How to implement AI in business: start with the data

A client came in to work with me recently, very excited, with a whole bunch of ideas. Since it was our first session, he didn't really know yet what he wanted to focus on, and he started throwing out probably seven different things he wanted to do with AI. I could tell how excited he was about the possibilities, about being able to achieve so much more than he currently could.

Maybe you have a list like that too.

Where do you start with AI when you have too many ideas?

You start by picking one goal. I had to slow my client down and tell him, let's focus on one thing first.

Quite often, I start with two questions: how can we make more money, or how can we fix the workflows inside your business?

Making more money means finding more customers or improving the relationships you already have. Fixing workflows is the operations side: smoothing out bottlenecks and closing workflow gaps.

With my client, we talked for about five minutes about which of his ideas to focus on, and we landed on his social media presence, specifically Instagram, because we thought it would have the most impact on revenue, the fastest. He already had some traction on Instagram, but he clearly didn't have a plan. His product is very visual, and his ideal customer spends a lot of time on Instagram. That's where they're likely to find his product, or even buy it.

So that gave us a direction. Then we started coming up with a bunch of ideas for what to do, and I quickly realized what he was missing: data. Before we worked together, his Instagram data was nowhere. He just had whatever he saw in the app.

How do you get your data ready for AI?

Organize the data you already have, or set up systems and processes to start collecting it. Without data, you have nothing but your own intuition and guessing.

Most of the time, I'm pulling data out of a CRM or a lightweight ERP system, or a service like Zendesk, Asana or monday.com, and it usually isn't meant for use with AI. So we extract the data from the systems it already lives in, use AI to organize it, and then use AI to pull out what we need and analyze it in different ways, without paying a third-party data scientist to help.

For my client, the data was only what he saw in his Instagram app, so we spent that first session building a system around it. We pulled together the history of his stories, reels and static posts, and even some old ad campaigns. We set up a process to keep collecting new data, and we automated the analysis.

Then we decided what a winning post looks like. His goal right now is presence, building reach first. We're not worried about conversion yet; that comes later in the plan. So we looked for the posts that got the most reach and the most impressions, then at shares and watch time. We chose those top-of-the-funnel numbers on purpose, instead of clicks or conversions, because they match the goal.

Now the plan is to iterate off those winners instead of guessing at new concepts. We know what's working, so we have a guiding light. And because the data is collected and organized, we can build a loop with Claude, the AI model we work in, so Claude knows where we've been, what we've tried and how it did, and we can keep making the content better for his audience.

I know having data may seem basic, and a lot of people already have apps that collect numbers for them. The difference here is that AI automated the collecting and shaped the data in a way that made sense to him specifically. Instead of learning another tool, he has workflows, spreadsheets and a dashboard built around his needs, so he's much more likely to use that data, and it's more likely to have an impact.

Customer history is another place I pull data from, when a company has it. One of my favorite things to do is extract the voice of the customer. As companies grow, leadership usually gets more and more disconnected from its customers, even though every business's job is to serve them. Without a good feedback loop from customers, it's hard for leaders to guide the company.

So if there's contact history in a CRM or a support ticketing service, or better yet, recorded phone calls, we get it transcribed and work with AI agents to extract the customer's own terms and language.

Speaking in your customer's language is very powerful in marketing. It can shape everything from ad copy to company communications to how your salespeople talk with customers. That's a well-proven direct marketing tactic that used to be hard to pull off without very fancy marketing software. With AI, as long as you have enough of those conversations, AI agents can find that language for you.

What tasks should you automate with AI first?

Automate the quick wins first: the ones with the most impact on your costs or your revenue, where AI mostly needs information you already have, organized. Here are three kinds I see.

Support tickets are one. A business we worked with had a lot of old technical support tickets. We took the data, organized it, and analyzed what the most common issues were. Then we used those common issues to front-load their customer support FAQ and answers.

Most tickets come from a few common issues, the 80/20 rule, so answering those up front has greatly reduced how much human support they need. When most customers can serve themselves, they get their answer a lot quicker without having to talk to anybody, and the company carries less of a support burden.

And that data keeps working. The common themes in those tickets can become videos, handouts and product inserts, or shape how the product and service pages on the website talk to a prospect, because now you're anticipating their objections as well as their needs. It makes you much more focused on the customer.

The data also tells you whether it worked. Once capturing and processing data is a normal part of your business, you can track and measure. On the support side, the measure is fewer tickets. If a first level of support answers the common issues, you count the tickets that still go to a person and watch whether they go down over time. With real evidence like that, you can be confident your decisions are steering the company in the right direction.

Accounting is the second kind, with a lot of quick, easy wins, most of them from getting a company's financial and accounting information together in an organized form, then letting AI work with it. AI makes a very good intermediary between a company's financial information and QuickBooks. It isn't necessarily challenging, but it's critical that the company centralizes its financial records first. Once AI has an organized data structure to work with, along with some specific prompting and workflow instructions, it can become very reliable.

The third quick win is getting the thoughts and information out of your head and into a structured format AI can use. At our AI agent workshops, I can see the light bulbs going off for many participants when we start creating the core documentation about their companies. Once they begin to identify and centralize the information about their company, like their customers, products and vendors, AI has a way to be very useful to them.

What shouldn't you automate with AI yet?

Hold off on the expensive, data-hungry projects until you have the data to support them. That usually means voice agents and trying to replace employee roles.

A lot of people want to get straight into the very sexy stuff. AI can do all of it. We've seen the headlines and watched the demos. But the reality is that these projects are very difficult to do. They're often quite expensive to put together, they take a lot of time, and sometimes the impact on your P&L isn't obvious.

A lot of people want to get straight into the very sexy stuff.

Voice AI is a good example. It works for me: Millie, my AI receptionist, lets me focus on deep work and my customer relationships while keeping my speed to lead extremely fast. But for a business with plenty of employees available to answer the phone and answer questions, an AI receptionist may not make sense, given how much time and money it can cost.

Then there's the data. The agents people show off and get excited about take a lot of it to train or tune properly, and many companies aren't capturing it, or haven't captured it consistently enough. Gartner's survey of data management leaders found that 63% of organizations either didn't have, or weren't sure they had, the right data management practices for AI. We can work with synthetic data, but it doesn't capture true human behavior.

I have this exact situation with a large client right now. He wants to implement voice agents, but his company has no recorded phone calls, so there's nothing to work with whatsoever.

I told him, let's work on other parts of the business first, and once we actually have some phone call data, we can start on the voice agents. He understood completely. He realized what he'd missed was time, all the calls he never recorded, and that data is going to drive everything in the AI era.

Now I'm working with him, his CTO and the IT department on a phone system that records and captures the calls, so we can feed them into AI. Later, those calls become the data that makes his voice agents useful to his customers, instead of just answering the phone and routing calls to human support agents.

So start capturing now, even if the data is messy. Data can always be smoothed out and massaged. But if you don't have it, you don't have it.

It all comes back to getting data, and getting it in a structure AI can use.

Whether your data is already organized, spread across a few systems or not captured yet, reach out to us and let's discuss it. We'll work with you on getting it ready, and on whether there's an AI tool or agent worth building on top of it.