Most of the AI advice aimed at small businesses is written by people selling AI tools. The claims follow a pattern — save twenty hours a week, ten times your revenue — and the examples are always a demo rather than a business that has been running the thing for a year.
We build AI features into real products. Some of them earned their keep. Several were quietly removed after nobody used them. This is what the difference looked like.
The pattern that separates useful from decorative
AI works when it is applied to a task with three properties: it happens frequently, the answer is knowable from information you already have, and a near-miss is still useful.
Answering the same twelve customer questions forty times a week fits all three. Deciding which supplier to sign a contract with fits none of them — it is infrequent, the answer is not in your data, and a near-miss is expensive.
Run any AI idea through those three tests before you spend anything. Most enthusiasm dies at the second one.
What genuinely works
Answering repetitive questions
The most reliable win. If your team answers the same questions about pricing, delivery, opening hours or process every day, an assistant grounded in your own documentation handles a real share of them at any hour.
The word that matters is grounded. An assistant answering from your actual content is useful. One answering from general knowledge will eventually invent a refund policy you do not have, and that single incident costs more than the tool saved.
First-draft writing
Product descriptions, listing copy, social captions, email replies. Genuinely faster, provided someone edits before publishing. The failure mode is not bad output, it is unedited output — publish enough of it and your site reads like everyone else’s.
Extracting structure from mess
Underrated and often the highest-value use. Pulling fields out of invoices, categorising incoming enquiries, summarising long threads, tidying inconsistent spreadsheet data. Boring, repetitive, and precisely what the technology is good at.
Translation and tone
For Nepali businesses selling internationally, drafting in one language and adapting for another is a real, immediate saving. Have a fluent human check anything customer-facing before it goes out.
What consistently disappoints
Fully autonomous customer support. Deflecting repetitive questions works. Removing humans entirely does not, and it damages exactly the customers whose problems were unusual enough to matter.
AI-written blog content at volume. Publishing large amounts of unedited generated content is a well-documented way to acquire a lot of pages that rank for nothing. Google’s guidance is about helpfulness rather than how text was produced — the problem is that volume-first content is rarely helpful.
Predictive analytics on thin data. Predicting which leads will convert needs a meaningful history of leads that did and did not convert. With two hundred records you get confident nonsense.
Chatbots on sites with no traffic. A support bot on a site receiving thirty visitors a month solves nothing. Fix visibility first.
The honest cost picture
| Approach | Setup effort | Ongoing cost | Suits |
|---|---|---|---|
| Off-the-shelf tools (writing, transcription) | Hours | Per-seat subscription | Individual productivity |
| No-code automation between apps | Days | Subscription + usage | Routine handoffs |
| Assistant grounded in your content | Weeks | API usage + hosting | Support deflection |
| Custom feature inside your product | Months | API usage + maintenance | Something core to the offering |
The cost people forget is maintenance. Models change, prompts drift, your own content moves on. An assistant nobody has reviewed in eight months is confidently answering questions from information you no longer stand behind. Budget for someone owning it, or do not build it.
Start where the work is repetitive
A sequence that tends to work:
- Write down what your team repeats. One week, note every task done more than five times. That list is your candidate set, and it is more reliable than any vendor’s use-case page.
- Pick the dullest item on it. Not the most impressive — the most repetitive.
- Try it manually first. Do the task with a chat tool by hand for a week. If that does not help, automating it will not either.
- Automate only after the manual version proves useful.
- Measure the thing you claimed. If the justification was hours saved, count hours before and after.
Steps three and five are the ones that get skipped, and skipping them is how businesses end up paying for tools nobody opens.
Two things to decide before you start
Where your data goes. If you send customer information to a third-party model, know what is retained and under what terms. For anything sensitive this is a decision to make deliberately, not discover later.
What happens when it is wrong. Not if. Every AI feature needs a defined failure path — a handover to a person, a “I’m not sure” response, a human check before anything is sent. Features without one eventually produce the incident that ends the project.
Frequently asked questions
Is AI worth it for a small business in Nepal?
For repetitive text and data tasks, often yes, and the cost is low enough to test cheaply. For anything requiring judgement about your specific market, it is a poor substitute for someone who knows the market. Start with the dull tasks.
Will an AI chatbot replace my support team?
It should not, and attempts to do so tend to go badly. The pattern that works is deflecting repetitive questions so your team handles the ones needing judgement, with a clean handover when the assistant is out of its depth.
Can I just use ChatGPT instead of building something?
For internal productivity, usually yes, and you should try that first. You need something built when the task must happen automatically, be available to customers, or work from your own private data.
Will AI-written content hurt my SEO?
Google evaluates whether content is helpful, not how it was produced. In practice, high-volume generated content is rarely helpful, which is why it usually underperforms. Generated first drafts edited by someone who knows the subject are a different proposition.
How do I stop an assistant inventing answers?
Ground it in your own content so it answers from your documentation rather than general knowledge, instruct it to say when it does not know, and have it cite the source it used. Test it deliberately with questions it should refuse.
What does it cost to add AI to an existing website?
A grounded assistant is typically a weeks-long project plus ongoing usage costs, and the scope depends mostly on how much content it must draw on and how many systems it integrates with.
If you want to try something
Pick the single most repetitive thing your team does with text, and do it by hand with a chat tool for a week. That week will tell you more than any vendor demo.
If it works and you want it running properly — grounded in your content, integrated with your systems, with a sensible failure path — that is what our AI chatbots and assistants work covers. We have also written about which AI features actually moved metrics in products we have shipped.
