AI in a Traditional Business: The 5 Places It Pays Back in Year One (and 3 Where It Burns Cash)

MIT's NANDA initiative studied enterprise adoption of generative AI in 2025 and found that roughly 95% of pilots produced no measurable return on the profit and loss statement. Not a small return. No measurable return at all. Around the same time, a Wine Business survey of an industry that has been optimizing the same process for two thousand years found that only 4% of respondents were using AI to track fermentation and wine supplies.

Read those two numbers together and you get the real state of play. Almost everyone is spending on AI. Almost nobody is spending it where it pays. And the traditional sectors, the ones with real inventory, real machinery and real customers, are barely at the starting line.

The failure rate is not a technology problem. It is a scoping problem: companies buy AI as a strategy instead of applying it to a specific, expensive, repetitive task.

Our group was founded in Azagra in 1890 and now runs eight divisions across wine, real estate, hospitality, mineral water, electrical installations, mobility and US distribution, selling into more than 75 countries with over 180 people. That is a portfolio of deeply traditional businesses. Over the last two years we have watched which uses of this technology actually paid for themselves inside twelve months and which quietly turned into a subscription nobody could justify. Here is the honest split.

The test to run before you spend anything

Before evaluating a single tool, put the proposed use case through three questions. If it fails any of them, you are buying a pilot that will join the 95%.

• Is the task high volume? A process you run four times a year cannot repay implementation cost, no matter how clever the automation. You need something that happens hundreds or thousands of times.
• Is the task language based or pattern based? Current models are extraordinary at reading, writing, translating, classifying and summarizing. They are mediocre at judgment calls that depend on context nobody has written down.
• Do you already have the data, in one place, in usable shape? Gartner has estimated that around 85% of AI projects fail because of poor or insufficient data quality. The model is almost never the problem. The spreadsheet nobody maintains is.

If you cannot name the exact task, the number of times per month it happens, and the hours it consumes today, you are not ready to buy anything.

Where it pays back in year one

1. Language heavy administration

This is the least glamorous and most reliable win in a traditional company. Reading supplier contracts and extracting the terms. Turning a technical specification into a customer-facing quote. Reconciling invoice lines against a purchase order. Drafting the first version of a tender response.

In our distribution and installations businesses, this kind of work consumes real hours from people whose time is better spent on customers. Forbes has reported that small and midsize companies save an average of around $7,500 a year from AI, with the top quartile saving more than $20,000. Those figures are unremarkable precisely because they come from unremarkable work.

The highest-return AI in a traditional business almost never touches the product. It touches the paperwork around the product.

2. Inbound response speed

The commercial value here is not the cost of writing a reply. It is the conversion difference between answering an inquiry in ten minutes and answering it in two days.

Drafting first responses to inbound inquiries, classifying them by urgency and routing them to the right person is a narrow, high-volume, language-shaped task. A human still reviews and sends. What changes is that nothing sits in a queue over a weekend.

3. Selling in languages you do not staff for

We sell into more than 75 countries. Historically, entering a market meant either hiring a native speaker before the revenue justified it, or accepting materials that read as though a foreigner wrote them, which in a premium category is worse than saying nothing.

Translation and localization of commercial material is now close to free and, with a native speaker reviewing rather than writing, close to native quality. That changes the arithmetic of market entry, which is the same discipline we described in our piece on choosing an international distributor.

4. Demand forecasting and inventory

The U.S. Chamber of Commerce has reported that family businesses using AI apply it most often to process efficiency (40%) and risk management (39%). Forecasting sits at the intersection of both.

Any business with seasonality, perishability or long lead times is currently forecasting with a spreadsheet and an experienced person's instinct. That instinct is genuinely valuable and should not be replaced. But a model that reads three years of shipments, weather, promotional calendars and lead times will surface patterns the instinct misses, and every point of forecast accuracy converts directly into working capital. The mechanics of why that matters are in our piece on the cash conversion cycle.

5. Institutional memory

A company founded in 1890 has an enormous amount of knowledge sitting in documents nobody can find. Contracts, technical specifications, supplier correspondence, maintenance records, decisions and their reasons.

Making that searchable in plain language is one of the few AI projects where the value grows with the age of the company. The older and more document-heavy the business, the larger the buried asset.

Where it burns cash

1. Anything that replaces craft judgment

There is a persistent fantasy that AI will make the expert decision. In wine, that means the blending call. In real estate, the site decision. In installations, the diagnosis of a fault that presents like three other faults.

