What Is AI Actually Good At In A Real Business?

A global bank just published the two AI systems it put into production, and both of them are so unglamorous that almost nobody running a small business would have picked them. Here is the question this article answers: of all the things you could point AI at, which ones actually pay?

The Hook

On October 1, 2026, Anthropic published what Barclays is doing with Claude. I read it expecting something clever. What I found instead was a bank quietly admitting that the two AI systems it trusted enough to run in production are the two most boring things in the building.

One finds information. One sorts incoming mail.

That is it. A Colleague Knowledge Assistant, built on retrieval, serving more than 16,000 staff and racking up over one million searches. And a Global Markets email triage system that classifies and routes roughly 120,000 emails every single day. Barclays also plans to have Claude Code in the hands of 50% of its developers by the end of 2026 and a majority of its engineers in 2027, but the two systems it chose to show the world are retrieval and routing. Source: Anthropic.

Here is the direct answer, before the first section break. AI is reliably good at two things in a real business: finding the information your people already own but cannot locate, and sorting incoming work so it lands in the right place. Almost everything else you have been trying is harder, riskier, and worth less.

Candidly, that hurts a little. I built a custom GPT for my team that I was proud of, and almost nobody used it, because it answered questions nobody was asking. Meanwhile three people kept asking me where a document lived. I had the glamorous problem solved and the boring problem bleeding.

The pattern holds at every size. The bank with 16,000 staff and the coach with two contractors lose time in the same two places.

Key Takeaways

The two AI use cases with the most documented production success are retrieval, meaning finding what you already own, and classification, meaning sorting what arrives.

Most AI projects fail for reasons that have nothing to do with the model and everything to do with picking the wrong process to automate.

The widely repeated claim that workers lose about 19% of the week searching for information traces back to a 2012 McKinsey estimate, not a recent measurement.

You should choose the process before you choose the tool, because the tool is the cheap part and the process is where the money is.

A boring, narrow, measurable AI deployment beats an ambitious one you cannot evaluate.

The Problem

You have been sold the wrong shape of problem.

Every demo you have watched starts with the tool. Here is ChatGPT, here is Claude, here is an agent builder, now go find something to do with it. That is backwards, and it is the single most expensive mistake I see in my community. Isar Meitis made exactly this argument on Leveraging AI episode 331 on September 29, 2026, with Corey Ganim: most companies start with the tool instead of identifying the highest-ROI process first. Source: Leveraging AI.

I did this for most of 2023. I would see a capability, get excited, and then go hunting for a place to put it. That is shopping, not strategy. It feels like progress because you are busy and learning, and it produces almost nothing you can point at six months later.

The difficulty is real, and I am not going to pretend otherwise. You cannot easily tell, from the outside, which AI use cases are solid and which are demo-ware. A tool that reliably sorts emails and a tool that claims to run your whole sales process are sold with the same font, the same confidence, and often the same price. You are being asked to make a technical judgment call without technical training, which is not fair to you.

And the honest version is worse. Nobody publishes their failures. You see the wins on LinkedIn. You do not see the 40 hours somebody burned on an agent workflow that broke the first time a client wrote a weird subject line.

So here is the reframe that changes everything. Stop asking what AI can do. Start asking which of your processes has three specific properties: it happens constantly, it has a clear right answer, and today a human does it from memory. Those three properties are the tell. Retrieval and routing both have them. Almost every failed AI project I have looked at is missing at least two.

The Evidence

Let me give you the numbers, with sources, including one place where the famous number does not hold up.

One. The information-finding problem is real, and the number everybody quotes is older than you think. You have probably seen a stat claiming knowledge workers lose roughly 19% of their week looking for information. The traceable primary source is the McKinsey Global Institute report The social economy, published July 2012, which states that the average interaction worker spends an estimated 28 percent of the workweek managing email and nearly 20 percent looking for internal information or tracking down colleagues who can help with specific tasks. That same report estimated a searchable record of knowledge could cut information-search time by as much as 35 percent. Source: McKinsey Global Institute. Note what that means: the figure you keep seeing cited as current fact is a 2012 estimate. It is directionally useful. It is not a 2026 measurement, and anyone presenting it as fresh data is being sloppy.

Two. A recent survey, with a named methodology, says the problem got worse. Coveo's fifth annual EX Relevance Report, released April 29, 2025, surveyed 4,000 US and UK adults at companies with 5,000 or more employees through Arlington Research. Employees reported wasting an average of three hours per day searching for information, said 42% of the information they sift through is irrelevant to their role, and 47% cited information scattered across multiple applications as their largest hurdle. Source: Coveo. It is vendor-published research, so read it with that in mind. The direction matches Barclays putting a retrieval system in front of 16,000 people.

