Search is full of a third question that ate 2024-2026: how do I use AI in my business? The answers online are mostly product demos. The failures we see in Uganda and on remote jobs look the same: a chatbot on a site nobody visits, staff pasting client data into a public tool, and a subscription no one opens after month two.
Front Star Digital offers AI-powered business solutions as a named service with a pilot and a production package. We also build the websites and systems the model would sit on. This article is the filter we use before anyone trains anything.
What AI is good at in a small firm
Ignore the keynote. In a company of 5-200 people, AI earns its keep on narrow jobs:
- Drafting - first versions of emails, job ads, FAQ answers, product descriptions. A person still sends them.
- Classifying - routing support mail, tagging leads, sorting documents into the right folder.
- Extracting - pulling totals, dates and names from invoices or forms into a database you already own.
- Summarising - turning a long report or a week of tickets into a briefing a manager will actually read.
- Retrieving - answering staff questions from your manuals, not from the open web, with citations you can check.
Notice the pattern: the system of record stays yours. The model speeds a step. If you do not have records, AI will confidently invent them. That is not transformation. That is a liability.
What AI is a poor substitute for
- Being found. Models do not replace Google visibility or a site that converts. Buyers still search and still judge you on a phone.
- A missing operations system. If stock is a notebook, an "AI dashboard" has nothing true to show. Fix website vs system vs app first.
- Cybersecurity. A chatbot will not patch servers, test backups, or stop a staff member clicking a fake invoice. That remains cyber security and data protection work.
- Headcount magic. If the process needs a human decision (credit, diagnosis, hiring, legal), the model can prepare a pack. It should not be the signer.
A simple test before you spend
Write one paragraph with four lines. If you cannot, you are not ready:
- Job: what a person does today, in verbs (reply to quotes, code invoices, file student papers).
- Volume: how often, and what a miss costs in money or hours.
- Data: where the truth lives now (Excel, email, a system we already built, paper).
- Metric: how you will know the pilot worked in 30 days (hours saved, error rate, time to first reply).
If the metric is "we will look innovative", stop. If the data is personal or financial and you have no policy on where it goes, stop until data protection is in the scope.
Cheap tools versus a business integration
Using a public assistant for a blog outline is cheap and often enough. Wiring a model into your POS, school system or ticket queue is a software project: authentication, logs, prompts you control, fallback when the API is down, and training so staff do not paste passwords into the box.
Our packages are deliberately small:
- Pilot - one workflow, evaluation metrics, risk notes. You can kill it.
- Production - integrated system, monitoring, staff training. Someone is named after go-live.
We do not claim we built the underlying model. We do not sell fake certifications. We will not put customer records into a tool we cannot name in the contract.
Where this sits next to our other work
AI projects fail when they are isolated. The combinations that survive look like this:
- A website that already ranks for the service, plus drafting help for new articles you still edit.
- A custom system or product (POS, HR, school, hospital) that holds clean tables, plus extraction or routing on top.
- Data analysis when the question is "what changed last quarter", not "make me a robot CEO".
- Training so staff know what they may paste, what they must check, and when to escalate.
If your bottleneck is access control on a physical site, that is hardware and policy - access control and smart security - not a language model.
WhatsApp, chat widgets and "AI customer service"
This is the pitch small firms hear most. A bubble on the website that "answers 24/7". It fails in three predictable ways:
- The site has no traffic, so the widget talks to nobody.
- The bot invents prices, opening hours or stock you did not publish.
- Angry customers still want a human, and you have no ticket trail when the bot is wrong.
A better pattern: keep WhatsApp or the form as the intake, store the message in a system staff already use, and let a model draft a reply the staff member sends. That is classification plus drafting, with a person accountable. It is slower to demo and much harder to regret.
If you do not yet have a public page that states hours, services and how to pay, skip the bot. Fix the website. Then talk.
AI will not replace hiring - and hiring will not replace a process
Owners ask whether they should hire a "data person" or buy AI. Different jobs. A person who understands your books can use a model to go faster. A model with no owner will still need someone to check totals before they reach a tax filing, a bank, or a board pack. Budget for the owner first. The tool is second.
The same rule applies to developers. An AI coding assistant does not make an unnamed offshore team into a local firm. We still name who owns the work. If we cannot, we do not bid it.
Risks we put in the quote on purpose
- Hallucination. The model will invent a policy or a price. Production flows need a human approve step or a lookup against your database.
- Confidentiality. Staff will try the fastest box they have. Training and technical controls have to match.
- Vendor lock-in. If all your knowledge lives in one chat history, you do not own a system. Export and backups stay in scope.
- Cost drift. Usage-based APIs are cheap in a demo and noisy in production. We estimate volume before you scale.
A 30-day path that does not embarrass you
- Pick one workflow with a metric. Write it down.
- Confirm where source data lives and who is allowed to see it.
- Run a pilot on staging, with real samples, not marketing copy.
- Compare against the old process for a fixed period. Keep or kill.
- Only then connect production, monitoring and training.
That is slower than a launch video. It is how you avoid paying for a widget your team walks around.
If you already know the workflow, send it
Bring the four-line brief: job, volume, data, metric. We will say whether this is an AI pilot, a system build, a website, or training - and we will not relabel the last three as AI to win the job.
Request an AI pilot · Contact us · Book a meeting
Email: support@frontstardigital.com
Phone: +256 779 971 024
Location: Kampala, Uganda - on-site and remote.
Questions people ask
How can a small business use AI?
Pick one workflow with a measurable outcome: draft replies you still approve, classify tickets, summarise reports, or extract data from forms. Do not start with an "AI strategy" slide. Connect it to records you already trust, then decide whether to keep it.
Is ChatGPT enough for my company?
A public chat tool is useful for drafts. It is not your CRM, not your accounting file, and not a place for customer data unless you have a contract and a policy. Business AI that matters is wired into your systems with roles, logs and an owner.
How much does AI for business cost?
A true pilot (one workflow, metrics, risk notes) is a scoped project. A production integration with monitoring and training costs more because it has to survive staff turnover. Quotes that only sell "access to a model" usually hide the integration work.
Will AI replace my website or my staff?
No. AI does not index you on Google and does not replace cashiers, nurses or accountants. It can speed drafting and routing if the process is already defined. You still need a site people can find and systems that store the truth.
