{"slug": "why-adoption-not-the-model-is-the-hard-part-of-ai", "title": "Why Adoption, Not the Model, Is the Hard Part of AI", "summary": "A new field guide argues that the hardest part of AI is not building the model but getting people to trust and adopt it, citing surveys from 2025 and 2026 showing roughly half of employees use unauthorized AI tools and only about a third find sanctioned tools meet their needs. The piece, drawing on studies of humanities and social science researchers, outlines a three-step approach to fostering adoption, emphasizing trust and alignment with existing work practices.", "body_md": "*The model is the easy part. Getting real people to trust it, use it, and let it change how they work is where most AI projects quietly die. This is a field guide to the human half of AI, the half almost nobody plans for.*\n\nThe model was good. That was never the problem. It predicted the thing it was built to predict, it passed every test, it shipped on time, and the dashboard it fed was clean and quick. Six months later, almost nobody was using it. Not because it was wrong, but because the people it was built for did not trust it, did not quite understand it, and quietly went back to the spreadsheet they had always used. The project was not killed by a bug. It was killed by the gap between a working model and a person willing to rely on it.\n\nThat gap is the hard part of AI, and it is the part the industry spends the least time on. We pour years into making models more accurate, and then act surprised when a perfectly accurate one sits unused. The uncomfortable truth underneath most failed AI projects is that the technology worked and the adoption did not. So this piece is about the other half of the job, the human half. What it actually takes to get a real person, with a real workload and real fears, to trust a machine enough to change how they work.\n\nBuilding the model is a surprisingly small slice of the work. The far larger slice is everything wrapped around it, the training, the trust, the change to how a team makes decisions, the slow business of convincing an expert that a number they did not calculate themselves is worth acting on. That last one is the heart of it, and it is worth sitting with.\n\nPicture a seasoned professional in any field where judgement is the job, a doctor weighing a diagnosis, a planner deciding where to build, a loan officer deciding who to trust with money. They have spent decades building an instinct for their work. Now a model hands them a confident answer, and the honest question forming behind their eyes is not is this accurate. It is why should I believe this, and who carries the blame if it is wrong and I went along with it. That is not ignorance or stubbornness. It is exactly the caution that made them good at their job. An AI tool that ignores it does not get adopted. It gets politely worked around.\n\nYou can see this playing out at the scale of whole organisations. Across surveys through 2025 and 2026, roughly half of employees admit to using AI tools their employer never approved, and in one survey the figure among senior executives was far higher again. The reason is quietly damning for how we roll these things out. In the same research, only about a third of employees said the officially sanctioned tools actually met their needs. People are not refusing AI. They are refusing the AI that was handed down to them, and reaching for the one that helps, approved or not. Adoption, it turns out, is not something you can mandate. People adopt what earns their trust and abandon what does not, and no amount of model accuracy changes that equation.\n\nTo make this real, here is one example case. It’s not one specific place, it’s a mix built from several. But every step comes from real studies on how humanities and social science researchers actually start using these tools, and those studies are listed at the end.\n\nA research team has a huge collection of documents. Millions of pages, going back decades, way more than anyone could ever read by hand. Someone gives them a powerful AI tool. On paper, it does exactly what they need. Months later, almost nobody has used it.\n\nThe tool itself was never the problem. It just spoke the wrong language, both in a technical sense and in how the researchers actually think and work, and they wouldn’t trust an answer they couldn’t check against their own way of doing things. Here’s how someone turns that around, in three steps.\n\n**Picking the tool.** The first step is choosing the right tool, and the smart move is the opposite of what you’d expect. Instead of picking the most powerful option, you pick the one that fits how these researchers already work, even if it looks less impressive. Here’s a simple test: pick it for the least confident person in the room, not the most tech-savvy one. If only the one person in the department who can code can actually use it, it hasn’t really been adopted, it’s just been set aside and ignored. That’s what fair, equal adoption really means, and the research backs this up. In a big study of humanities and social science researchers, what actually got people to use a tool wasn’t how powerful it was. It was how well it matched the way their field already thinks, and whether people they respected were already using it.\n\n**Testing with real people.** The second step is building the tool together with the researchers, not just handing it to them. That doesn’t mean one workshop at the start and calling it done. It means shaping the tool around how they actually work, then testing it in a specific way. You sit next to the person on the team who’s least comfortable with it. You watch them use it. Every time they pause, frown, or misread something on the screen, you write it down and fix it before it goes any further. The question isn’t “does the software work.” It’s “can a nervous professor get an answer they actually trust.” This matches what researchers really do, they only trust an AI’s answer once they can check it against their own methods, so the tool has to earn its place in their work instead of replacing it.\n\n**Sticking to ethics.** The third step runs underneath the other two, from day one. An old archive can hold sensitive material, personal records, unpublished work, culturally important items. So privacy and ethics aren’t just a checkbox before launch. They’re rules that stay in place through every version of the tool. Sensitive data stays where it’s allowed to be. A person checks the answers that matter. There’s a record of what the tool did and why. And here’s the part people miss: being careful isn’t the enemy of getting people to use the tool. It’s actually part of what earns the trust that makes people want to use it. A researcher relies on a tool much more easily when they can see their sources are safe, and that they, not the machine, are still responsible for the conclusions.\n\nPut these three together, and the tool that used to sit unused becomes one the team actually reaches for. Look closely at what changed and what didn’t. The tool was just as capable on day one as it was at the end. What was added was fit, testing, and trust. The model itself was never what was holding it back.