I build AI systems for a living, so you might expect me to say every business should use AI for everything. I won’t, because it isn’t true, and pretending otherwise is part of why so many AI projects fail.
In 2025, a widely discussed MIT study of enterprise AI found that about 95% of generative AI pilots were delivering no measurable impact on profit and loss. The models weren’t the main problem. Most of those projects were aimed at the wrong job, built on the wrong foundation or never connected to how the business actually works.
Before you spend money on AI, check yourself against these seven warning signs. If several apply, wait, or fix something else first.
1. You can’t name the problem in one sentence
“We need an AI strategy” is not a problem. “Our team spends 20 hours a week copying enquiries from email into our CRM” is.
If the reason for an AI project is that competitors are doing it, or the board asked about it, stop. Find one specific, repetitive, measurable task that costs you time or money every week. If you can’t find one, you don’t have an AI project yet. You have a curiosity, which is fine, but don’t pay for a build to satisfy it. If you want help finding the real problem first, AI consulting agency is a better first purchase than a build.
2. Your data is scattered and nobody owns it
This is the most common reason AI projects disappoint, and the least discussed.
Many businesses buy ChatGPT or Claude for their team and assume they now “have AI.” Then someone asks a real question about the business, such as “which clients haven’t ordered in six months?”, and it can’t answer. Nothing ever gave it access to the company’s own information.
The model is like a brilliant new employee who has never been shown the filing cabinet. If your customer data lives in a CRM, three spreadsheets, an old accounting system and a decade of email, with the same customer spelled four different ways, no AI can fix that by being smarter. Someone first has to gather the information, make it consistent and put it where the AI can reach it.
If that describes you, your first project isn’t AI. It’s organising your data. That work makes every future AI project cheaper.
3. Your key tools have no way to connect
Here’s a test that takes about 90 seconds. List the main software your business runs on and check whether each one has an API or official integration option (search the product name plus “API”).
If your core tools do, AI can almost certainly connect to them cleanly. If they don’t, the AI has to work the software the way a person does, by clicking through screens. That’s possible, but it’s fragile, more expensive, and breaks whenever the software changes.
If your most important system fails this test, think about whether replacing that system should come before automating around it.
4. A single mistake would be very costly and hard to catch
AI models do not give the same answer every time, even to the same question. That’s how the technology works, not a defect a vendor can patch. It’s also why lawyers keep getting sanctioned for filing AI-invented court cases.
That doesn’t rule out AI in high-stakes work, but it changes the design. If one wrong answer could cause serious harm and you have no practical way for a human to check the output before it takes effect, don’t automate that step. Use AI to prepare the work, and let a person decide.
5. The work depends on relationships, not repetition
Some customer interactions are the product. A financial adviser calming a client in a market crash. A senior account manager rescuing a relationship. A clinician delivering bad news.
Klarna is a useful example. In 2024, the company said its AI assistant was doing the work of hundreds of customer service staff. By 2025, its chief executive acknowledged that focusing so heavily on cost had hurt quality, and the company began making sure customers could always reach a human.
The lesson isn’t “AI customer service is bad.” AI is excellent for routine questions, out-of-hours calls and the first steps of a conversation. It’s that when the relationship is what customers are paying for, AI should support your people, not replace them.
6. The volume is too low to justify it
Automation pays off through repetition. A task that happens five times a month, takes ten minutes each time and rarely goes wrong costs you less than an hour a month. Building, testing and maintaining an AI system for it will almost never pay back.
A rough rule: if a task doesn’t happen at least daily, or doesn’t consume several hours of someone’s week, put it lower on the list.
7. You’re expecting a finished system in two weeks
Simple automations can go live quickly. But if you want AI that makes decisions and takes actions across several systems, plan for two to three months. Most of that time isn’t building. It’s testing against real situations, finding the cases nobody anticipated and adjusting how the system hands off to your staff.
If your expectations are set by a demo built in an afternoon, you’ll be disappointed by any honest vendor and over-promised by a dishonest one. Reset the timeline first.
Signs that you are ready
Now the other side. You’re probably a good candidate for AI if:
- You can point to a repetitive task that eats real hours every week
- The information it needs is already digital and reasonably tidy
- Your main tools can connect to other software
- A mistake would be easy to spot and fix, or a person can review before anything final happens
- You have a number to measure: hours saved, response times, error rates
If you tick most of those, AI is likely to pay for itself, often within months. For a quick check against these criteria, an AI readiness assessment will score where you stand.
What to do if you’re not ready yet
Not being ready isn’t failure. It usually means one of three things comes first:
- Map the process. Write down how the task actually happens today, step by step. You’ll often find improvements that need no AI at all.
- Organise the data. Pick one source of truth for customers and key records.
- Fix the tools. Replace the system that can’t connect to anything, if it’s central to how you work.
Do those, and your eventual AI project will be faster, cheaper and far more likely to land among the 5% that pay off.
.
