Let’s take a moment and go back in time, courtesy of Mr. Peabody’s WABAC machine, to the decade spanning the mid-1990s to the mid-2000s. That was when Microsoft introduced Access to the corporate environment. It rapidly spread as thousands of office workers eagerly began building homegrown database applications to support their jobs. Information that had existed in Excel spreadsheets and manual processes was transformed into data entry forms and reports, housed in tables—all parts of a relational database.
In some cases, those applications grew to encompass more than just the department that built them. In many other cases, data center managers frequently (sometimes angrily) quashed the notion of a desktop database accessing mainframe datasets and providing information to unknown, unverified end-users.
Nonetheless, despite inadequate security controls, Access proliferated until most of those applications became orphans as their owners left for other jobs, leaving no one who understood how to maintain them.
I am now getting a sense of déjà vu, because many small businesses are developing their own Artificial Intelligence (AI) projects. What makes me wonder if this is Access all over again is that a “real” AI pilot project requires tremendous effort, whereas some of the developed apps I’ve seen lately are less than stellar.
Let me start by establishing a baseline. AI implementation projects for small businesses (SMBs) are a huge group effort. They require detailed research among various business process owners.
Soon after the start, it often becomes clear that the SMB’s processes are poorly documented and inconsistent, and that not all components are easily available or accessible (i.e., they live in various files, email threads, and individual staff members’ brains). That is usually because what is commonly known isn’t written down or codified.
This disjointed (or insufficient) information prevents them from building the right environment for an AI pilot. Whenever I’m involved in an effort like this, I usually find it necessary to enforce the equivalent of a technology stabilization pre-project phase – an “information stabilization,” if you will – to ensure the data that will be provided to the AI agent is consistent, easily accessible, correct, and approved.
If you can document a manual workflow – and have it performed without error or external interference (meaning someone had to do something outside of the documented steps), then you are safe to proceed with a pilot. Otherwise, you must go back and further improve and refine the manual workflow. Sometimes, you must also perform additional cleanup and documentation.
I am sure some business owners will object to spending time (and money) on this preliminary phase. If they do, there is no likelihood of a successful AI project. Because one thing is certain: AI agents will confidently tell you things that are flat-out wrong. You don’t want to base your business on something like that.
In rare cases, after completing the information stabilization phase, a business owner realizes they have streamlined several disorganized processes and decides not to continue with AI agent implementation. After all, saving slightly more at this point will likely be offset by the actual cost of AI agent tokens in the long run.
It almost goes without saying that it takes significant time to become proficient at any task. When you work with AI agents, learning how to prompt the agent to obtain the results you want is the most meaningful skill. Because AI agent tokens cost money, people sometimes shirk developing this skill. If you don’t take the time to build proficiency, you might end up with a worthless workflow or invalid/incorrect results. You also must consider what your customers would think about invalid invoices or incorrect work orders; worse still, if your financial calculations turn out to be fabrications. At each step, you must manually verify the results. And you cannot become complacent about how often you check those results once you start using the AI project in production. One mistake can have drastic consequences.
Agentic AI, where most things are automated up until it is time for a person to be involved, is the goal around the corner. But I am going to be a Luddite and say, let the enterprise world of Fortune 100 companies – that have the funding – blaze a trail here. Let’s learn from all their mistakes and take advantage of their coffers to perform these projects – before we embark on that phase of the AI SMB journey.
I’m looking to see how many of you raise your hand if you want me to help you begin an AI pilot. Otherwise, I’ll wonder if these new AI projects will go the way of Microsoft Access applications.
Thanks, and safe computing!