Personal Insight From One Gen AI Journey as You Embark on Yours

I would have tried to understand more precisely what the estimates were of the team's work and what the inputs to that estimate were. We had six weeks of availability, and we got the initial deliverables done in two or three weeks.

But I could have been more prepared to have more scopes of work lined up and ready to go, if I had had a better understanding of what the team was going to be able to do in the time allotted. I'm not sure if we could have fixed that because of the nature of Gen AI projects. It was new to all of us, including the people on my team who had to estimate the work.

We probably could have also better described what success of the proof of concept looked like. We went into it with an “anything is better than nothing” approach, because any results that we get will at least help us know what not to do when we do Gen AI work internally. Literally, any knowledge was good knowledge.

I'm not sure if we would have done anything differently. We had two executive sponsors for this effort — me and our vice president of AI — thinking about this a lot. We had been putting together a list of internal and external use cases that we had running for five or six months before we started this engagement. So, I feel like we had a robust list of at least high-level use cases.

The trap you can fall into is seeing Gen AI as this shiny object; there’s a temptation to apply it to everything. But it’s super important to focus on things that are immediate needle movers for your customers — making it fast and easy for the customer. Because if you're not doing something that makes an impact for your customer, then why are you doing it? It's a waste of time.

Remember, for many, AI feels like magic. People have higher expectations for magic. So don’t present something to your user that just doesn't work or is half-baked. If they get something and it works and feels like a win for them, that's a victory. But anything that requires them to do work or change their usage patterns or anything else that requires mental effort feels like a loss.

(Specific details about the use cases ServiceTrade tackled are in the case study here​​.)

First, our executive team fully bought in on AI for our internal operations and our product roadmap. So, the question for us was which workflow improvements would create the best outcomes for our customers for which they will pay. If your Gen AI project gives you the data needed to prove the benefits for the end user workflow are strong enough to be monetized, you’re probably good to go. If you can’t, then you should think about another project.

Second, naming the solutions for which AI didn’t pan out as strong solutions is just as valuable to know, so you’re able to create those solutions in other ways.

We felt tons of urgency. My biggest concern is that we can't ship meaningful capabilities to our customers fast enough, losing a competitive advantage on, at the very least, doing simple Gen AI things like taking a pile of unstructured text and writing a summary of it.

In the meantime, there were some other projects that we started internally slightly before being used by our customers in a closed beta format. I expect we will end up taking the asset summary capability to market in the same way within the next couple of months.

So, there's a lot of urgency to get usage and customer adoption.

Everybody is — us included — being expected by our customers and by our investors to explore how Gen AI could make our business internally more effective and bring our customers more value.

We have many different use cases in our customers’ businesses of people not having the time or bandwidth to consume all the information they need to do their job effectively. Because there's too much content or they don’t know what to pick out and what to ignore. It all kind of boils down to, “I'm not going to read all this documentation and all this history and all these comments. I can't do that 1,000 times a day.” That's the root of several different use cases that our customers face.

Gen AI is good at taking big piles of content and turning it into five bullet points. This is a no-brainer approach to several of our customers’ problems.

Yes. We were convinced that Gen AI would be just the thing to help us with route optimization. We wasted the better part of a week on something completely not a problem where Gen AI could be a solution.

But it was good that we did that because we quickly learned that Gen AI is great at fluency and bad at precision. So, anything you need precision for, don't do it. Don't even try.

It almost feels like a hammer in search of a nail. You’ve got a hammer; so now you say, “What can I hammer down?”

You must make space for distraction, particularly with bleeding-edge technology. Because if you don't, then you won't produce any neat, novel innovations. That's the sort of exploring you do and finding the stuff that doesn't work to help you discover the things that do work you didn't expect.

It all comes back to knowing what sort of problem you're trying to solve for the customer and what use case you're going to hit. Start with the problem you're trying to solve and work backwards, such as a text content consolidation problem. That’s a good use of Gen AI. But, if it’s a difficult math problem, that's not a good use of Gen AI.

The place where we felt that most acutely was when we needed specific technical knowledge about Gen AI techniques that we didn't have. The landscape is moving so fast that it was literally not possible to keep up by reading documentation — by doing the things you would normally do to understand how to do something.

For example, there’s a technique called retrieval augmented generation, or RAG, that’s common when you have a big set of content and you want AI to answer questions based on, for instance, your knowledge base or your support site. We needed to focus on our content and how to pull pieces out of it to solve the problem. But we didn't know how to do it. And it was hard to get a definitive answer on how to do it.

Having somebody come in with expertise was super valuable. It ended up being valuable for other projects, too. Many techniques we use on our internal use cases are ones we learned from the SoftServe data science team.