← Blog

Are quick wins possible with AI? Yes, with tradeoffs

Bennett Data Science Quick Wins Artificial Intelligence

AI projects usually take a long time, but a useful result in a few months is possible if you scope it carefully. This post covers what a full project involves, how to plan a quick win, and the risks that come with cutting corners.

I’m talking about two to four months, where eight to twelve is typical, ending in something a company can use to automate work a person normally does, like matching a customer to the right product or answering a question during a purchase.

Why projects usually take so long

AI projects take time, especially at companies or divisions that don’t already use predictive analytics. Here’s the ten-step process we’ve used for years:

  1. Business understanding
  2. Data assessment and understanding
  3. Choosing candidate models
  4. Data preparation
  5. Building the pipeline
  6. Modeling
  7. Evaluation
  8. Deployment
  9. Monitoring and maintenance
  10. Optimization

Getting through the first five steps can take months, and that’s before any modeling happens in step six. Most of the time goes into building a solid pipeline that feeds clean, reliable data to the data scientists, and later into putting their models into production. I cover that part in putting AI into production.

So what if we skipped a lot of that work and built something that shows a quick win and actually works? It’s possible, but we’d be skipping important steps. An example shows what that means.

An example: predicting churn at a telecom company

Say a telecom company wants to know when a subscriber is about to leave for another carrier. One way to speed things up is to ask a key question in an early stakeholder meeting: “If you knew which subscribers would leave tomorrow, how would you use that information?”

Would a spreadsheet be enough? Or does a churn score need to appear in an existing system or dashboard?

If a spreadsheet is enough, the project gets much faster than one that has to plug results into an established platform. We’re likely talking about months saved.

Assume the spreadsheet is enough, and that the data is easy to pull each day. We build it, and every night we analyze the previous day’s data and deliver a report of the customers most likely to leave. It takes a couple of months.

What can go wrong

Here are the downsides of a quick-win project, with the questions to ask:

  • Do people know how to use it? If no one taught the retention team how to use the churn list, skepticism about the model can stop it cold.
  • Is the incoming data changing? Without checks for consistency, the pipeline can quietly start producing wrong results.
  • What happens as things change? Without monitoring, the model will get less accurate over time and deliver less and less value. I wrote about that in models need ongoing maintenance.
  • Does it create more work? A churn list adds calls for the people who contact customers. If it feels like a temporary patch, people lose faith in it, and that hurts future analytics projects too.

Is a quick win worth it?

Every case is different. Picking a project with a high payoff and watching for problems downstream can show early value and build support for bigger projects later. Just know that a quick win walks a fine line between useful and wasted. Zac Ernst makes a similar case in when “quick wins” in data science add up to a long fail.

Today we build that balance into how every engagement starts. A fixed-fee Discovery takes ten business days and shows what the result is worth before you commit to a full build. You can see the steps in how we work with clients.

Let's talk about what you want to improve

Talk to an engineer