AI for Small Business—Boom or Bust?
Brian Long
AI is the biggest revolution since the Industrial one!!! AI is overhyped and the bubble is just about to pop!!! Which is it? It can’t be both, or can it? Even if that’s a bit unclear at the moment, it is very clear that billions of dollars are being poured into it. Billions upon billions even while 95% of generative AI pilots are failing according to an [MIT report](MIT report: 95% of generative AI pilots at companies are failing | Fortune).
Well, which is it, boom or bust? Weirdly, it’s a bit of both. And for small and medium businesses that’s great news even. Let’s talk about why.
The Hype Cycle Keeps on Cycling
The hype cycle is a model that the research firm Gartner has used for decades to track emerging technologies. It goes like this: A new technology emerges, we’ve never seen anything like it and it seems like it can do everything. But then everyone piles onto the inflated expectations and reality comes crashing in. Then, the hype around the technology falls off and falls into what Gartner calls “the trough of disillusionment.” Projects get canceled and budgets get cut. The media media, ever fixated on The Next Big Thing™ writes obituaries. Then the interesting part happens: the people organizations who actually understood what the technology is actually for keep quietly using it. They just quietly keep on building with it to create strategic advantage and mature the technology and it turns into foundational technology.
The internet went through this exact cycle. The dot-com crash of 2000 didn’t mean that the internet was a fad, but it did mean that pets.com was a bad idea. The companies that actually understood how to use the web to accelerate a good business model kept on building and are now dominant.
Generative AI is reaching the end of the hype cycle and entering the trough. The MIT report mentioned earlier did show that most enterprise AI pilots fail to show measurable returns is real, but the cause is the interesting part. The study found the problem was organizational, not technological.
So, What is AI Good For Anyway?
First we need to clarify what we mean by “AI”. When we say AI is “Artificial Intelligence” we might be forgiven for thinking we actually mean intelligent like a person. That’s the goal, but not where we are right now. You may hear that referred to as “AGI” or general intelligence. What all of the hype now is all about language and is called a Large Language Model, or LLM. Granted language is critical for intelligence and especially for appearing intelligent. The human still has to provide the actual intelligence and the LLM acts as a language processor.
That reframes the entire picture of what it is and what it is good at. It’s not a search engine, database of facts, or a digital employee. It’s a language processor, a tool that takes in language and transforms it into different language out like a food processor takes in vegetables and gives you back chopped, blended, or pureed.
That framing tells you where it will shine: summarizing, drafting, rewriting, translating between tones and formats, or even another language. It will extract the important parts from an impenetrable wall of text, or turn your rough notes into something presentable. Most small businesses run on exactly this kind of language work. Customer emails, estimates, job postings, review responses, proposals, policies, social media postings; it’s a huge chunk of the week and almost none of it your reason for starting a business in the first place.
It also reveals where to need to be careful. A language processor doesn’t know things the way a reference book contains facts. It generates plausible language, something that sounds nice. That means you must verify facts, numbers, and especially anything with legal weight. You wouldn’t ask a food processor to plan your menu. You shouldn’t hand an LLM work where you cannot check the output.
The failed enterprise projects mostly ignored these realities. They asked the technology to be something it isn’t. You don’t have to make that mistake.
Getting Good Output: A Thorough Briefing, Not Magic Words
Disappointment with AI comes down to this: type in a vague one-line request, get a generic or made-up answer, and conclude that it’s useless slop. That would be like handing a new employee a sticky note that says “deal with the customer thing” and being surprised when the result is nowhere near what you actually wanted.
The fix isn’t to use magic words. It’s to give a thorough briefing just as you would give a skilled temp on his/her first morning. We’ll build one in four parts: task, role, context, and examples.
Let’s begin. Meet Maria, who owns a twelve-person plumbing and AC company. A customer just left a frustrated three-star review: the technician was great, but the appointment started two hours late and nobody called ahead. Maria wants to respond well, but has eleven other things to do before lunch.