These decisions rest on context that has never been written down, which means it has never been in the training data. Worse, the model will produce a confident, fluent, wrong answer, and a confident wrong answer is more dangerous than no answer.

Use AI to give your experts better inputs. Do not use it to replace the output of thirty years in the same trade.

2. Custom model building before the data is clean

Every year a vendor proposes a bespoke model trained on your proprietary data. It sounds like a moat. In a company whose data lives in four systems that disagree with each other, it is a very expensive way to discover you have a data problem.

McKinsey's framing is useful here and it is worth memorizing: the 10/20/70 rule. Roughly 10% of the effort in a successful AI program goes to algorithms, 20% to technology and data, and 70% to people and process change. Companies that spend as though the ratio is reversed are the ones writing off pilots.

3. Company-wide licenses with no named owner

The most common waste we see is not a failed project. It is a subscription bought for everyone, used properly by four people, and renewed automatically because cancelling it would look like retreat.

Every tool needs a named owner, a specific process it is meant to improve, and a number that should move. If nobody owns it, nobody measures it, and it becomes exactly the kind of unexamined cost we wrote about in cutting costs without cutting muscle.

The measurement that separates the 5%

MIT Sloan research found that firms with low AI intensity, below a 25% threshold, saw essentially zero annual revenue growth attributable to it, while firms above that threshold approached 24%. The lesson is not "spend more." It is that scattered, half-committed adoption returns nothing, and that the companies which win pick fewer processes and go all the way through them.

Practically, that means three things:

• Baseline before you buy. Record the hours, error rate or response time today. A benefit you cannot compare to a starting point is a story, not a return.
• One process, all the way to production. A pilot that never changes how the work is actually done is a demo. Demos do not appear on the profit and loss statement.
• Budget for the 70%. Training, process redesign and the awkward months where people run the old way and the new way in parallel are the real cost.

Key Takeaways

• Roughly 95% of generative AI pilots deliver no measurable P&L return (MIT NANDA), and the cause is scoping, not technology.
• Qualify every use case on three tests: high volume, language or pattern based, and data you already have in usable shape.
• The reliable year-one wins in traditional companies are administrative: document handling, inbound response speed, localization, forecasting and searchable institutional memory.
• Around 85% of AI projects fail on data quality (Gartner). Clean the data before anyone proposes a custom model.
• Never point AI at the craft judgment that took your experts decades to build. Point it at the inputs feeding that judgment.
• Remember the 10/20/70 rule: 10% algorithms, 20% technology and data, 70% people and process.
• Every tool needs a named owner, one target process and one number that must move, or it becomes a renewing subscription nobody defends.

Frequently Asked Questions

How much does AI cost for a small business?

Far less than most owners expect for the tools, and far more than they expect for the change. Per-seat assistants typically run in the range of $20 to $60 per user per month, and most small companies need very few seats. The real budget line is implementation: process redesign, training and the parallel-running period, which usually exceeds the software cost in year one.

How can I use AI in my small business?

Start with one repetitive, language-heavy task that happens hundreds of times a month, such as drafting quotes, answering routine inbound inquiries, or extracting terms from supplier documents. Measure the hours it takes today, run it for a quarter with a human reviewing every output, and only expand once you can show the change on a number.

What is the 10/20/70 rule for AI?

It is McKinsey's estimate of where effort goes in a successful AI program: about 10% on the algorithms, 20% on the technology and data, and 70% on people and process change. It is the single most useful correction to the instinct that buying the tool is the project. The tool is roughly a tenth of the work.

What industries are adopting AI the most?

Technology, financial services, professional services and media lead by a wide margin, because their core product is already information. Traditional sectors lag sharply. A Wine Business survey found only 4% of respondents using AI to track fermentation and supplies, which tells you both how far behind these industries are and how much unclaimed advantage is sitting there.

Why do so many AI projects fail?

Three causes account for most of it: data that is fragmented or unreliable, pilots that are never taken through to production, and use cases chosen because they sounded impressive rather than because they were expensive and repetitive. Gartner has put the data-quality share alone at around 85% of failures.

Pick one process this quarter

Do not write an AI strategy. Take the single most repetitive, document-heavy, high-volume task in your company, measure what it costs you today in hours and errors, and run one tool against it for ninety days with someone's name on the result. That is how the 5% got there.

See how the Manzanos Enterprises group operates across eight divisions and more than 75 countries, and if you want the longer view on modernizing a company without damaging what makes it valuable, read our related piece on adopting technology in heritage businesses without losing their soul.