Where to start with AI in your company: diagnose before you buy

The most expensive mistake: buying before understanding

You've probably seen it happen. The owner watches an impressive demo, a competitor mentions they're "already using AI," and within a few weeks the company has a customer service bot, a proposal tool, a meeting transcription app and three chatbot subscriptions. Each one solves a slice of the problem. None of them talk to each other.

The tool isn't the problem. The order is. AI layered on top of a messy process doesn't fix the process: it speeds up the mess. If orders come in by text, get retyped into a spreadsheet and then double-checked from memory by a manager, a bot that answers customers faster will just promise things the operation can't deliver.

That's why I recommend one simple rule: no new tool before a diagnosis. A diagnosis doesn't mean an 80-page report. It means knowing clearly where the bottleneck is, what it costs and what has to be true for AI to actually help.

What the research says about AI projects that don't pay off

These numbers deserve a careful look, because some of them spread widely and not always with the right context.

MIT NANDA: the famous "95% with no return"

In August 2025, MIT NANDA's report "The GenAI Divide: State of AI in Business 2025" made headlines by claiming that 95% of organizations were getting zero return from generative AI. Fortune summed it up: only about 5% of pilots were reaching rapid revenue acceleration; the rest stalled with no measurable impact on the bottom line.

The report itself says the gap between winners and everyone else "does not seem to be driven by model quality or regulation" but by approach. Lead author Aditya Challapally told Fortune that the companies doing well tend to "pick one pain point, execute well, and partner smartly."

Now the criticism. The study is preliminary: it rests on interviews with 52 organizations, 153 leaders surveyed at conferences and a review of 300-plus public initiatives, all between January and June 2025. It defines success as deployment beyond the pilot with measurable KPIs, assessed six months later. The Marketing AI Institute pointed out that this bar is narrow, ignores efficiency gains, and that the report itself calls its figures "directionally accurate," based on interviews rather than official company reporting. In other words: don't treat 95% as a precise statistic. Treat it as a signal that a pilot without focus and without measurement rarely turns into results.

Gartner: abandoned after proof of concept

In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The reasons it cited: poor data quality, inadequate risk controls, escalating costs and unclear business value. In February 2025, Gartner went further, predicting that through 2026 organizations will abandon 60% of AI projects that aren't supported by AI-ready data.

RAND: a misunderstood problem comes first

RAND interviewed 65 experienced data scientists and engineers and in August 2024 published the five leading root causes of AI project failure. The first and most common: people misunderstand, or miscommunicate, what problem AI is supposed to solve. The others include not enough of the right data, chasing the latest technology instead of the user's real problem, weak infrastructure and problems too hard for AI to solve.

BCG: 70% of the work is people and process

BCG describes its 10/20/70 rule: only 10% of the value of an AI transformation comes from the algorithm, 20% from data and technology, and 70% from redesigning the work, along with culture, governance and how people and AI work together.

The 6-step roadmap

This is the path I recommend for anyone who wants to start with AI without wasting money. It follows the same logic as the NoOne method: map and diagnose first, then prioritize, and only then build.

Step 1: map where time and money leak out

Before you talk about AI, look at a normal week in your company. Leaks almost never show up on the P&L. They show up as fatigue, delays and "let me take care of it." Look for these signs:

• Rework: the same information gets typed two or three times (from a text thread to a spreadsheet, from the spreadsheet to the system, from the system to an email).
• Customers waiting: quotes that take days, simple questions left unanswered because "the person who knows" is busy.
• Knowledge only one person has: special pricing, customer history, the right way to do something. When that person goes on vacation, things grind to a halt.
• Decisions stuck with the owner: discounts, purchases, exceptions, proposal approvals. Everything goes through you, and the line keeps growing.
• Reports no one trusts: each department has a different number for the same thing, and the meeting turns into an argument over which spreadsheet is right.

For each sign, write down three things: who feels it, how often it happens each week and how much time it eats up. It doesn't need to be exact. An honest estimate is enough to show where the biggest hole is.

Notice that several of these signs share a root: the company's knowledge lives in a few people's heads and in scattered conversations. If that sounds familiar, read about company memory and the business second brain, because that's where AI starts to have something solid to stand on.

