If you funded an AI initiative this year, the honest odds are that it did not work. The reason has almost nothing to do with the model.
MIT NANDA's research found that 95 percent of generative AI pilots produce no measurable impact on profit or loss. IDC tracked 33 hospital AI pilots and found that only 4 reached production, an 88 percent failure rate at the exact moment the project was supposed to start paying for itself. Gartner expects more than 40 percent of the agentic AI projects running right now to be cancelled by the end of 2027. None of that is a story about AI not working. It is a story about the wrong tool getting picked for the job, over and over, for a reason that has nothing to do with engineering.
The reason is buzz. Somewhere in the last two years, choosing an AI architecture stopped being an engineering decision and started being a status decision. Nobody gets excited announcing that the company built a lookup tool. Everybody gets excited announcing an autonomous agent. So teams reach for the agent, the custom model, the flashiest possible label, whether or not the actual business problem called for it. That mismatch, not model quality, is where most of this year's failure statistics come from.
Nobody gets excited announcing a lookup tool. Everybody gets excited announcing an agent. That is how good money ends up solving the wrong problem.
There Are Only Seven Real Jobs an AI System Can Do
Strip away the marketing language and every AI system on the market today is doing one of seven things. Knowing which one you actually need, before anyone shows you a demo, is the single highest-leverage AI decision a leadership team makes all year.
Look things up and answer from them (RAG)
Think of this as an AI with a library card. Instead of guessing from memory, it finds the actual document, contract, or record first, then answers from what it found. It is the workhorse pattern in serious companies right now, showing up in somewhere between 51 and 73 percent of enterprise deployments depending on the survey, because it can point to its source. This is the right tool anytime the job is answering questions from information you already have. It is the wrong tool when a company sells it to you as something smarter than a very good lookup system.
Sound like your company, consistently, at scale (fine-tuning)
This is coaching, not education. It teaches an AI your tone and your format without teaching it new facts. Only about 9 percent of production systems use this as their primary method, and it is the most commonly misapplied pattern on this list. A fine-tuned system can state something completely wrong in the exact same confident, on-brand voice it uses for something correct, which makes the mistake harder to catch. If a vendor says they fine-tuned a model to “know” your product catalog, they used the wrong tool. That is a lookup job, not a coaching job.
Do both at once, with a paper trail (hybrid)
This is the combination: an AI that talks like your team and looks things up before it answers. It shows up in clinical documentation, finance, and legal, anywhere a regulator wants both a professional tone and proof of where an answer came from. If your industry has an auditor, this is very likely the pattern you actually need, not a flashier standalone system.
Complete a multi-step task on its own (agentic)
This is the category generating the most excitement and the least agreement on how far along it actually is. Depending on definition and survey, somewhere between 11 and 31 percent of enterprises have a true agent in production. Gartner projects 40 percent of enterprise applications will embed some task-specific agent by the end of this year, which sounds like momentum until you notice that is an embedding statistic, not a success statistic. The tell for whether something is a real agent or a flowchart wearing an AI costume: does it choose its own next step and recover from a failed one, or is a human-written script doing the actual decision-making with an AI answering questions at each stop. This is the category where a rushed, guardrail-free deployment is most likely to end up in next year's cancellation statistics.
Make a narrow numeric prediction (traditional machine learning)
Fraud alerts, demand forecasting, predicting when a machine will fail, occupancy patterns on a sensor network. This technology has existed for more than a decade and quietly still does most of the real work, now wearing an AI-powered label it did not need. It is often the cheapest, fastest, most accurate option for a well-defined numbers problem, and it is the pattern most likely to hide in plain sight during diligence because nobody thinks to ask if it is actually new.
Train a genuinely new model from nothing (foundation model pretraining)
This is building a mind from birth rather than hiring one already educated. It costs an enormous amount of money and a rare kind of talent, and outside a handful of frontier labs and heavily capitalized physical-world robotics companies, almost nobody actually does it. If a mid-market company tells you they trained their own model, ask directly whether that meant starting from zero or customizing an existing open model. Those two claims are separated by roughly two orders of magnitude in cost, and the gap between them is where a lot of thin technical claims like to hide.
