{"id":4410874,"date":"2026-09-29T17:51:35","date_gmt":"2026-09-29T15:51:35","guid":{"rendered":"https:\/\/noesysai.com\/?p=4410874"},"modified":"2026-09-29T17:51:35","modified_gmt":"2026-09-29T15:51:35","slug":"the-distance-from-using-ai-to-creating-value","status":"publish","type":"post","link":"https:\/\/noesysai.com\/index.php\/2026\/09\/29\/the-distance-from-using-ai-to-creating-value\/","title":{"rendered":"The distance from simply using AI to actually creating value"},"content":{"rendered":"

By Ga\u00ebl Gioux and Theos Evgeniou<\/em><\/p>\n

In McKinsey\u2019s State of AI survey published this August, 89 percent of respondents say their organization uses AI regularly in at least one business function. The figure is often quoted as a sign that adoption is succeeding in those organizations. Our interpretation is different. It only tells us that somebody, somewhere in those companies, has access to a tool and might be using it to draft an email or translate an article. One can really talk about success in adoption when actual value is created, often when the work itself changes around the technology, and current evidence suggests that has barely started.<\/p>\n

In the same survey, 37 percent say AI has contributed to their EBIT, essentially at the same level that figure stood a year earlier, and about 6 percent attribute 5 percent or more of their EBIT to it. Our NoeSysAI corpus of 2,573 published AI use cases, the one behind our monthly AI Brief<\/a>, points the same way. Forty-two percent of them apply AI to a single task. About one in five reaches across a whole process, and those are four times as likely as single-task cases to require heavy builds. Usage spreads fast because a license is easy to buy. Adoption that creates value is slower because it asks a company to change how it works.<\/p>\n

At NoeSysAI, we spend most of our time closing the gap between usage and actual value creation. In the work we ran this year, in industry and in services, the companies that could show something at the end had done the same three things, and none of the three had much to do with which model they picked.<\/p>\n

What follows discusses each of these three key practices we identified with our customers. We also draw from our corpus of 2,573 AI cases to compare our clients\u2019 progress with what happens out there.<\/p>\n

One. The business case you write first can be a lot larger than you think<\/h2>\n

This pattern comes up often. Consider a firm that sells through channels it does not fully control. Somebody checks the listings whenever there is time, but not often enough, hence there is real leakage \u2013 that nobody has ever put a figure on. The \u201cAI opportunity\u201d is obvious: use AI to do this work automatically and continuously.<\/p>\n

But that framing often sets the ambition and potential value creation low, while underestimating the human involvement and monitoring necessary to deliver this value. The first question to ask is \u201cwhat else can using AI enable here?\u201d, including activities we could not do before. A machine watching listings runs at a cadence and a breadth no person is able to match, so the first thing that changes is coverage, and the value shows up as significantly reduced leakage and at a fraction of the initial cost. The second question is how the human job changes. Supervision of a monitoring system is not the task that was being done before, and it needs different skills on a different rhythm. The third is the one nobody put a value on at the start. Watching a market continuously produces a record of that market, and that record turns out to be worth something to marketing, to pricing and to whoever is supposed to know what competitors are doing.<\/p>\n

So the business case moves through several stages, and companies that stop at the first one both miss the full AI potential and measure the wrong indicator.<\/strong> In McKinsey\u2019s 2026 survey, the companies that attribute a significant share of their EBIT to AI, about 6 percent of respondents, are also the ones that changed how the work is done: nearly three quarters of them say they have fundamentally redesigned workflows because of AI. Among all the other companies surveyed, only about one in four has done so. An earlier edition tested 25 organizational practices and found that workflow redesign had the biggest effect on EBIT impact. What we have described above is what redesign looks like from the inside.<\/p>\n

Our NoeSysAI corpus seems to confirm both the narrower focus on productivity and the lack of clear results. Of the 2,573 use cases we have, only one in ten states both a starting point and an end result. Productivity is the headline value lever in 63 percent of them but only 22 percent report a measured or independently verified outcome. The rarer cases built around cash, such as working capital or cost that actually leaves the company, report a measured result more often: 34 percent of them. One likely reason is that a financial effect is easier to measure than, say, time saved.<\/p>\n

This is also where datasets like our corpus help, and also where their limits become clear. They are useful as a source of starting ideas: more than 2,500 projects in our data that other companies have implemented, which can be searched by industry, function and the data they used, help our clients get started. They also provide a first benchmark: the order of magnitude achieved by a comparable case and what it took to build. But the real value we create with our clients sits further along, which rarely makes it into an article, case study, benchmark, or database<\/strong>.<\/p>\n

Two. Precise answers need structured data, and most of what a company knows is still in documents<\/h2>\n

A customer service operation we worked with wanted an assistant that could answer customers\u2019 questions about their contracts. The first version read the contracts as prose and produced answers that were plausible but sometimes wrong, which in that setting is worse than no answer at all. Getting it to the required precision meant extracting the features of each contract and holding them as structured fields, so the assistant could look up an entitlement instead of inferring one.<\/p>\n

That work had a second effect nobody had planned for<\/strong>. Writing down what each clause means forces a single interpretation, and the moment we asked for one, it turned out that different departments had been reading the same clause differently for years. The disagreement had been invisible because each interpretation was defensible on its own and was aligned with each department\u2019s own, sometimes conflicting, objectives. The data preparation improved the assistant, and it also improved the company\u2019s understanding of its own contracts, which we think was the more durable outcome.<\/p>\n

