Start with one painful task: how a small business can actually put AI to work
Updated 15 Nov 2026, 16:50 SAST
The short version: Most organisations that say they “use AI” have not yet made it part of everyday work. McKinsey’s 2025 global survey found that 88 percent of respondents reported regular AI use in at least one function, yet nearly two-thirds had not begun scaling it across the enterprise, and only 39 percent reported any enterprise-level EBIT impact. Smaller firms lag: 29 percent of companies with under $100 million in revenue were in the scaling phase, against nearly half of those above $5 billion.[1] For a shop with a handful of people, the useful question is not “How do we adopt AI?” It is “Which painful task will we fix first, how will we measure it, and who checks the output before anything goes out?” The plan below draws on published surveys and, where labelled, on our own builds for small, often one-person, businesses — many of them on the road in South Africa, where WhatsApp is already the front desk.
Why most shops get stuck
McKinsey’s survey, fielded from 25 June to 29 July 2025 among 1,993 participants in 105 countries, found that use is widespread and scale is not. Approximately one-third of respondents’ companies had begun to scale their AI programmes. High performers, McKinsey wrote, are more likely to redesign workflows, not just add a chatbot on the side.[1]
In July 2024, Gartner predicted that at least 30 percent of generative-AI projects would be abandoned after proof of concept by the end of 2025, “due to poor data quality, inadequate risk controls, escalating costs or unclear business value.”[2] That was a forecast, not a later measured rate. The reasons it named — fuzzy value, rising cost, weak data — are the same reasons a small pilot quietly dies.
The OECD, in its 2026 report on AI and skills, said a lack of skills was holding back adoption, “especially” among small and medium-sized firms. Around 40 percent of employers in manufacturing and finance who had not yet adopted AI named skills as the main reason, “as do more than half of SMEs that are not yet using generative AI.”[3] In the OECD’s more detailed SME survey, half of non-users (49.8 percent) said employees did not have the right skills.[4]
We have already reported on this training gap in The AI training gap: what businesses say is hardest right now, and how to teach the shop floor. This article is the next step: what to implement, and in what order.
Analysis. The pattern in these surveys matches what we see when a small firm tries to “get on AI.” The owner signs up for a chatbot, staff try it on a few emails, nobody writes down whether it saved time, and three months later the subscription is still there and the painful job is unchanged. The failure is usually in the choice of starting point, not in the model.
Start with one painful task
In training work we have done with shop-floor teams, the first lesson is always the same: pick one problem, not a list. Write down what went wrong this week, in plain words. A tool can help you group the list and ask how often each item happens, what it costs in time, and who it affects. It cannot tell you which problem matters most. That call belongs to the owner and the people who do the work.
A good first task has three tests:
- You can say it in one sentence: what is happening, where, how often, and what it costs. No causes, no blame. If you cannot say it in one sentence, it is still too big. Cut it down.
- It is small enough to finish in a month, or at least to have a working version in 30 days.
- You can measure it before you change it. A tally on a notepad is enough: how many enquiries went unanswered after hours; how many invoices went out late; how many hours the owner spent chasing payment.
From our own build. The jobs that survive this test tend to be unglamorous: a WhatsApp that sits unanswered while the owner is on a job; a quote that never becomes an invoice; a slip photographed and then lost in a camera roll; a reminder that nobody sent. Those are better starting points than “install AI across the business.”
Do not change the process while you are still measuring the baseline. The 30/60/90-day plan below starts with that count.
The jobs AI actually handles well
Analysis. AI is uneven. It is strong at drafting, sorting, transcribing and reminding. It is weak at judgement, at prices, at promising a slot, and at anything you would be ashamed to send if it were wrong. In our own builds we treat four jobs as the core, and we keep a person in the loop on every outbound message.
Front-desk enquiries and WhatsApp
For many small South African firms, the front desk is already a phone. A Standard Bank study of informal and township businesses, compiled with the research firm Foshizi from interviews conducted between March and May 2025, and reported by the Sunday Times, found WhatsApp was the most popular marketing and communication tool, used by 74 percent of respondents. Nokhanyo Ntshingila, a Standard Bank senior manager for enterprise and supplier development, said firms mainly based the choice on cost: “WhatsApp is the go-to tool for its ease and low cost.”[5]
From our own build. A workable front door looks like this. A message arrives — a form, an email, or a WhatsApp chat. It is logged, then classified as a quote request, a booking, a simple question, or something that needs the owner. The system can draft an acknowledgement and a reply. It should not invent a price, confirm a time, or send anything to the customer until the owner has approved it. Frequently asked answers can go out automatically only when the owner has already written them and has said they may. If the owner already lives in WhatsApp, the approval should appear there, or in a single inbox they already check.
Invoices and paperwork
Capture on the phone, draft in the system, confirm before it leaves. A voice note at the end of a job — the customer’s name, what was done, a rate — can be transcribed and turned into a draft quote or invoice. A photo of a slip can be filed and turned into a draft expense record. Overdue invoices can be chased on a fixed rhythm, for example at 7, 30 and 60 days, by email rather than from the owner’s private number. Incoming payment notices from the bank can be matched to open invoices, with a short unmatched list sent back to the owner.
