The Number Nobody in AI Likes to Quote
In 2025, researchers at MIT looked at how generative AI was actually performing inside businesses. The headline that came out of it was brutal: about 95% of enterprise AI pilots delivered no measurable return.
If you are a New Zealand small business owner who has been told AI will change everything, that number deserves a hard look. Because the honest version of it is more useful than either the hype or the doom. AI is not failing because it does not work. It is failing because most businesses cannot get it across the last mile, from a clever demo to something that actually runs the business. This guide explains why, what it means for a small business specifically, and how to be in the 5% that gets a real result.
What the 95% Actually Means
It is worth being precise, because the headline gets stretched.
The study (MIT's "State of AI in Business 2025", from its NANDA initiative) looked at enterprise generative-AI pilots and found that the large majority showed no measurable impact on the bottom line. It did not say the AI gave wrong answers. It did not say 95% of companies regret trying. It said most pilots never turned into a durable change in how the business made or saved money.
The researchers were also clear about the cause. It was not model quality. They called it a "learning gap": generic tools work fine for an individual typing into a chat box, but they stall inside a business because they do not learn or adapt to that business's actual workflow. The technology was rarely the problem. The fit was.
So the takeaway is not "do not bother with AI". It is "the demo is the easy part, and almost everyone underestimates the rest".
Why AI Projects Fail: the Last Mile
Engineers have a name for this. The last mile is the final, unglamorous stretch between a thing that works in a demo and a thing that works in the real world. With AI, the demo is now astonishingly easy. The last mile is where the failures live. Here is what it actually consists of:
- Integration. A real business runs on a stack of tools: email, calendar, accounting, a job-management or booking system, a website. AI only earns its keep when it is wired into those, not sitting in a separate chat window you have to copy and paste from.
- Workflow fit. Your business does not run like the average in a training dataset. AI has to bend to how you actually quote, book, follow up and invoice, including the local details. A generic tool that makes you change how you work usually gets quietly abandoned.
- No one to land it. A pilot needs someone to take it from "interesting" to "live", and then keep it running. Big companies have teams for this. Most do not.
- Data readiness. The information the AI needs is often spread across spreadsheets, inboxes and someone's head. Waiting for perfectly clean data is a fantasy, but ignoring data entirely guarantees a stall.
- Trust and control. No sensible owner hands an unsupervised AI the authority to send, post or spend. Without clear guardrails and a human approving the important actions, the pilot stalls at the trust question and never goes live.
Notice that only one of these five is about the AI itself. The other four are about getting it into your business. That is the last mile.
The Most Useful Finding for a Small Business
Buried in the same MIT research is the line that matters most if you run a small business.
Projects where companies bought or partnered for their AI succeeded far more often than projects companies tried to build internally. Roughly two thirds of the bought-or-partnered efforts worked, versus only about a third as many of the internal builds.
Read that again, because it is counterintuitive. The companies that handed the job to people who do this for a living got a result far more often than the ones who tried to do it themselves. For a small business with no dedicated tech team, that is not a small detail. It is the whole game. Trying to implement AI alone, on top of running your business, is the most reliable way to join the 95%.
Why Small Businesses Stall Even Harder
Everything above is harder, not easier, when you are small.
A large enterprise that fails an AI pilot has a team, a budget and another quarter to try again. A small business has none of that slack. The barriers show up clearly in the New Zealand data. The AI Forum NZ found that while a large majority of NZ organisations are using AI in some form, only around one in eight have scaled it across the business, and nearly half are still stuck in pilot or exploration mode. Industry surveys put the top barriers as a lack of internal skills, messy or disconnected data, and uncertainty about governance, with cost close behind and most workers reporting no AI training at all.
In plain terms: the typical NZ small business does not stall on AI because the owner is not smart enough or the tools are not good enough. It stalls because there is nobody whose job is to take AI the last mile. The consultant's slide deck arrives, and then it sits there, because actioning it was always going to need a team the business does not have.
How the Big Players Solved It
Here is the interesting part. The most advanced AI companies in the world hit exactly this wall, and their answer tells you what actually works.
