AI Implementation for Small Business: the order to do it in

AI implementation for a small business works best in a fixed order: fix the process first, then pick one task that already has a number attached to it, then buy the specific tool rather than build a general one, then train the people who will touch it daily. The research on failed rollouts points at sequence and organisational design, not at the technology.

Almost every South African owner asking about AI has already been told the failure rate. It gets quoted at them in the first two minutes of a sales call, usually as a reason to buy something. The figure is real, it has a source, and the source says something more useful than the headline does. It is worth reading properly, because the actual findings are close to an instruction manual for the order to do this in.

The 95 percent figure, and what it actually counted

The number doing the rounds is that 95 percent of AI pilots fail. It comes from one document: The GenAI Divide: State of AI in Business 2025, published by Project NANDA at the MIT Media Lab, with a stated research period of January to June 2025.

The report's own sentence is narrower than the headline. It reads that "95% of organizations are getting zero return", and that "just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact".

Read that twice, because the gap between the two phrasings is the whole point. "No measurable impact" is not the same claim as "failed". A pilot with no baseline recorded before it started cannot produce a measured result afterwards, whatever it did. The report's own methodology section says success was "defined as deployment beyond pilot phase with measurable KPIs" and that ROI was measured six months after the pilot. If nobody wrote down the before, there is no after to compare it to.

The sample is worth knowing too, because it is smaller than the confidence of the headline suggests. The methodology is 52 structured interviews, 153 survey responses gathered at four industry conferences, and a review of more than 300 publicly disclosed AI initiatives. The report itself lists its limitations plainly, including that the sample "may not fully represent all enterprise segments or geographic regions" and that organisations willing to discuss their AI failures may differ systematically from those who declined.

None of that makes the finding worthless. It makes it a finding about approach rather than a verdict on the technology, which is exactly what the report says: the divide "does not seem to be driven by model quality or regulation, but seems to be determined by approach".

The contradiction inside the source, and why it matters to you

Here is something that does not appear in the write-ups, because it only shows up if you read the section rather than the summary of it.

Section 3.4 of the report covers where the money goes. Its takeaway line states that "50% of GenAI budgets go to sales and marketing, but back-office automation often yields better ROI". Four sentences later, the body of that same section states that "Sales and marketing functions captured approximately 70 percent of AI budget allocation across organizations in our survey".

Fifty in the takeaway, seventy in the body, same section, same survey. We could not reconcile the two from the published document. The underlying method is also softer than either number implies: the report says GenAI spend "is not yet formally quantified across organizations", so executives were asked to allocate a hypothetical hundred units across functions. That is a stated preference, not an audited budget.

So treat the exact percentage as unreliable and the direction as solid. Whether it is half or seven tenths, the money is going to the visible, top-line functions, and the report attributes that to easier metric attribution rather than to better returns. Its own summary line for the pattern is "Budgets favor visible, top-line functions over high-ROI back office".

That single sentence is the reason the order matters. The default instinct is to point AI at the thing customers can see. The evidence says the return is sitting in the part nobody photographs.

The order to do it in

1. Fix the process before you automate it

The report's stated barrier is not infrastructure, budget or talent. It is that most systems "do not retain feedback, adapt to context, or improve over time", and that tools fail through "brittle workflows" and "misalignment with day-to-day operations". A tool cannot align itself with a workflow that only exists in somebody's head.

Write the process down first. Who does it, what triggers it, what they type, where it goes, what breaks. If it cannot be written in ten steps, it is not one process and it is not ready. This is the same finding behind why AI projects fail in South Africa, and it is the cheapest step in the sequence because it costs an afternoon and no software.

2. Pick one task that already has a number attached

The 95 percent problem is substantially a measurement problem. So start where a number already exists, before anything changes: invoices sent per week, days to get paid, hours spent on payroll capture, quotes that never got followed up.

If you have to invent the metric to justify the project, you have picked the wrong task. Pick a different one. The measurement is not paperwork, it is the only thing that will later tell you whether to expand or stop. What a realistic first period looks like is covered in AI automation ROI in your first 90 days.

3. Look at what your team already uses without telling you

One of the more useful findings in the report has nothing to do with strategy. It says that while only 40 percent of companies had purchased an official LLM subscription, workers at over 90 percent of the companies surveyed reported regular use of personal AI tools for work.

Your staff have almost certainly already picked a tool. Ask them which one and what for, before you procure anything. It is free research into which tasks are genuinely painful, done by the people who feel the pain, and the report notes that forward-thinking organisations are "learning from shadow usage and analyzing which personal tools deliver value before procuring enterprise alternatives".

