Auto Alpha AdvisoryTwo of my own builds make the point better than any survey. One reads several dozen news and research feeds every morning, filters out the noise and leaves a summary waiting before I open my laptop. It has run daily and unattended for months. The other was a multi-agent build: four bots, a shared knowledge base, voice interaction. It demoed beautifully. I now start it by hand, when I remember, which is close to never.
Same builder, same tooling, opposite fates. The difference had nothing to do with sophistication. The boring one is attached to a clock. The interesting one is attached to my memory, and memory is not a scheduler.
That is the rule I would defend against most of what gets sold as business automation: an automation survives when it attaches to something that already happens on a schedule, and it dies when it needs a human to remember it. As a CA(SA) I am in the habit of asking what a thing costs to keep rather than what it costs to build, and on automation that question sorts the field on the first pass.
I have written already about which automations to set up first and about what you are actually buying from an automation company. This is the part that never makes it into the proposal: what those automations look like six months after the invoice is paid.
Owners are not hostile to automation, they are indifferent to it
The businesses I speak to have not rejected automation. They have never had a reason to think about it as a category at all.
Survey headlines make this hard to see. U.S. Chamber of Commerce research published in August 2025 put small-business use of generative AI at 58%, up from 40% a year earlier and 23% in 2023. But that number counts anyone who has typed into a chat window. Typing into ChatGPT is not an automation. It still requires you to remember to open it, which is the exact failure mode this whole post is about. On the definition that matters — something that runs without you — most of the South African owners I meet are at zero, and they are not upset about it.
What actually predicts survival
An automation survives when its trigger comes from outside the person who benefits from it. Everything else is detail.
A scheduled job has a trigger nobody has to supply. It fires at 6am on Tuesday whether anyone thought about it or not. An assistant you have to open has a trigger made of intention, and intention is the least reliable component in the system. It degrades under deadline pressure, which is precisely when the automation was supposed to earn its keep.
This is also why the failure numbers were bad long before language models arrived. Writing in 2020, Pegasystems founder Alan Trefler cited an EY study finding that between 30% and 50% of initial robotic process automation projects failed outright — the previous generation of automation, with no AI in it anywhere. Trefler's diagnosis: "RPA can't fix bad processes – it just speeds them up." The mechanism has not changed. If the underlying task was already skipped half the time, automating it produces a script that gets skipped half the time.
Which automations are structurally durable
The durable ones share one property: their cadence is set by someone other than you.
Scheduled filtering and reporting. A morning digest, a weekly exception report, a nightly data sync. The calendar supplies the trigger. The failure mode here is silent death: the job stops, and because nobody was waiting at a specific minute, weeks pass before anyone notices. The fix is cheap and almost always skipped — the job must send something even when there is nothing to send. An empty digest is a heartbeat.
Statutory and deadline-driven work. VAT201 preparation, payroll runs, EMP201 submissions, annual returns. SARS sets the date, so the trigger is enforced by an outside party with penalties behind it. These are the most durable automations a small business will ever own. Their failure mode is nastier than stopping: a rate or form change upstream that the script does not know about, so it keeps running and produces confidently wrong output on time.
Event handoffs between two systems already in daily use. Enquiry form to CRM record to notification. The trigger is a customer, not an employee. Failure mode: someone renames a field on either side and the mapping breaks. Trivial to fix if it fails loudly, fatal if it fails quietly.
Which ones are structurally fragile
Every fragile shape has the same tell: a person has to decide to use it.
The assistant you invoke. The prompt library, the internal tool with a nice interface, the bot you open when you need it. Failure mode: you forget it exists. Inside a month the browser tab is faster than remembering the workflow.
Chat over internal company knowledge. Failure mode: staff ask a colleague instead. Asking a person is socially cheaper and the answer arrives with accountability attached, which no chatbot offers.
Multi-step agent chains where step three consumes step two's free text. Failure mode: compounding error with no schedule to catch it. Recovery needs one human who understands the entire chain, and there is only ever one of those. When they are on leave, it is off.
Anything whose output needs editorial judgment before use. Failure mode: reviewing takes as long as doing. The queue fills, then the queue stops being opened. I have built this one for myself more than once and I would not claim immunity.
Why the demos that sell best die fastest
A good demo optimises for the ten minutes you are watching. Survival is decided by the two hundred mornings afterwards, and nothing about a demo tests that.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, on escalating costs, unclear business value or inadequate risk controls. Anushree Verma, Senior Director Analyst at Gartner, put it directly: "Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied." In the same January 2025 poll of 3,412 attendees, 19% said their organisation had made significant investments in agentic AI. Gartner also estimates that of the thousands of vendors marketing agentic products, only about 130 are real; the rest is what it calls agent washing — existing chatbots and RPA with new labels.
What "automation company" means in South Africa
Search for automation companies in South Africa and most of the first page is industrial, not office. Systems Automation & Management, head office in Gauteng, describes itself as "one of the leading integrators of PLC's, SCADA's and Fieldbus systems in South Africa", working in power generation, mining, automotive and food and beverage. PCMP in Roodepoort sells turnkey industrial automation — PLC, SCADA/HMI, MES, industrial networking — to manufacturers.
That is a different product from what an office-bound owner means by the word. It is also, quietly, the grown-up version of the argument above. A PLC on a bottling line is the purest scheduled automation there is: the trigger is the line itself, the failure is loud and immediate, and nobody has to remember to run it. When an office automation is still alive at six months, it has usually become more like SCADA than like an agent.
How to tell within a month, and the cheapest test
Count runs, not outputs.
Instrument the thing on day one so every execution writes a line: timestamp, trigger, result, and whether a human touched it. At 30 days you want two numbers — how many times it fired on its own schedule, and how many times a person had to start it, unblock it or correct it. If human touches are running above roughly one a month, it is already dying. You have bought a chore with a dashboard attached.
Three other tells at the one-month mark. Does anyone besides the builder notice when it is late? Does it fail loudly, or does it fail by producing nothing? Is its output already an input to a meeting or deadline that exists anyway?
And the cheapest test of all, which costs nothing and belongs before you spend anything: put the task on a calendar reminder and do it by hand for two weeks. If you skip it manually, you will not maintain it automated. You will have moved the skipping into a log file nobody reads. If you do it fourteen times out of fourteen and resent every one, you have found something worth automating, and you now know its schedule, its inputs and its outputs, which is most of the specification.
The uncomfortable conclusion is that the automation worth paying for is usually the one too boring to demo. When a supplier shows you a live agent doing something impressive on a call, ask what it will be doing on a random Tuesday in March, and who will notice if it stops. If the answer is vague, you are being sold the ten minutes. Mapping which of your processes actually have a schedule behind them is most of what an AI readiness audit does, and it is worth doing before anyone writes code.