Three. Retrieval has a peer-reviewed production result attached to it. LinkedIn's engineering and research team published a paper at SIGIR 2024 describing a retrieval-augmented system combined with a knowledge graph built from historical support issues. After roughly six months in production with LinkedIn's customer service team, it reduced median per-issue resolution time by 28.6%, taking the median from about 7 hours to about 5 hours. Source: arXiv. That is a named company, a measured result, and a published method. It is the same shape as the Barclays Colleague Knowledge Assistant.

Four. Classification and routing has the loudest documented result, and the most honest caveat. Klarna announced on February 27, 2024, that its AI assistant had handled 2.3 million conversations in its first month, two thirds of all customer service chats, cutting average resolution time from 11 minutes to under 2 minutes, with an estimated $40 million profit improvement for 2024. Source: Klarna. Klarna later pulled back and rehired human agents for complex work, as Fast Company and others reported. Both halves matter. Sorting and first-pass handling worked. Replacing judgment did not.

Five. The failure data points straight at process selection, not technology. RAND's report The Root Causes of Failure for Artificial Intelligence Projects reports that by some estimates more than 80 percent of AI projects fail, roughly twice the rate of non-AI IT projects, and names the top root cause as misalignment on the project's purpose, with chasing the technology instead of the problem close behind. Source: RAND. MIT's NANDA initiative, in The GenAI Divide: State of AI in Business 2025, reported that about 95% of enterprise generative AI pilots produced no measurable P&L impact. Source: Fortune. Be careful with that 95%. It came from a limited set of interviews and public deployment reviews, it has been contested, and there is no single agreed AI failure rate; the honest statement is that credible estimates range from about 80% down to far lower depending on what you count as failure.

Put those together and the conventional narrative looks thin. The story is not that AI does not work. The story is that AI works narrowly and the market keeps selling it broadly.

The Solution

Here is the thing. The two production systems Barclays disclosed are not a bank story. They are a template, and it scales down to a business with one employee.

I call it retrieval and routing, and the framework is three questions asked in this order.

Question one: where does my business lose time knowing things? This is retrieval. Every business accumulates answers it cannot find: pricing exceptions, onboarding steps, how to handle a refund, which proposal template is current. The symptom is interruption. If your team asks you something you have already answered in writing somewhere, you have a retrieval problem, and retrieval is the best-documented thing AI does. Barclays built it for 16,000 people. You build it for four.

Question two: where does my business lose time sorting things? This is routing. Email, inquiries, applications, support requests, leads, invoices. The work arrives unlabeled and a human reads each one to decide where it goes. Barclays classifies roughly 120,000 emails a day. You might classify 60. The economics work at both scales because the cost is per decision and the decision is identical.

Question three: what am I tempted to build instead, and why? Ask it out loud. You are probably tempted by strategy, creative, or a full agent that runs something end to end. I am not saying never. I am saying those are the projects that show up in RAND's failure column, because the right answer is fuzzy and you cannot tell whether the output was good.

Two things from this week suggest the industry is catching up. Yedric.ai took second place on Product Hunt on October 1, 2026, with 199 upvotes, for an embeddable agent that lets people control software they already own using natural language instead of learning where each feature lives. Source: Product Hunt. That is retrieval applied to a user interface. And on September 30, 2026, Gemini 4 Argon took first place on Zapier's AutomationBench with 51.3%, a benchmark built around business processes rather than coding. Hold that number: the best model scored barely over half on realistic business-process tasks. The frontier is competent at narrow work and mediocre at sprawling work. Your project plan should match.

My own proof case is unglamorous. The highest-return AI thing I run is not content and not strategy. It is a system that answers questions about my own material so my team stops routing them through me, and a sorting layer on inbound messages so the ones that need me actually reach me. Candidly, I resisted building both for over a year because they were not interesting enough to be worth my attention. That year was expensive.

Practical Steps

Track interruptions for five business days. Keep a running note every time someone asks you something you have already answered, or you hunt for a file longer than 60 seconds. Do not fix anything yet. You are building evidence, and five days is enough to see the pattern.
Count your inbound streams and label them by hand for two days. Write down every category your incoming email, forms, and messages actually fall into. If you cannot name the categories, AI cannot sort into them. This step is the whole project; the software is an afterthought.
Pick exactly one process, and make it the one with a right answer. Not the one that would impress people. If a competent new hire could be told whether the output was correct, keep it. If the evaluation would be a matter of taste, drop it for now.
Gather the source material into one place before you touch a tool. For retrieval, that means your actual documents, policies, pricing, and past answers in a single folder or knowledge base. Garbage in, confident garbage out. Most retrieval failures are document failures.
Use this prompt to pressure-test your choice before you spend a dollar. Paste it into Claude or ChatGPT and answer honestly.

