A company buys ChatGPT, Claude, or switches on Copilot inside their Microsoft environment. Licenses get assigned, someone schedules a training session, and then comes the question nobody planned for.
What do we do with it?
That question is the problem, and it shows up after the money is spent. These platforms do not solve problems on their own. They have to be given a problem to solve, and someone has to understand that problem well enough to set the platform up for success.
The old software purchase model trained us wrong
For decades, we bought software designed to do one thing. We identified a problem, purchased the tool built for it, and the tool did that one thing. The problem was defined before the purchase order was signed.
AI breaks that pattern. It arrives with a world of knowledge beyond our own comprehension and no idea what your business does. It waits for instructions.
You are no longer taking a tool off the shelf as-is. You are building the tool to fit the problem, which means you had better know the problem first.
RAND Corporation studied why AI projects fail, based on interviews with 65 data scientists and engineers. They found that more than 80% of AI projects fail, roughly twice the failure rate of conventional IT projects. The number one root cause was not the technology. Organizations misunderstand or miscommunicate what problem needs to be solved, so models get optimized for the wrong metrics or never fit the business workflow at all.
Read that again. The most common reason AI fails is that nobody defined the problem.
What operational drag actually looks like in a business
Before applying AI to anything, a company needs to find where its operations are dragging. Asana surveyed more than 10,000 knowledge workers and found that 60% of their time goes to what they call work about work: communicating about work, searching for information, switching between apps, managing shifting priorities, and chasing the status of things.
That is the drag. It is sitting inside your company, and most of it has never been written down.
Here are the questions I ask to surface it:
- Which tasks are repetitive?
- Which tasks can be automated?
- Which tasks never get done because they seem too big to start?
- What troves of information do you hold that you have been unable to pull insights from?
- What administrative work needs doing that is not getting done regularly or effectively?
- Which software systems sit separately when they should be sharing information?
- How could one department benefit from data owned by another, and could AI connect that data, generate reports, and take action based on activity inside another team’s platform?
Answer those honestly and you have the start of an AI implementation plan. Work the list in order of what is costing you the most.
A quoting problem that almost cost six figures
I worked with a client eager to implement AI. Once we started, we found operational gaps that had to be filled first, and the questions were basic ones.
Where was their data stored? How was it shared across the team? What happened to the information they gathered from customers? Was any of it in a CRM, or was it scattered through email threads and spreadsheets?
They also had a problem providing accurate quotes. The pricing information lived partly in a spreadsheet and partly in the heads of different people around the company. One detail in that process had almost cost them hundreds of thousands of dollars.
The CEO was hands-on enough to know the quoting risk was real. What he had not considered was that AI could weigh every moving part before producing a price, and that none of it would work until the information was organized somewhere AI could reach.
So we did that work first. We located the data, organized it, and then surfaced it at the moment each person needed it, so the company could operate on shared knowledge instead of individual memory. The quotes got more accurate because the small details stopped slipping through.
That client is not unusual. Gartner found that 63% of organizations either do not have the right data management practices for AI or are unsure whether they do. The “or are unsure” part matters most. A company cannot fix a condition it cannot describe.
Leadership is not where the answers to operational drag live
Leaders often cannot point to their own drag. They are too far from daily activity to see where the work snags.
Go to the people doing the work. Approach them in a way that is not threatening to their jobs, and help them understand you are looking for ways to make their work more accurate and less tedious.
Then talk to middle management, because that is where the best view sits. They understand what has to happen strategically and tactically, they are close to the outcomes, and they tend to feel less threatened by the idea of AI handling work more efficiently.
Expect more drag to surface during the build itself. Questions come up about where a piece of data lives, and the answer reveals a gap nobody had named. You have to locate the data before AI can use it.
What AI gives you once the operational drag is found
The benefits arrive in a specific order, and the order matters. Nobody gets speed, accuracy, or scale from buying the platform. Those come from applying it to something already identified as broken.
Fewer missed details. The quote that almost cost that client six figures was not a math error. It was one detail sitting where nobody thought to look. When information lives somewhere AI can reach, small things stop falling through.
Shared knowledge instead of scattered knowledge. Adding people to a company does not spread what the company knows. Hire ten more and you have ten more people asking where the file is. AI becomes the expert your whole team can talk to at any hour of the day, and it only fills that role when its knowledge is rooted in the context of your business.
Work that was never feasible before. Real-time market research on every deal used to be off the table because nobody had the hours. Connecting two software platforms meant a queue with IT and a timeline that killed the idea. With AI, both are reasonable now, and that is capacity you did not have at any headcount.
Initiatives that stopped getting shelved. Some of the best ideas inside a company die because the work required to launch them looks too heavy. Break the idea into its actual tasks, then look at which of those tasks AI can carry. A number of shelved ideas turn out to be feasible after all.
A measurable win. Leaders underrate this one. When you start from a known problem, you already have a before number, so you can prove the after.
How AI implementations end up labeled failures
Skip the operations step and you risk creating more noise and more work without solving any of the drag you already had. It happens two ways: you build new processes without considering operational need, or you point AI at a process that was broken to begin with.
Optimizing a broken process gives you a faster broken process.
The measurement problem is worse. Start from a brand new process and you have nothing to compare against, so you spend months trying to figure out whether the time and effort was worth it. You end up testing the process rather than the AI, and the AI gets written up as a failure for a problem it was never pointed at.
Finding and solving operational drag is measurable. You knew the problem, you can measure the reduction, and AI looks like the win it actually was.
The line between an operations fix and an AI fix
An operations fix is what you see when you look at inefficiency in your business processes, the activity of your employees, and the flow of information through your organization.
An AI fix is really an implementation. You implement AI to improve other areas of the business, and operations is the first place to point it.
AI is not there to fix your business. We implement AI to make businesses operationally more effective, more efficient, and ultimately more profitable.
What to do first, before you touch an AI tool
If you already own the license, the first move is not a prompt. It is a document.
Build a deep summary of what your company does and what it is trying to achieve. Include your business structure, what you sell, and how you sell it. Then get five things in order:
- A library of your standard operating procedures
- A list of your technology stack
- A list of your products and services
- Your pricing methodology
- A summary of your business and what it does in your industry
The SOP library is the biggest win of the five. Get one written for each repetitive task or project and AI can identify where it adds efficiency, because it finally has enough context to see how your work actually flows.
Context is what makes AI perform. Garbage context into AI produces garbage output from AI, and no amount of prompting fixes a model that knows nothing about your company.
My final thoughts: the companies getting real returns from AI did not pick a better platform. They found their drag first, organized what AI needed to see, and then implemented. It all comes back to operations.
Think about the seven questions above and ponder which would be hardest for your team to answer honestly. That answer is probably where your drag is hiding.