In McKinsey\u2019s 2025 survey, as reported in Stanford\u2019s 2026 AI Index, 74 percent of respondents counted inaccuracy as a relevant AI risk, up from 60 percent a year earlier, which puts it level with regulatory compliance and ahead of cybersecurity. In our experience, inaccuracy rarely starts with the model. It starts with knowledge that sits in documents and has never been turned into governed data.<\/strong> Our corpus of AI cases also supports this. Of the cases that name the systems the work touches, three in four reach into unstructured material. Unstructured here means documents and messages, as opposed to the tables of an ERP or a data warehouse. And the effort follows that mix: when a use case has to combine documents with structured data, about four builds in ten are heavy, against one in five when it stays inside structured data.<\/p>\n

That is a real cost for any AI program. That is also where competitive advantage lies.<\/strong> With models available to everyone on the same terms, what remains to differentiate on is the knowledge only you hold, and in most companies that knowledge is sitting in documents nobody ever had to formalize.<\/p>\n

AI is also a very practical tool to approach the challenge of structuring and governing company knowledge. In the case above, the features were extracted from the contracts with AI and checked by people, a job that would have been hard to do by hand at a reasonable cost. Our NoeSysAI corpus shows the same thing at scale. Of the cases that touch documents, about one in seven uses AI to pull structured content out of them first, and those report a measured outcome 36 percent of the time, against 22 percent for the cases that point AI at the documents directly. It also matches what we saw in the field: data precision and quality come from the structuring, and AI can now do much of that with appropriate processes and guidance.<\/p>\n

Three. Somebody is deciding what the machine does, and it should be you<\/h2>\n

In most companies, the split between what the machine does and what the people keep is being made by whoever configures the machine. It is rarely decided deliberately or documented, and almost never revisited, and the people accountable for the outcome often learn about it afterward.<\/p>\n

We treat that split in AI projects as a decision to be made at every step of a process, with four possible answers.<\/strong> Delegate the step to the machine. Augment the person doing it. Keep it human. Forbid it. At one of our industrial clients, choosing the tooling and settings for a machine sits definitively in the second: the system proposes a setup and a person confirms it, because a wrong setup is paid for in scrapped material and lost machine hours, or worse, in operator safety. Elsewhere, signing and sending a letter is kept entirely human, by choice, because that signature commits the company.<\/p>\n

Every step where a person still decides is also a place to learn<\/strong>. When the machine augments, each human override potentially reveals something the system failed to weigh properly. When a step is kept human, or forbidden, a system can still run alongside and record what it would have proposed. Documenting why each decision went the way it did, and comparing it with that proposal, builds a record of what drives the judgment. That record does two things. It makes the eventual case for changing the stance, if it should change. And it tells us what a person needs to know to do the job, which can be equally valuable.<\/p>\n

Our NoeSysAI corpus is revealing here. Across 2,403 use cases, 57 percent augment a person and 43 percent hand the step over. How much gets handed over varies a lot with where the work sits. In customer operations, 58 percent of the cases give the step to the machine. In legal and compliance the figure is 29 percent, and in healthcare 28 percent. We interpret this as reflecting the cost of error more than the difficulty of the task: an error in front of a patient or a regulator costs more than one in a customer reply.<\/p>\n

How the work starts matters too. When a step starts automatically, because an invoice arrives or an alarm fires, AI takes it over in 53 percent of the cases. When a person has to ask for it, as with a question put to an assistant, the figure drops to 35 percent. Work that already runs inside a process is probably easier to hand over than work that starts in someone\u2019s head.<\/p>\n

However, one needs to read such data with care. A corpus of published use cases can only show where AI was used, not where companies deliberately chose not to use it. A step a company decided to keep human, or to forbid, produces no deployment and therefore no evidence. So no benchmark and no market scan will ever tell you what your peers chose to protect.<\/strong><\/p>\n

How these numbers were produced<\/h2>\n

The corpus held 2,573 real use cases drawn from public sources on 23 September 2026. It grows every day, so every figure here is frozen at that date. Classification is automated and manually checked only in part. The pool also inherits the bias of anything published: companies write up what worked, vendors write up what sells, and failures are of course underrepresented (if represented at all). These percentages describe what gets written about, and only indirectly what gets built.<\/p>\n

The three questions<\/h2>\n

If you take one thing from this, take the three questions rather than the statistics. In our experience they are what adoption is made of, and a company that can answer all three has moved well past using a tool.<\/p>\n

\n

1.<\/strong>\u00a0\u00a0What would this be worth if we stopped counting saved hours and looked instead at what the work produces, some of which may be unexpected or previously impossible?<\/p>\n

2.<\/strong>\u00a0\u00a0Which unstructured knowledge, in documents, messages or people\u2019s heads, does this depend on, and who governs it, from agreeing on what it means to keeping it current and usable?<\/p>\n

3.<\/strong>\u00a0\u00a0Who decided which steps the machine handles versus which a human does, and where is that decision recorded?<\/p>\n<\/div>\n

Ga\u00ebl Gioux is Managing Partner at NoeSysAI. Theos Evgeniou is co-founder of NoeSysAI and Professor of Technology and Business at INSEAD.<\/em><\/p>\n

Sources<\/h2>\n