[The Job]
Help me decide whether [PROCESS I AM CONSIDERING AUTOMATING] is a good first AI project for my business, or whether I should pick something else.

[The Background]
My business is [WHAT YOU DO] with [NUMBER] people. The process happens about [FREQUENCY]. Today it is done by [WHO DOES IT] and takes roughly [TIME PER OCCURRENCE]. The information it depends on currently lives in [WHERE THAT INFORMATION LIVES]. When it goes wrong, the consequence is [CONSEQUENCE OF AN ERROR]. My other candidate processes are [LIST TWO OR THREE ALTERNATIVES].

[The Deliverable]
A direct verdict in the first sentence: good first project, or not yet. Then score my process from the perspective of someone who has watched AI projects fail, against four criteria: frequency, whether there is a verifiable right answer, how much the needed information is already written down, and the cost of an error. Then tell me which of my candidate processes you would actually start with and why. Be blunt. If my pick is wrong, say so plainly.

[The Questions]
Ask me any questions you have.

Build the smallest possible version and give it one week. One folder of documents, one category of email, one user. If it does not help inside a week at small scale, it will not help at large scale; it will just fail more expensively.
Measure the thing you were losing, not the thing the tool reports. Count interruptions avoided, or minutes from arrival to correct destination. Compare to your five-day baseline from step one. A tool's own dashboard will tell you it was used, which is not the same as useful.

Frequently Asked Questions

Is AI only useful for big companies like Barclays?
No. The two use cases Barclays disclosed, internal knowledge retrieval and email classification, both scale down cleanly because the cost is per decision, not per company. A four-person consultancy loses the same proportion of its week to finding documents and sorting inbound work as a bank with 16,000 staff does.

What is RAG, in plain language?
RAG stands for retrieval-augmented generation. It means the AI looks up your actual documents before it answers, instead of answering from general training. That is why it can tell you your refund policy rather than a generic one. Barclays built its Colleague Knowledge Assistant this way, and LinkedIn's published support system uses the same approach.

Why do most AI projects fail?
According to RAND's research, the leading cause is misalignment on the project's purpose, followed by chasing technology instead of a defined problem. Models are rarely the bottleneck. Projects fail because nobody agreed what success looked like, the needed information was never written down, or the task had no verifiable right answer.

Should I start with ChatGPT, Claude, or an automation platform?
Start with the process, not the tool. Once you know which process you are fixing, the tool choice is usually obvious and often reversible. Picking a platform first is what Isar Meitis and Corey Ganim identified as the common failure pattern on Leveraging AI episode 331, and it is the expensive way to learn.

How long before a retrieval or routing system pays for itself?
Faster than you expect, if it is narrow. LinkedIn's published system cut median support resolution time 28.6% within about six months of production use. For a small business, a one-week pilot on a single document set or one email category is usually enough to see whether the time savings are real.

The Close

I went looking for something clever in that Barclays announcement. What I found was a bank with effectively unlimited budget choosing the two most boring AI projects available, and I think that is the most useful thing published this quarter.

You have been trying to make AI impressive. Stop. Impressive is what the vendors need. What you need is one process that happens every day, has a right answer, and currently runs out of somebody's memory. Fix that and you will get hours back this month. Chase the autonomous agent that runs your marketing and you will join the 80 percent that RAND is writing about.

You are not behind because you have not learned enough tools. You are behind because nobody told you which problems were solvable, and the people selling you tools had no reason to narrow your ambition. I am telling you now, including the unflattering part: I know this because I spent a year building the exciting thing while the boring thing quietly cost me more than fixing it would have.

If you want the full process-selection and build framework, along with the prompt library and the walkthroughs we use inside AI Insiders, that is exactly what the membership exists for. Come get the shortcut I had to buy with my own wasted year.

Find the boring thing. It is where the money is.

About the author: Jonathan Mast is the founder of White Beard Strategies LLC, where he provides AI coaching and mentorship to entrepreneurs, consultants, and small business owners. He runs a Facebook community of more than 500,000 members, leads the AI Insiders membership, created the Perfect Prompt Framework, and speaks regularly on practical AI adoption. He spent most of a year building AI systems nobody used before admitting that his own highest-return deployment was a boring retrieval system that answers questions about documents he had already written.

What Is AI Actually Good At In A Real Business?

Contents

The Hook

The Problem

The Evidence

The Solution

Practical Steps

Frequently Asked Questions

The Close

Sources

Do I Have to Tell My Clients When I Use AI on Their Work?

Why Is My AI Bill So High When Most of What I Ask It Is a Yes-or-No Question?

How Do I Get My Business in Front of Customers When They Never Leave ChatGPT?

Why does my AI agent work in the demo and fall apart the moment real customers touch it?

About the Author

Jonathan Mast