\n\nHere is the most useful idea in this entire piece, and it reframes almost every stalled rollout. When people resist a new AI tool, the instinct of most organisations is to assume a knowledge gap. They do not know how to use it yet, so we train them. And the training lands with a thud, because the problem was almost never knowledge.\n\nA helpful way to see this is the **ADKAR **model, a widely used change-management framework from the firm Prosci. It breaks any individual’s path through a change into five stages that have to happen in order. Awareness of why the change is happening. Desire to take part in it. Knowledge of how to do it. Ability to do it under real conditions. And Reinforcement to make it stick. The order is the whole point. You cannot skip a stage, and training only serves the third one, Knowledge.\n\nNow the insight. Fear of AI is almost always a Desire problem wearing a Knowledge costume. People often know perfectly well how to click the buttons. What they lack is the willingness, and the willingness is missing because they are afraid. Afraid the tool is there to replace them. Afraid of looking slow or incompetent while they learn it. Afraid of being blamed when the AI makes a mistake they signed off on. Afraid of losing a skill they take pride in. None of that is answered by another tutorial. Throwing training at a fear problem is like handing someone a thicker instruction manual when what they actually said was I am not sure I want to do this.\n\nThis is why the quiet signs matter more than the loud ones. When people go silent, avoid the tool, or say I just have not had the time to learn it yet, that is very rarely a schedule problem. It is Desire, dressed up in a more acceptable excuse, because I am afraid this will cost me my job is a hard thing to say out loud in a meeting. Respond to it with more training and you confirm that the fear was never heard, which quietly makes the resistance worse. The fix is almost embarrassingly human. Name the fear directly. Be honest about what the tool will and will not change for that person’s role. Say the thing everyone is thinking before they have to.\n\nThere is a revealing twist buried in those shadow AI numbers. The people using unapproved AI most enthusiastically are often the senior ones, the executives and managers. The same people who might publicly worry about risk are privately reaching for whatever helps them get through the day. That is not hypocrisy so much as proof that the human relationship with these tools is messy and emotional for everyone, top to bottom, and that it responds to trust rather than to policy.\n\nIf mandates fail and training alone misses the point, what works. The pattern is consistent across the fields that get this right, and none of it is about the model.\n\nIt starts by meeting people where they are, which means building the tool around how a team already works rather than forcing the team to reshape itself around the tool. The most reliable way to do that is to bring the least confident users into the design early, as genuine participants rather than an audience told about the change after the fact. A blunt rule of thumb holds up well here. If the tool works for the person who is most nervous about it, it will work for almost everyone.\n\nTrust, the thing everything hinges on, is built by making the reasoning visible. An answer that arrives from a black box invites the question says who. An answer that can show its working, here is why this route scores high, here are the sources this claim rests on, invites a decision instead. This is also where a human staying in the loop stops being a slogan and starts being the mechanism. When people can see the reasoning and know a person still reviews the high-stakes calls, the tool stops feeling like a replacement standing over them and starts feeling like a capable assistant they remain in charge of. That shift, from replacement to assistant, is most of the battle.\n\nAnd it changes what you measure. The tempting metric is usage, logins and query counts, because it is easy to count. But usage is a treacherous signal. A high number can simply mean people were told to log in, and a low number can mean one confusing step early in the process is blocking everything downstream, not that anyone dislikes the tool. Numbers tell you what is happening and almost never why. The organisations that get adoption right measure the harder thing, whether people’s trust and confidence are actually rising over time, gathered by sitting with users and watching them work, not just reading the logs. The goal was never a login. It was a person who reaches for the tool because they believe it will help.\n\nIt is worth naming that everything above is also what people mean by Responsible AI, approached from the human side rather than the compliance side. Fairness, transparency, accountability, and human oversight are not a separate box to tick after the model is built. They are the very things that make trust possible in the first place. An expert trusts a tool that can explain itself, that treats different groups fairly, that keeps a person accountable for the serious decisions. Build those in from the start and you are not slowing adoption down with ethics. You are creating the exact conditions under which adoption can happen at all. Guardrails built into the system, rather than bolted on at the end, are what let people move quickly and trust the output at the same time.\n\nSo the question to end on is not the one most AI projects ask. It is not did we build an accurate model, and it is not did people log in. Those are the easy questions, and passing them is no guarantee of anything. The real test is quieter and much harder. Do the people this was built for trust it enough to change how they work. Everything that matters about adoption lives in that sentence. The model was always going to be the easy part. The hard part, the human part, is the whole job, and it is the part worth getting right.\n\n[Why Adoption, Not the Model, Is the Hard Part of AI](https://pub.towardsai.net/why-adoption-not-the-model-is-the-hard-part-of-ai-1a6c49180ff2) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.", "url": "https://wpnews.pro/news/why-adoption-not-the-model-is-the-hard-part-of-ai", "canonical_source": "https://pub.towardsai.net/why-adoption-not-the-model-is-the-hard-part-of-ai-1a6c49180ff2?source=rss----98111c9905da---4", "published_at": "2026-08-05 13:10:58+00:00", "updated_at": "2026-08-05 13:36:56.104392+00:00", "lang": "en", "topics": ["artificial-intelligence"], "entities": [], "alternates": {"html": "https://wpnews.pro/news/why-adoption-not-the-model-is-the-hard-part-of-ai", "markdown": "https://wpnews.pro/news/why-adoption-not-the-model-is-the-hard-part-of-ai.md", "text": "https://wpnews.pro/news/why-adoption-not-the-model-is-the-hard-part-of-ai.txt", "jsonld": "https://wpnews.pro/news/why-adoption-not-the-model-is-the-hard-part-of-ai.jsonld"}}