Auto Alpha AdvisoryThe first piece of my own work I handed to AI was the one eating 1–2 hours of every day: reading the news across 80-plus crypto holdings and deciding what mattered. In March 2026 I built an automated pipeline to do that research, and when I built it, it cost about R220 a month to run. I didn't get there by asking what Claude could do. I got there by looking at which recurring task was eating my day.
That's my answer to a question I see asked in different forms: is AI actually useful for a small business, or does working out what to use it for just become another job? It tends to become another job when you start with the tool. Start instead with a costed list of the work you already do, and within a few weeks you'll know whether AI is worth anything to your business, and where. Sometimes the answer is "not yet", which is a perfectly good result. Wandering around inside a subscription looking for a purpose is the outcome to avoid.
Why AI turns into a second job
Most firms meet AI tool-first. Someone buys seats, or the owner signs up after a demo, and the team is told to find uses for it. The hunt for uses, and the checking of whatever comes back, lands on top of the work that was already there.
The survey data looks a lot like that. Upwork's July 2024 press release reported on workers in the US, UK, Australia and Canada, surveyed in April and May 2024: 77% of employees using AI said the tools had added to their workload, and 47% said they had no idea how to achieve the productivity gains their employers expected. The survey didn't compare tool-first with task-first rollouts, so this next part is my reading rather than its finding. Tell people to "use AI" without saying on what, and numbers like those are what you'd expect.
You wouldn't run a cost-saving programme that way, buying the machine first and then walking the floor for something to point it at. You'd find the cost first.
Cost your week before you open a tool
I'm a CA(SA), so I treat this as cost allocation. It's activity-based costing, applied to one week of your own business.
Pick a normal week: not month-end, and not the week half the team is on leave. For five days, everyone (you included) notes each recurring task that takes more than fifteen minutes. Write down what it is, how long it took, how often it comes round and where its inputs came from. You want the shape of the week, not a timesheet audit.
Then cost each line at hours × loaded hourly rate. Loaded means salary plus everything else it costs to employ that person. I've worked through the loaded-rate calculation for South African payroll in Automate or Hire, so I won't repeat it here. Cost your own hours too. In an owner-run firm the owner's time is usually the most expensive in the building and the least often priced.
By the end, every line should carry two numbers: hours a year, and what those hours cost in rand.
Rank by how often it repeats and how rule-bound it is
Cost tells you what a task is worth fixing. It doesn't tell you whether AI can fix it, so score each line on two more things.
Start with frequency. Daily scores 3, weekly 2, monthly or less 1. A daily task gives you hundreds of chances a year to test and correct a setup, where a quarterly one gives you four.
Then score how rule-bound it is. If you could write the steps on one page for a new hire, that's a 3. Mostly rules with the odd judgment call is a 2. If it depends on reading people or on experience nobody has written down, it's a 1.
Multiply the two. Anything scoring 6 or more makes the shortlist; sort the shortlist by annual cost and take the top line. One check overrides the scores. If the inputs arrive on paper or only exist in someone's head, the task isn't ready for AI yet. Getting the input into digital form comes first, and that's often the smaller fix.
Here's how it might look for a hypothetical eight-person accounting practice in Claremont:
- Answering repeat client questions about filing deadlines: 5 hours a week, frequency 3, rules 3, so it scores 9.
- Chasing outstanding client documents by email: 3 hours a week, frequency 2, rules 3, so it scores 6.
- Drafting engagement letters: 2 hours a week, frequency 2, rules 2, so it scores 4.
- Partner review of files: 6 hours a week, frequency 2, rules 1, so it scores 2.
Partner review takes the most hours and scores the lowest. That's normal. In a professional firm the expensive work is usually the judgment, which is why it's expensive. The first line wins: five hours a week is around 240 hours a year, and nearly all of it follows rules.
Starting with one task isn't timid, either. The US Census Bureau's April 2026 working paper on how firms use AI found that 65% of firms using it in worker tasks kept it to three or fewer.
How I chose mine
I didn't need a spreadsheet to find my first task, but it holds up on the method above. Portfolio research ran daily, so it scores 3 on frequency. Most of it followed rules I could write down: check these sources for these assets, flag anything material, ignore the rest. Deciding what counts as material takes some judgment, so call it a 2. That's a 6, which makes the shortlist, and at 1–2 hours a day it was costly enough to rank near the top. Every input was already digital, too: X, news sites, Telegram groups and price data.
So I built one pipeline around that one task. It uses the Claude API for relevance scoring and synthesis, Tavily for news search, FireCrawl for X, CoinGecko for prices and AgentMail to send the briefing by email. The build log has the detail.
On working days alone, the 1–2 hours a day it took when I built it come to roughly 20 to 40 hours a month. Put any hourly rate you like on those hours and set the total against R220. You won't need a spreadsheet for that comparison either. The build time was a real cost and belongs in the same sum, but most of it was paid once, while the saving comes back every day.
What I didn't do was open Claude and ask what it could do for me. I picked the task, and the task picked the tools.
Sometimes the honest answer is "not much yet"
A recurring argument on small-business forums is that small businesses don't care about AI automation. I take it seriously, because the best data I've found on the question partly agrees.
The US Census Bureau has asked businesses about AI in its Business Trends and Outlook Survey since 2023, when only 3.8% of US businesses said they used it to produce goods and services. The Minneapolis Fed's May 2026 summary of that data is blunt: "Among those not planning to use it, the most common reason cited was that AI did not apply to their operations." The Census Bureau's own update the same month, covering December 2025 to May 2026, found 37% of firms with at least 250 employees using AI, against fewer than one in five firms with four or fewer. Among firms with fewer than 20 employees, use didn't change significantly over the period.
These are American figures, and I haven't seen a South African dataset this detailed, so I won't guess at the local split.
Some of those owners are right. If your costed week turns up no task that repeats at least weekly, follows rules you could write down and adds up to more than a few hours, AI won't change much for you yet. A five-person firm whose work is site visits and client relationships could log a week and find nothing worth handing over. That's a legitimate result, and it cost a week of notes rather than a year of subscriptions.
I'd rather say that plainly, because it's what makes the rest of this believable. If an adviser tells you every business has a dozen AI use cases waiting, ask to see their log.
Keep the log, though. Run it again when headcount grows, or when clients start sending you structured data your team used to retype.
What to do once the log picks a task
Before you buy anything, test the top task for two weeks with a chat tool you already have, on real inputs, and time it against the old way. If the saving shows up, it's worth building properly. If it doesn't, try the second line or stop there.
Keep it to one task until that one works. The second job tends to appear when you go looking for uses everywhere at once.
If you'd rather do the mapping with someone, my automation work starts with the Map, a free 90-minute session that maps your workflows and ends in a plain go or no-go. The AI automations page sets out how it works.