None of that should go out raw. Transcription mishears names and amounts; a document scan misreads a total. From our own build. We leave every new document in draft until the owner confirms it. Payment chases are held for a “send it” or “don’t send.” The bookkeeper, or the owner’s tax practitioner, still files the returns. The aim is fewer lost slips and fewer late invoices, not a robot that runs the books.
Scheduling and ops check-ins
The useful product here is not a clever agent. It is a short daily list.
From our own build. A morning digest that the owner can read in a minute beats a dashboard they never open. Ours are written as a handful of actions, with money at risk and today’s commitments first: jobs on the diary, an overdue invoice, a reminder that still needs approval, a booking that did not show. If a section has no data, the digest should say so. It should not guess a stock level or invent a total. Anything that would message a customer is a suggestion, not a send. Start with the digest, then add one chase when the digest is trusted.
Staff questions
Even a one-person business eventually has a helper, and the questions are predictable: can I take Friday off, yesterday’s hours were nine not seven, a certificate or a vehicle disc is about to expire.
From our own build. A staff inbox can log the request, draft a holding reply, and put a one-line summary in front of the owner. It should never approve leave, change pay, or confirm a rota on its own. If the record is missing a leave balance or an expiry date, it should say so and ask, not guess. Disputes about pay, discipline or injury are not chatbot work; they go straight to the owner.
Choosing tools versus building
Analysis. Most small firms do not need to build a model, and most do not need a custom app on day one. The useful split is: use what you already have; buy a ready tool for a standard job; build only the glue between them.
Use what you already have. If the team already writes on WhatsApp and email, start there. A general chatbot is enough to help draft a quote, a job note, or a polite payment reminder, provided someone checks the text.
Buy a ready tool when the job is standard. Invoicing, document filing, calendars and basic transcription are solved problems. Paying for a known product is usually faster than writing your own, as long as the owner is not forced to learn a new desktop for each one.
Build, or wire, a custom flow when the work lives between tools. The painful job is rarely “create an invoice.” It is “I sent a voice note from the van, I need a draft invoice, I want to approve it on my phone, and I want it emailed to the customer without opening three other systems.” That glue is where low-code workflow tools — the n8n or Zapier class — earn their keep. From our own build. We reuse four shapes: intake (something arrives, it is logged, the next action is decided), chase (a nudge on a rule), digest (a roll-up for a person), and a gate (a hard stop before anything outbound). The automation runs the chase. A person keeps the conversation.
If a ready-made tool already does most of the job, buy it and live with the gap for 90 days. Only build the gap if it is still painful after you have measured it. The owner should have one login, or no login at all beyond WhatsApp.
Getting the staff on board
The related article on the AI training gap sets out why general “intro to ChatGPT” sessions fade, and why the training that matters is shared practice on real jobs.
The owner has to use the tool first, on the same job the staff will use. If only the office sees the drafts, the floor will not trust them. Appoint one checker on the floor and one in the office, or, in a one-person business, make the owner the checker until the error rate is known.
Write a short rule-book, one or two pages: what may be pasted into a tool, who checks the output, what is recorded, and what happens when the tool is wrong. Practice during paid hours, on last week’s real jobs, with private details removed. Check the arithmetic; transcription and document tools misread handwriting and can add up a column wrongly. Do not promise that AI will replace the team. That is a fast way to lose the people you need to check the work.
Costs, and why simpler usually wins
We are not going to put a price on a build here. Costs move, and a figure that is right for one firm is wrong for the next. What the surveys do show is that unclear value, not a lack of software, is what kills a pilot.
From our own build. For a small shop, the expensive part is usually the time spent setting the work up and supporting it, not the monthly fee on a chatbot or a workflow tool. Complexity is a cost. Every extra login, dashboard and automatic send is something the owner has to watch. Keep the first version boring: one task, one channel, one person who can switch it off. Prefer a tool the team already opens, and a draft that waits over a message that leaves on its own. Add a second job only when you can point to a number that moved — unanswered chats, days to invoice, hours spent chasing. If a vendor leads with a guaranteed saving, or cannot tell you how a person stops a bad send, walk away.
Mistakes we keep seeing
From our own build, and from the surveys above, these are the patterns that waste the first three months.
Starting with everything. A front door, a voice-to-invoice flow, a staff inbox and a daily digest, all switched on in the same week, means nobody knows what failed.
Skipping the baseline. If you did not count how late the invoices were, you cannot say the new flow is better.
Letting the bot send. The moment a system can invent a price, confirm a slot, or approve leave, it will eventually do so. Draft and hold is slower. It is also how you keep customers.
Buying a desk the owner never opens. A powerful back office is useless if the owner is in a van and the only screen they trust is WhatsApp.
Starting with “agents.” McKinsey’s 2025 survey found plenty of curiosity about AI agents, and much less actual scale: 62 percent of respondents said their organisations were at least experimenting with agents, but in any given business function, no more than 10 percent said they were scaling them.[1] A small shop should first make a simple intake or chase reliable.