A prompt with nothing built underneath it
Worth naming because it shows up constantly and is not really a seventh capability so much as the absence of one. A system prompt sitting in front of an existing AI model, with no retrieval, no training, and no proprietary data underneath it. It is the fastest, cheapest thing to build, which is exactly why it is the easiest for a competitor to copy. If this is the whole product, that is the finding, not a footnote.
Match the Job to the Tool, Not the Tool to the Buzz
This is the checklist version. Before approving any AI initiative, name the actual job first, then check which pattern fits it. Most real business problems land in the boring middle of this list, not the exciting end of it, and that is fine. The differentiation was never going to come from having an agent. It comes from retrieval quality, data discipline, and picking the right box in the first place.
| If the job actually needs… | Reach for… | Not this, even though it is louder |
|---|---|---|
| Answers grounded in your own current documents, with a source you can point to | RAG | A custom-trained model. You do not need to teach it facts, you need it to look them up. |
| Consistent tone, format, or classification on a narrow task, run thousands of times a day | Fine-tuning | A general-purpose chatbot wrapper. It will drift in tone and cost more per query at volume. |
| A regulated, high-stakes answer that must both sound right and prove where it came from | Hybrid RAG plus fine-tuning | Either one alone. Compliance needs both the paper trail and the controlled voice. |
| A system that completes a multi-step task end to end, with recovery when a step fails | Agentic, but only with guardrails and a rollback plan already built | An agent with no human checkpoint. This is where most of 2026's cancelled pilots came from. |
| A numeric prediction from clean, structured data: failure rates, demand, utilization | Traditional ML | Generative AI. It is slower, more expensive, and less accurate for this exact job. |
| A genuinely novel capability nothing else on the market can do | Foundation model pretraining | This is the only box on this table almost no one should check. Verify the claim before you believe it. |
The Trust Gap Is an Architecture Problem Wearing a Training Problem's Clothes
A piece in Healthcare Today this week made the case that healthcare's AI gap is about confidence, not technology. Joint research from Corti and YouGov, covering nearly 2,500 healthcare professionals across the UK, Germany, France, Denmark, and the US, found that 74 percent supported using AI in their work, but more than half said they did not trust the tools currently available. Fear of clinical error was the leading barrier clinicians named. The article's conclusion is that healthcare needs to invest in workforce readiness and training, not just deploy more tools.
That diagnosis is half right, and the missing half is the one this piece is about. Training a clinician to use a system well cannot fix a system that was never built to be trusted. A prompt-only wrapper cannot show its work. A fine-tuned model can be wrong in a confident, on-brand voice that makes the error harder to catch, not easier. An ungoverned agent making its own next-step decisions in a clinical workflow is exactly the scenario a cautious clinician should distrust. The confidence gap the article describes is, in a meaningful number of cases, a rational response to the wrong architecture being deployed, not a training deficit in the person using it.
The pattern built for this exact problem already exists. Hybrid RAG plus fine-tuning gives a clinician both a controlled, professional voice and a traceable source for every claim, the two things a person needs before they can reasonably trust an output they did not generate themselves. Training closes the gap between a good system and a confident user. It cannot close the gap between a bad system and a cautious one. Healthcare's AI adoption problem will not resolve by teaching people to trust tools that were not built to earn it.
The Question That Separates Real from Rehearsed
If someone tells you their product runs on “proprietary AI,” there is one follow-up that gets you further than any demo. Ask them to walk through which of these seven their system actually is, specifically, and what they gave up to build it that way. A team with a real answer gives it to you in under five minutes, tradeoffs included. A team without one reaches for the marketing language again. That reach is the finding.
I apply this same discipline to the stranded post-M&A IoT and RTLS assets at the center of Johnson Holdings Group's acquisition thesis. A product line that underperforms inside a large strategic is rarely underperforming because the technology is weak. It is usually organizational structure or a go-to-market mismatch that has nothing to do with what the product can actually do. The AI-powered label creates the same distortion in reverse. It lets a thin technical asset look stronger than it is, and it lets a genuinely solid asset built on unglamorous classical ML look weaker than it is. Either way, the label is doing work the architecture should be doing instead.
Ask which of the seven their system actually is. A team with a real answer gives it to you in five minutes. A team without one reaches for the marketing language again.