Task: Say Exactly What you Want
The task is the deliverable, specific and to the point. It’s the thing that should exist when the AI is done that doesn’t exist now. Vague tasks get mediocre results. “Help me with this review” might get a tangent lecturing on the importance of reputation management.
Compare this:
Write a public response to this customer review, under 120 words, that apologizes for the late arrivval, explains that we’ve changed our dispatch process, and invites the customer to call me directly.
Notice everything that sentence pins down: format (a public review response), length, and the three key points to hit. The single biggest upgrade most people can make to their results is simply finishing the sentence: “When you’re done, I will have a ___ that I can use to ___”
Role: Tell it Who to Be
An LLM can write in a thousand voices, so tell it which is yours. Assigning a role sets the tone, vocabulary, and perspective in one stroke:
You are the owner of a small family run plumbing and AC company writing in a warm, direct, no-corporate-speak voice. You take responsibility and never sound defensive.
Without this, Maria gets the standard corporate apology: “We sincerely apologize for any inconvenience this may have caused.” Every customer has read that sentence a hundred times—and believed it zero times. With the role, the response sounds like a person who actually answers her own phone.
Context: Give it What it Can’t Possibly Know
Here’s the part the “language processor” model makes obvious: the AI knows the language, but it knows nothing about your business, your customer, or this situation unless you give it exactly that information. Context is the background a trusted employee would already have:
Background: The customer is Dave R., a repeat customer (three jobs over two years). The delay happened because an earlier job ran long and our old scheduling system had no way to alert the customer. As of this month we use automated texts when a tech is running more than 20 minutes behind. Here is Dave’s full review: [paste review].
Context is where generic output becomes your output. It’s also the part easiest to skip because it feels like extra typing. It isn’t extra! It’s the actual job. Thirty seconds of pasting beats four rounds of “no, that’s not quite right.” And context is the key to the truly useful part of AI we are building now.
Examples: Show, Don’t Just Tell
The fastest way to communicate style is with a sample. If Maria has a past response she was proud of, she includes it:
Here’s a response I wrote last year that captures the tone I want: “Frank you’re right, and I’m sorry. No excuses: we missed the window we promised you. I’ve credited the service call and your next maintenance visit is on us. Maria”
One good example outperforms three paragraphs of adjectives. Describing your voice as “friendly but professional” leaves a thousand interpretations while showing it leaves just one. This works for formats too; show it a past proposal, estimate, or job posting, and it will match the structure without being told.
Stack all four pieces together and Maria’s “AI experiment” stops producing fortune-cookie filler and starts producing a response she would actually post. All in about two minutes, with one quick read-through before it goes live. If she wants to retype it to make sure it truly uses her voice, maybe there is a word or punctuation that she wouldn’t use, the task is done in five minutes and on with the day. Task, role, context, examples. The same briefing works tomorrow for the job posting, the estimate follow-up and the holiday-hours email.
The Bust Must Be to Have Room to Boom
Now, notice what Maria did not do. She didn’t hire a consultant or buy an expensive enterprise platform. She simply found a piece of recurring language work, briefed the tool properly , and kept the judgment call, the final read, for herself. That’s what the winning side of the hype cycle looks like at small business scale: not a moonshot effort, just a growing list of two-minute wins that used to take twenty. And, all those small wins add up to make a small business boom.
The trough of disillusionment is when the this advantage is built, precisely because the hype is gone and nobody is watching. The quiet is needed to build the thing that actually works.
This article is the first in a series. In the coming weeks we’ll dig into where to find those recurring language tasks in your own operation, how to verify AI output efficiently, and when you know you are ready to go beyond the chat window. At Synapses Technologies, we help small and medium businesses scale enterprise technologies to fit them, finding the practical fit for AI, without the hype and past the backlash. It’s what we do, so when you are ready to find your first two-minute win, we’d love to help.