Building your AI knowledge base: A recipe for propane marketers

In my last column, I suggested that a propane marketer’s most valuable artificial intelligence (AI) asset is its knowledge base.

For the second article in this series, I’ll explore how to create your knowledge base.

I do almost all the cooking in our home, and I sometimes do a bit of thinking while preparing a meal. And it occurred to me while thinking about this column that building an AI knowledge base is a little like following a Julia Child recipe. You read the list of ingredients and think, “Well, this doesn’t seem too difficult.”

Until you start cooking.

The ingredient list for an AI knowledge base is deceptively simple: Gather the information your company uses to answer questions, train employees, make decisions, serve customers and operate the business. Put it in digital form. Organize it. Decide who can see what. Add AI.

Bon appétit.

Unfortunately, the actual recipe quickly gets complicated.

The good news is that you already have most of the ingredients. You just need to figure out where you put them. Your company’s knowledge is scattered among binders, spreadsheets, software notes, old emails, safety manuals, pricing sheets, service tickets, filing cabinets and employee memories – not to mention the occasional note written on a restaurant or bar napkin brought home from a propane convention.

▶ Step 1: Get the ingredients onto the counter

Start by getting as much of that information as possible into digital form. We began digitizing our propane company’s records and eliminating paper forms seven or eight years before we ultimately sold the business.

Documents were relatively easy. Historically recorded telephone calls buried in third-party databases were harder (we decided to start fresh with a new, AI-supported system). Hardest of all was the accumulated knowledge walking around in employees’ heads. We’re pretty sure we left some – perhaps a lot – of that ingredient behind.

▶ Step 2: Don’t throw everything into the same pot

Some information can be used freely: your website, public safety information and manufacturer manuals. Some belongs in an employee-only knowledge base: operating procedures, forms and training materials.

Customer and financial information requires much tighter controls. And some information – employee health records and tax returns come to mind – probably doesn’t belong in an AI knowledge base at all without very specific reasons and permissions.

External information needs its own consideration. Regulations, Propane Education & Research Council and National Propane Gas Association resources, manufacturer manuals, state association materials and weather data may be public, but public doesn’t necessarily mean accurate.

The rule is simple, even if implementing it isn’t. Decide what information AI may use, who may use it and for what purpose. That’s your firewall.

▶ Step 3: Taste before serving

Don’t try to build the entire company knowledge base at once. Pick one process – answering common customer questions, organizing service documentation or helping new employees find procedures.

Give the AI the info it needs, then start testing. When an answer is wrong or incomplete, figure out what source material is missing or weak and fix that before moving on. Then, like the dishes after a good meal, rinse and repeat.

This may be the most important part of the process. In a kitchen, you can follow Julia Child’s instructions perfectly, but the finished dish still won’t taste right if you leave out a key ingredient. AI works the same way. Leave important information out of its knowledge base, and the finished product won’t be what you need or expect.

The most difficult ingredient may be the knowledge nobody has written down. Much of what skilled, experienced employees actually do involves judgment they’ve accumulated over years and now exercise almost without thinking. Getting them to identify and explain these important decision points is surprisingly difficult. AI itself can help by asking questions, identifying gaps and turning that institutional memory into usable information.

So, yes, the recipe is more complicated than the ingredients suggest.

Christopher Caywood is a board member of Guardian Propane Partners LLC. Contact him at ccaywood@guardianpropane.com.

Featured homepage image: everythingpossible/iStock / Getty Images Plus/Getty Images

---

Related articles

An AI primer for propane retailers

Connecting the dots: From operating efficiency to enterprise value

Relationships first, technology second