Take it slow with AI

I want to get a little dense with tech next week. We will talk about knowledge graphs and how they are better than LLMs for many industry-focused uses. But before that, there is the ‘Part 2’ of ‘Enterprise AI Adoption’ that we need to talk about. Last week, I looked at it from the vendor’s POV. This week, its from your POV.

AI is a bit unpredictable like the french quarter of New Orleans, which I had visited recently. It’s all fun and dandy on the main streets. Everything is magical. If you go to one of the side alleys, someone can take you out and your wallet before you say “Laissez les bons temps rouler”

• What if AI spews non-sense to my customers and even worse, what if we get sued for wrong information?

• What happens to our data?

• Who do we work with? No one seems to know this well!

• Our vendors are either old-school or don’t want AI to reduce their billing

We get these questions from potential customers in traditional industries. Assuming you find a studio that is data and AI savvy, they are going to be fast. They work in cycles of days and you work in cycles of quarters or even years. How do you reconcile? Where do you start?

Take it slow.

AI is easy to build. What’s hard is conceiving what delivers the most value, in a non-threatening way. That wisdom still comes from you and not AI. The right question to ask is

“What can we launch to our customers in two months that we can test deeply for a month?”

Let’s say you run a power utility. Of course, there is great value in automating customer record updates in your CIS system using agents and delivering information back and forth with your ERP. But that’s high stakes and takes time.

But what if it is outage reporting and billing inquiries? What if a chat bot or a human-like voice agent handles all the front-line questions for your residential customers all day, every day?

(If you are on the other side, I know you are side-eyeing me for this ‘chatbot’ idea. What’s interesting is not what generates value. What generates the best value is not always to easy to get buy-in for. What gets buy-in is not always where clean data is. I can go on and on, but your side eye will strain.)

A very limited, low-to-no integration use case like a chatbot with a billing system behind it has less moving parts. It is easy to build, easy to put guardrails around, and it demo-es very well.

In fact that is what we did in New Orleans last week. Our partner (Exceleron) launched Billie, the billing and outage chatbot for utilities. Please ignore the background noise, unless you understand what they are saying in Bavarian accent. If its anything good, tell me.

Billie now works with any registered US phone number and it is currently being tested with utilities. If I have to bring out a lesson from reflecting on it, Billie went from concept to demo in 3-4 weeks and to production in another 2-3 weeks.

The best way to incorporate AI within your organization is to get familiar with it in your midst and within your workflows. We picked a low stakes, easily reversible idea where there were alternative means for customers to get their job done.

Once you gain confidence, you may expand the footprint to hook outage status responses back to the customers. You may hook in a message to the field service team. It truly is a case where the surface area expands from win to win, month to month.

But, there is a catch.

You don’t build any system by focusing only on the parts. You need the whole to work well for the system to be stable. At the minimum you need to avoid reworks, and at the extreme end you don’t want multiple AIs and traditional systems crossing wires.

So plan for two tracks:

1. Rapid cycles of ‘experiment to production’ of low-stakes, reversible features

2. Deliberate design of your enterprise AI roadmap starting with processes, systems, integrations, blind spots, customer journeys, and value maps

The second track informs the first track while the first track gives legitimacy, urgency, and budget for the first track. Do a cycle of release and then kick off a roadmap exercise. Pick an easy win and then go for an enterprise-wide roadmap. Iterate over time instead of making it out a ‘point in time’ exercise – because AI is evolving fast, protocol level building blocks are falling in place, and new winners at the infra-level are beginning to show up.

So if things are evolving should you take it slow and revisit later? Well, who do you think I am? So you know my answer. But seriously, your ChatGPT-using residential customer is not going to put up with IVR for long and the nosy neighborhood competitor might just decide to get on with building.

As I said earlier, let’s talk about knowledge graphs next week. They are cool, unlike the AI that gives off a Frankestein-gets-Alzheimer’s vibe.