It is also a data question, not only a productivity one. If customer information is already moving through personal accounts, you have a POPIA exposure that predates any AI project you were planning. Handle that first.

4. Buy the narrow thing rather than build the general one

This is the sharpest number in the report and the one most relevant to a business without a development team. In its sample, "external partnerships with learning-capable, customized tools reached deployment ~67% of the time, compared to ~33% for internally built tools".

The report is careful here and so are we. It flags those figures as self-reported, and states directly that the difference "may reflect organizational capabilities rather than implementation approach alone" and that the correlation "does not necessarily prove causation". Organisations that choose partners may simply be better organised to begin with.

Even discounted for that, the direction is the right bet for a small South African business. Build-it-yourself is where most of the sample's effort went and where less of the deployment came out. Scope narrow and buy specific, and put the vendor through the questions in the AI vendor checklist before you sign anything. If you are still choosing between candidates, how to pick the three AI tools that get used covers the shortlist stage.

5. Train the people who will touch it every day

Last, and never first. Training before a process exists teaches people a tool they cannot apply. Training after the process is written and the tool is chosen teaches them their own job with one step removed.

Keep it to the people who touch the task daily. A general AI session for the whole company is exactly the sort of visible activity the report's investment-bias finding warns about: it looks like progress and moves no number. AI training for employees covers what that scope looks like in practice.

How this stops being a manual job

The sequence above is mostly admin, and admin is the part that quietly comes back every month unless something holds it. Four concrete mechanisms, none of which need an AI tool to work:

• The process map is captured once and reused. The ten steps you wrote in stage one become the specification you hand a vendor, the test you check their demo against, and the training script in stage five. Written once, used three times. Most of the cost of stage one is recovered before you buy anything.
• The baseline number is recorded, not remembered. One row per month in one place, captured on the day rather than reconstructed later. This is the single step that separates a project with a result from a project with an opinion, and it is why the 95 percent figure is as large as it is.
• The tool register is a standing document. Every tool in use, who owns it, what data it touches, when it renews. It answers the shadow-usage question in stage three, and it is the same register POPIA already expects you to be able to produce.
• The review date is a scheduled event, not an intention. Ninety days after go-live, the baseline is compared to the current number and the decision is expand, adjust or stop. Diarised at the moment you sign, not at the moment you remember.

Once those four exist, the work that used to be re-derived every time a new tool came up is already on file. The next AI decision costs an hour instead of a month, because the process map, the numbers and the register are all already there.

What this research does not tell you

Stated plainly, because it changes how much weight to put on it. The MIT NANDA sample is drawn from enterprises and mid-market organisations, largely American, gathered at industry conferences. There is no South African dataset of comparable scope that we could find, and a business with nine people is not a scaled-down version of one with nine thousand.

What carries across is the mechanism, not the percentages. Money follows visibility rather than return. Unmeasured work cannot be shown to have worked. Narrow bought tools outperformed broad built ones in a sample that was mostly trying to build. Those hold at any size. The specific numbers are evidence for the direction and nothing more, and anyone quoting the 95 percent at you as a verdict on AI has not read past the first line of it.

Frequently Asked Questions

What is the first step in AI implementation for a small business?

Writing the process down. Before any tool is evaluated, document who does the task, what triggers it, what gets typed and where it goes. The most commonly cited cause of stalled rollouts is misalignment with day-to-day operations, and a tool cannot align to a workflow that was never written.

Do 95 percent of AI projects really fail?

Not as stated. The MIT NANDA report found 95 percent of organisations showed no measurable profit-and-loss impact, which is a different claim from failure, since a pilot with no recorded baseline cannot be measured either way. The sample was 52 interviews and 153 survey responses.

Should a small business build its own AI tool or buy one?

Buy, in almost every case. In the MIT NANDA sample external partnerships reached deployment about 67 percent of the time against about 33 percent for internal builds. The report notes those figures are self-reported and may reflect organisational capability rather than the choice itself.

Which part of the business should AI start with?

The part nobody sees. The same research found budgets concentrating on sales and marketing while noting that back-office automation often yields better returns, and attributes that bias to easier metric attribution rather than to actual value. Start where the admin leaks, not where the work shows.

How long before an AI implementation shows a result?

Set the first review at 90 days and compare against the number you recorded before you started. Without that baseline there is nothing to compare, which is the mechanism behind most reported failures rather than the technology underperforming.

Where to start

If you already know which task is bleeding hours, start at stage one and write it down this week. If you do not, that is the actual first problem, and it is what a business autopsy is for: four hours on your real processes, working out where the time goes before anything gets bought. You can book a discovery call and we will tell you honestly whether AI is the right next step for you, or whether something duller and cheaper would move the number further.