No written stopping rule. If nobody can say when a draft must go to a person, the system will eventually continue silently. In our own work, a failed check is retried once; a second failure goes to a human and the flow stops.
Overlapping tools for the same job. Two reminder systems, or two places where a lead can sit, guarantee that something will be chased twice, or not at all.
Treating a chatbot session as the plan. A good answer in a chat window is not a process. The process is what happens the next time the same job arrives, including when the owner is busy.
A 30/60/90-day plan
Adjust the pace if you are a one-person business doing the work in the evenings.
Days 1 to 30 — pick, measure, draft. Write the one-sentence problem. Count the baseline for at least a week, without changing the process. Choose the smallest tool that can draft a useful output — a chatbot for text, or a simple intake that logs WhatsApp or email. The owner checks every draft. Nothing is sent automatically. Write the one-page rule. If you have staff, name the checker and practise on real jobs during paid hours, as set out in Getting the staff on board.
Days 31 to 60 — wire one chase or one digest. Once drafts are trusted, add a single follow-up: a payment reminder held for approval, a booking reminder, or a morning list of what needs action. Keep the human gate. Review the error rate once a week — wrong names, wrong amounts, messages that should not have been drafted — and fix the rule-book. Do not add a second job yet.
Days 61 to 90 — prove it, then add one more. Compare the new count with the baseline. If the number has not moved, do not scale; change the task or the tool. If it has moved, add one adjacent job only — slip photos once invoices are reliable, or a staff leave inbox once the morning digest is part of the day. Decide whether you still only need ready-made tools, or whether a small piece of glue is worth building, as discussed in Choosing tools versus building. Rehearse what happens when the tool is wrong, and how you switch it off.
After 90 days you should have one working process, one measured improvement, and a short written rule. That is a better foundation than a stack of unused subscriptions.
The South African angle
The international surveys are useful, but they are not a picture of a plumber in Gqeberha or a two-person workshop in Germiston.
SAP Africa, reporting in May 2025 on research done at the end of 2024 among mid-size and enterprise companies in Kenya, Nigeria and South Africa, said every organisation surveyed expected demand for AI skills to rise in 2025, and expected some AI-related skills gap. Organisations in South Africa were “most likely to experience AI skills shortages (53%), followed by Nigeria (50%) and Kenya (43%).” Where skills were simply not available, 68 percent of South African respondents were using automation to augment their AI skills, SAP said, the highest of the three countries.[6]
PwC South Africa’s study of AI in manufacturing operations, based on a survey of more than 400 operations executives across Europe, the Middle East and Africa, found high local optimism and thinner results. The published report said 81 percent of South African respondents expected AI to increase operating profits by at least three percentage points by 2030, and that “only 4% globally and 15% from South Africa report that they have already achieved significant financial benefit and return on investment (ROI) from AI in operations.” A later summary on PwC’s South Africa page used a 13 percent local figure for the same finding.[7][8] We have used the 15 percent figure from the published report, and we flag the difference rather than averaging the two.
On the ground, the Standard Bank township study is the more useful clue for a small firm: WhatsApp first, websites later, cost as the deciding factor.[5]
From our own build. That is why we design for the phone, not the desk. The owner is often on the road, and the enquiry arrives as a voice note or a WhatsApp message. Bank payment notifications arrive by email. Power and connectivity are uneven outside the main hubs, so a process that dies when a laptop cannot open a dashboard is a poor process. Afrikaans and English both show up on slips; a document tool has to be able to read more than one language. The owner should not be asked to open a stack of software accounts just to accept a job.
Why it matters for your business
The surveys are easy to read as someone else’s problem — big-company pilots, C-suite skills shortages, manufacturing returns. The useful reading is narrower. Most firms that “use AI” have not yet made it pay. Skills are a real barrier, and most small firms that are not using these tools also do not train for them. The shops that get through the first 90 days are the ones that pick a single job, measure it, keep a person in the loop, and refuse to let a draft become a send on its own.
The rest of the stack can wait. The unanswered WhatsApp, the late invoice and the morning list cannot.
This article is for information only and is not financial or investment advice.
Spotted an error? Tell us and we'll correct it.
Sources
- McKinsey, The State of AI: Global Survey 2025 ↩
- Gartner, 29 July 2024: “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025” ↩
- OECD, 2026: “AI and skills” ↩
- OECD: “Are SMEs prepared for generative AI? Generative AI and the SME Workforce” ↩
- Sunday Times, 19 October 2025: “WhatsApp tops for township SMEs,” reporting the Standard Bank Township Informal Economy Report 2025 (research by Foshizi) ↩
- SAP Africa, 20 May 2025: “Does Africa Have the Skills to Benefit from AI?” (report: Africa’s AI Skills Readiness Revealed) ↩
- PwC South Africa: “AI in operations: Revolutionising the manufacturing industry” (PDF) ↩
- PwC South Africa page for the same study, which states a 13 percent local figure for significant financial benefits and ROI ↩
Featured deep dives
Longer follow-ups on the stories behind the week's headlines, with every source linked. Free to read.