin or capacity associated with it. Instead, when PepsiCo applied AI and digital twins to the design of plants and warehouses, it stated that it had increased throughput by 20% and reduced necessary investment between 10% and 15%.
BMW offers a similar contrast: it used AI to project on a Serie 8 works created from 50,000 images, without associating the initiative with a business metric; at its Regensburg plant, a predictive maintenance system prevents more than 500 minutes of assembly downtime each year. In all cases there was AI. Only in the second could the change in a variable regulating the system be observed.
An Inefficiency is Not a Constraint
Here appears an important distinction: an inefficiency is not the same as a constraint. A company can have dozens of slow tasks, uncomfortable tools and improvable processes, but not all these problems are limiting its capacity to sell, produce, deliver, collect, decide or grow.
The constraint is the point that sets the pace of the system. It can be a machine, but it can also be a validation, data that never arrives correctly, an internal rule, a person with scarce knowledge, a decision no one wants to make, insufficient commercial capacity or even lack of demand. Improving anything else can save time or effort; improving that point changes what the system is capable of achieving.
Intervening at the Point Regulating the Flow
That is why "what can we do with AI?" is too broad a question. It is better to replace it with another much more demanding one: "what is currently limiting an important result?". The temporal nuance is also essential, because a constraint is not eternal. When it is resolved, the limit shifts and a solution that works can make visible the next problem.
If we automate document review, perhaps the new limit appears in approval. If we speed up offer preparation, it might appear in delivery capacity. If we reduce billing errors, perhaps we discover that the main problem lies in how commercial agreements are formalised. Improving a company does not consist of installing a solution and considering the job done, but of observing how the flow changes after intervening.
Before adding technology, it is also useful to understand why the critical point performs below its capacity. Perhaps it does not need more power, but fewer interruptions, complete information, a clear separation between normal cases and exceptions, explicit decision criteria or that the rest of the organisation stops sending it faulty, incomplete or out-of-sequence work. Only then does it make sense to ask if its capacity needs expanding and if AI is the best way to do so.
This order matters. If ignored, AI becomes a sophisticated layer over an intact cause; if respected, it can gather dispersed information just before a decision, read documentation that blocks a process, detect deviations before they become incidents, apply repeatable criteria and send only exceptions to a person. It can also preserve context between departments and allow professionals with more specialised knowledge to focus their attention on the cases that truly require judgement. The difference is not so much in the technology used as in the place it occupies within the flow.
Six Questions Before Adding AI
To evaluate an AI opportunity seriously, start with six questions.
1. What result do we want to change?
The objective cannot be "to use AI". It must be expressed as a recognisable business change: reducing response time, increasing capacity, preventing errors, protecting margin, improving conversion, shortening collection or giving more reliability to a decision.
2. Where does the waiting, rework or uncertainty really accumulate?
It does not have to be where there are the most complaints. You must follow the work from beginning to end and observe where it stops, goes back or needs extraordinary intervention.
3. Why does it happen?
The first answer is not enough. "Human error" is not a sufficient cause and "missing data" neither; you must continue until you find the system condition that produces the error, the absence of the data or the dependence on a person.
4. Does that point limit the entire result or is it just a local nuisance?
This is probably the most uncomfortable question, but also the one that prevents investing in improvements that produce an apparent saving without modifying the business result.
5. What is the minimum intervention that can change the flow?
It could be AI, but it could also be an integration, a rule, a better defined responsibility, a change of sequence, mandatory data or the elimination of a step that adds no value.
6. How will we know the system has improved?
We will not know by the number of licences, prompts, agents or documents generated, but by the total time, capacity delivered, errors avoided, margin recovered, decisions resolved or the service received by the client.
The Cost of Getting the Problem Wrong
This way of thinking is especially relevant today. For years, building a technological solution was expensive enough to force companies to choose; now producing a demo, connecting a model or launching a pilot is much easier. That is good news, but it also reduces the cost of getting the problem wrong.
We can optimise more things, faster and with less initial investment. We can generate local improvements throughout the organisation and fill it with assistants that save minutes without modifying any of the variables determining its performance. The less scarce technology is, the scarcer focus becomes.
Recent evidence points in the same direction. Stanford's AI Index 2026 observes that the largest productivity gains are concentrated in structured, measurable jobs with results that can be monitored. Meanwhile, McKinsey's survey shows that organisations achieving the most impact do not limit themselves to adding tools, but redesigning the workflows where AI must operate and connecting the intervention to growth, innovation or business efficiency objectives. They do not use AI for AI's sake; they change a way of working.
Technology is Not the Starting Point
That is why at Monreal & Meadow we do not start by looking for where to place a technology. We follow the work, observe how information circulates, understand where decisions stop and distinguish a symptom from a cause and an annoyance from a constraint. Only then do we decide if the answer is AI, automation, an integration, an internal tool, an operational change or simply stopping doing something.
AI should not be the starting point, but, when appropriate, a precise part of the solution.
Cited Sources
- McKinsey — The state of AI in 2025: Agents, innovation, and transformation.
- Stanford HAI — 2026 AI Index Report: Economy.
- Project NANDA — The GenAI Divide: State of AI in Business 2025, preliminary findings.
- PepsiCo — Doritos Silent.
- PepsiCo — AI and digital twins in plant and supply chain operations.
- BMW — The Ultimate AI Masterpiece.
- BMW Group — Smart maintenance using artificial intelligence.