Rather than ship software and hope customers could implement it, they started embedding their own engineers inside the customer's business. Palantir pioneered the model years ago and called the role the forward deployed engineer. In 2026 OpenAI launched a dedicated deployment company, informally "DeployCo", whose entire purpose is to put forward deployed engineers inside large organisations to build and run AI in production, even acquiring a firm to bring in a ready team of deployment specialists. Anthropic built its own version with applied AI engineers.
The pattern is unmistakable. The frontier of the industry is not racing to build a smarter model. It is racing to solve deployment, the last mile, by sending people to embed, understand the real workflow, and build the thing that fits. They worked out that the bottleneck was never the model. It was getting it landed.
If you want the full explanation of the role, we wrote one: what a forward deployed engineer is, in plain English.
What This Means for You
The catch, of course, is that those forward deployed engineers cost a fortune and are aimed at banks and Fortune 500s. A small business cannot hire OpenAI's deployment team.
But the approach is what matters, and the approach scales down. You do not need a smarter model. You need someone to do for your business what DeployCo does for a big one: sit with you, learn exactly how your work flows, build the AI around it, connect it to the tools you already use, and stay accountable for getting it live. That is the entire idea behind a forward deployed engineer for a small business, and it is what we do at BestAI, scaled to a price a small business can actually approve.
How to Actually Implement AI in Your Business
If you want to land in the 5%, here is the practical sequence, whether you do it with us or anyone else.
- Start with one task, not a strategy. Forget "an AI strategy". Name the single repetitive task that wastes the most time this week. That is your pilot.
- Demand a working version, not a deck. The deliverable should be something running that you can touch in days, not a document describing what could be built.
- Test it somewhere safe first. A first version should run where it cannot affect your real business, so you can see it work before you trust it.
- Keep a human in the loop. Nothing gets sent, posted or spent without your approval, and every action is logged. Trust is earned one supervised step at a time.
- Connect it to your real tools. The value appears when it is wired into your email, calendar, accounting and website, not sitting in a separate app.
- Get someone to own the last mile. This is the step everyone skips and the reason most pilots die. If you do not have an internal team, have a partner build and land it. The MIT data says that is the more reliable path anyway.
For the everyday jobs this applies to, and how we run it end to end, see our AI automation for NZ small business page. If your need is really one focused tool rather than a workflow, My Tool, 48hrs is the faster route.
FAQ
Is it true that 95% of AI projects fail? The figure comes from an MIT study published in 2025, and the precise claim is narrower than the headline. It found that about 95% of enterprise generative-AI pilots delivered no measurable return on the bottom line. It does not mean AI does not work. It means most pilots never make it past the demo into something that changes the business. The cause is usually the last mile, not the model.
Why do most AI projects fail? Rarely because the AI is not capable. They fail at the last mile: connecting the AI to the systems a business already runs, fitting it to the real workflow, having someone to get it live, and trusting it enough to give it real work. The MIT study called the core issue a learning gap, where generic tools do not adapt to how a specific business actually operates.
What is the AI last-mile problem? The last mile is the gap between a working AI demo and AI that is genuinely wired into how a business runs. A demo is easy. Integrating it with your email, calendar and accounting, shaping it to your workflow, getting your team to trust it, and keeping it running is the hard 90%. Most projects stall there.
How does a small business actually implement AI? Start with one specific, repetitive task rather than an AI strategy. Get a working version into a safe test first, keep a human approving anything that goes out, then connect it to the tools you already use and train your team. Most NZ small businesses do not have an internal team for this, so the realistic path is to have an outside partner build and land it for you.
Should a small business build AI in-house or get help? The same MIT study found that buying or partnering on AI succeeded far more often than building it internally. For a small business with no dedicated tech team, getting an outside team to build and land it is usually both cheaper and far more likely to actually work than trying to do it alone.
Next Steps
- Pick the one repetitive task that wastes the most time in your business this week. That, not "AI", is where to start.
- Book a free 45-minute chat. We will tell you honestly whether AI is the right answer for that task, and what it would take to get it actually running.
- No obligation. If AI is not the right tool for it, we will say so.
BestAI is an Auckland-based custom software and AI automation company serving New Zealand small businesses. We bring the embed-and-build approach big AI labs use, and land AI that actually works, from $399. Get in touch or see how we do AI automation.