Find out how much of your company still depends on you.Take the free index

Step 2: pick a single area to start

With the list in hand, the temptation is to tackle everything. Don't. Pick one area. To decide, use a simple matrix with three criteria, each scored from 1 to 5:

• Pain: how much the problem costs in money, lost customers or burnout. A 5 means "this hurts cash flow every month."
• Frequency: how often it happens. A 5 means "every day, several times." A rare problem, even a serious one, pays back little when automated.
• Ease: how repetitive the process already is and whether the data already exists somewhere. A 5 means "always the same steps, and the information is in a spreadsheet."

Add up the three scores for each item on your list. The highest total is your first area. If there's a tie, pick the one with the higher ease score: the first win needs to come fast so the team believes in the rest.

In practice, first areas tend to be things like answering quote requests, organizing the sales pipeline, reconciling payables and receivables, or putting together the weekly report. For examples of how real companies chose where to begin, read how companies put AI to work.

Steps 3 and 4: fix the minimum and run a measured pilot

Fix the minimum data and process

You don't need a six-month data project. You need the minimum so AI isn't working in the dark. For the area you picked, make sure of three things:

1. One place for the data. If customer history lives in five group chats and two spreadsheets, pull it into a single source, even a simple one.
2. One written way of working. The current steps, on one page. If everyone does it differently, agree on the right way before you automate.
3. One person in charge. Someone on the team who owns that process and will follow the pilot.

This is exactly the point of Gartner's warning about AI-ready data: without that foundation, projects tend to die after the proof of concept.

A short pilot, with the yardstick set first

Keep the pilot short, two to four weeks, with a before-and-after metric defined before you start. Examples: average time to answer a quote, hours per week spent on the report, number of orders retyped by hand, number of decisions waiting on you.

Measure the "before" for one week without changing anything. Then run the pilot and measure again, the same way. If you didn't measure the before, any result is just an opinion. And opinions are what get AI projects dropped at the first tough meeting.

Steps 5 and 6: a human approves, then you expand

At first, AI prepares and a person approves. It drafts the reply to the customer, someone reviews it and sends it. It builds the report, someone checks it and signs off. That protects the customer, teaches the team to trust the system and shows where AI gets things wrong so you can adjust the rules.

Over time, routine low-risk work can go out directly. Anything important, like pricing, discounts, contracts and money, keeps going through approval. The goal isn't to take the owner out of the picture. It's to have the owner decide only what matters, with the right information in front of them.

Only once the first area works, with before-and-after numbers in hand, is it worth expanding to the next one. And here's the difference between buying apps and building a system: the second area reuses the data, rules and memory of the first. That's the idea behind Business AIOS: a core with the company's memory and rules, and areas added one at a time.

Signs that the pilot is ready to grow

Growing too early is the most common way to turn a good pilot into a big problem. Before taking the idea to another area, it is worth checking that the signs below hold for several weeks in a row, and not just on a good day.

• One person answers for the result. They can say what improved, what got worse and what still depends on adjustment.
• The step by step is written down. Someone else can repeat the process without asking the person who built it.
• Errors are recorded. There is a place where you can see what the AI got wrong and what was corrected.
• The team uses it because it helps. If use only happens under pressure, the pilot has not yet solved a real problem.
• There is a simple measure. Time saved, rework avoided or faster response, compared with before.

If any of these signs is missing, the next step is to fix it, not to widen the scope. A small, reliable pilot convinces the company better than a large, fragile one.

Start with the diagnosis: the IDD and your 30-day plan

A good first gauge is how much the company depends on you. For that, NoOne built the Dependency Index (IDD), a free self-assessment quiz that gives you a score from 0 to 100. The higher the score, the more the operation relies on the owner to keep moving. It takes a few minutes and already points to where the bottlenecks that land on you tend to be. Take it here.

If your diagnosis shows the problem is bigger than one area, that everything goes through you and that information is too scattered for an isolated pilot, that's when a system makes more sense than another app. You can see how NoOne works or, if you prefer, start with just the quiz and take your time deciding the next step.

Cesar Augusto Carlos

Founder, NoOne Systems

Has been building several businesses with digital marketing since 2008. Today he designs systems that bring data, processes and AI together so owners can see the whole company and decide what matters.