Do you need AI, or better software?
A lot of “we need AI” turns out to be a workflow that never got fixed, data scattered across six places, or a product nobody has defined yet. Here is how to tell the difference before you spend anything.
Vinerals TechnologiesWorkshop notes
Right now a lot of owners are getting the same message from a board member, a competitor’s LinkedIn post, or a nephew who is very online: you need AI. The pressure is real, and it is not stupid. Some of it will matter. The trouble is that “we need AI” is a wish with a budget attached, arriving before anyone has said what the AI is actually for.
We have sat through enough of these conversations to notice a pattern. A good share of the time, the thing being described as an AI problem is a workflow that never got fixed, data that lives in six places, or a product nobody has defined clearly. AI does not fix those. It usually just runs them faster, and charges you more for the privilege.
So this is a piece about slowing down for one afternoon. Where AI genuinely helps a small or mid-sized company today, where it quietly does not, and how to spend a first quarter finding out without lighting money on fire.
The gap between the hype and what actually helps
The demos are built to look like magic. A system that reads any document, answers any question, runs the business while you sleep. In a controlled demo, with clean inputs and a friendly question, it works. In your office, with a supplier PDF that someone scanned crooked in 2019, it is a different afternoon.
Here is the frame we find useful. Broad, do-everything AI is still unreliable for anything you would stake real money on. Narrow AI, pointed at one well-defined task, is quietly good and getting better every few months. The wins we actually see at SMEs are almost always narrow.
Pulling structured fields out of messy documents.
Invoices, receipts, intake forms. A person still checks the uncertain ones.
Drafting first versions.
Replies, summaries, product descriptions. Someone edits before anything ships.
Sorting and routing.
Triaging support tickets, tagging leads, flagging the one email that needs a human today.
Search over your own material.
Asking a question across your contracts, manuals, or past projects and getting pointed to the right page.
None of that replaces a department. All of it saves a few hours a week on work people already resent doing. That is a genuine return, and it is the return worth chasing first.
When “we need AI” is really something else
A fair chunk of AI requests are another problem in costume. Nobody is being dishonest. AI is the word in the air, so every ache gets that name. A few we hear often.
“We want AI to answer customer questions.”
Usually the answers live in one person’s head and a Slack channel from 2022. Write them down first. A search tool over a good knowledge base beats a chatbot over a bad one.
“We need AI to predict demand.”
Sometimes you do. Often you have never plotted last year’s sales in a single view. A clear dashboard answers the question you actually have this week, for far less.
“AI should automate our intake.”
Frequently the intake is three forms, two inboxes, and a spreadsheet that four people edit at once. Get that into one path first. Then decide whether a model needs to sit anywhere in it.
“We want AI to understand our data.”
That one is usually a data problem wearing an AI hat. If your records are scattered and inconsistent, no model saves you. It just gives you confident answers that happen to be wrong.
The tell is simple. If a mid-level employee with a clear process and a decent tool could do the job, you have a software or workflow problem. Reach for AI when the task genuinely needs judgment over language or images, at a volume a person cannot keep up with.
The first projects that tend to pay back
When AI does earn its place at an SME, the first project tends to share a shape. Small, measurable, low stakes if it goes wrong, and sitting on top of data you already have.
Document extraction with a human check.
Invoices, purchase orders, forms. High volume, repetitive format, a person reviewing the ones the system is unsure about. You can measure the hours saved inside a month.
Drafting inside a tool people already use.
First-draft replies in the support inbox, first-draft summaries of long threads. The person stays in the loop, so a bad draft costs a delete, not a customer.
Internal search over your own documents.
Staff stop asking the same colleague the same question. Low stakes, because the tool points to a source they can open and verify.
And a few that usually do not pay yet.
Customer-facing automation with no human in the loop, early on.
The failure is public, and it is your brand wearing it.
Predictions that move real money.
Dispatch, pricing, inventory. Not before you have clean history and someone who can sanity-check the output.
A big internal platform with AI wired through every screen as version one.
You will spend a year learning what a two-week pilot would have told you.
A pattern worth stealing: keep a person in the loop for the first version of anything. It is cheaper, it is safer, and it teaches you where the model actually breaks before you trust it on its own.
The data question, asked before you spend
Most AI projects that stall do not stall on the model. They stall on the data underneath. The models are close to a commodity now. Your data is the part that is yours, and the part that decides whether any of this works.
Before you start, you want honest answers to a few plain questions.
Where does it live?
If the answer is “spread across email, a shared drive, and three SaaS tools,” then collecting it is the first project, and it is not a small one.
Is it consistent?
The same customer spelled four ways, dates in three formats, fields people fill in when they feel like it. A model grounded on that will be confidently inconsistent.
Do you have enough of it?
For pulling fields out of documents, you need a real pile of examples, including the ugly ones. For anything predictive, you usually need more than a year of genuine history.
Are you allowed to use it?
Customer data under Québec’s Law 25, anything sensitive, anything a contract restricts. Decide where the data can go, and whether it can leave your walls, before you hand it to anyone’s model.
You do not need perfect data. You need to know the shape of what you have. Half of what looks like an AI project turns out to be a month of quietly cleaning up records, and that month is often worth more than the model that sits on top of it.
A calm plan for the first quarter
If you want to start without betting the year, here is a shape that has worked. Thirteen weeks, roughly, in three moves. Bend the weeks to your reality.
Weeks 1 to 3: pick one task and measure it.
Choose a single annoying, repetitive job. Write down how long it takes today and how often it goes wrong. Without that baseline, you will never know whether the project helped.
Weeks 4 to 6: get the data honest.
Pull the relevant records into one place. Clean the obvious mess. Settle the rules and the privacy constraints. This is the unglamorous part, and it is where most of the value hides.
Weeks 7 to 11: build a small pilot with a person in the loop.
One task, one team, a human checking the output. Keep it off the critical path so a bad day is a nuisance instead of an incident.
Weeks 12 to 13: compare against the baseline and decide.
Did it save the time you hoped? Where did it fail? Now you can expand it, narrow it, or stop, with evidence instead of a feeling.
A quarter is enough to learn whether AI helps this specific corner of your business. It is short enough that being wrong stays cheap. And it usually surfaces the workflow and data fixes that were the real work all along.
If the honest answer at the end of the quarter is “the workflow was the problem, and now it is fixed,” that is a win, even when no model ships. Plenty of our AI conversations end right there, with a company that spent a little to avoid spending a lot on the wrong thing.
We build software by hand for SMEs, and part of that work is saying so when the thing you need is better software, cleaner data, or a clearer process before any AI enters the picture. If you are under pressure to do AI and you would like a plain read on which parts are real for you, we are glad to sit down and work through it. Bring the task that annoys you most. We will start there.
