An AI workflow is a fixed sequence of steps, written by you, with a model doing the part that involves reading, writing or classifying language. Something arrives, the model interprets it, your system does something with the result, and a person checks the bits that need checking. That is it. No autonomy, no decisions the model makes on its own, no surprises in the bill. If you want the contrast, that last part is exactly what separates a workflow from an AI agent.
AI workflow automation is unglamorous by design, and it is where almost all of the value we have seen a mid-sized business get from AI actually comes from. The hard part is not building one. The hard part, and the reason so much money gets spent for nothing, is picking the right job to point it at.
The News24 and Lula Small Business Survey of 1,088 South African SME owners in January 2026 found that more than two-thirds are either using AI daily or actively experimenting with it. That survey measures adoption and experimentation, not production: in our own client work, very few of those businesses have an AI workflow running unattended, on real data, every day, without someone babysitting it. That is the gap this post is about.
What an AI workflow actually looks like
Concretely, because the abstraction is useless.
A supplier invoice lands in an inbox. The workflow reads the PDF, pulls out the supplier, the invoice number, the line items and the total, matches it against the open purchase order, checks the arithmetic, and either files it as clean or puts it in an exceptions queue with a one-line note explaining what did not reconcile. A human looks at the exceptions queue. Nobody looks at the clean ones.
Note the shape. The model does one thing: turn a messy document into structured fields. Everything else is ordinary software doing ordinary software work. The person is placed at the only point where judgement is needed, and they see fifteen invoices a week instead of two hundred.
That is the pattern. The model is a component, not the system. Which is also why the whole thing can be tested, replayed and explained, and why the truth about what you owe your supplier still lives in your accounting system rather than in the model's output. We have written at length about why your AI should never be the source of truth, and workflow design is where that principle either gets respected or quietly broken.
Five tests for whether a job is worth automating
We run this filter with clients before anyone scopes anything. A job needs to pass most of it, and the first two are close to mandatory.
1. It happens a lot. Daily beats weekly, weekly beats monthly. This sounds obvious and gets ignored constantly, because the jobs people complain about loudest are often the annual ones. Volume is what turns a small per-item saving into a number that appears in your management accounts. A task that runs twice a month will not repay the effort of automating it, however irritating it is.
2. It is words-shaped. The model has to be doing something it is genuinely good at: reading unstructured text and turning it into something structured, or turning something structured into readable text. Extracting fields from documents, classifying incoming messages, summarising a call into notes, drafting a first version of a standard reply. If the difficult part of the job is arithmetic, scheduling, or looking something up in a database, that is a software problem with a well-understood, cheaper, more reliable solution, and you do not need AI for it.
3. A competent new starter could do it from written instructions. This is the sharpest test in the list. If you could hand the job to someone in their second week with a page of instructions and expect a good result, a workflow will do it. If the job depends on knowing that this particular customer always writes their order numbers backwards, and that Thursday deliveries go to the other depot, and that Johan's quotes are done differently, then the instructions do not exist yet. Write them first. Sometimes writing them is the whole project, and you discover you did not need AI at all.
4. Being wrong is cheap and visible. Automate the jobs where a mistake surfaces immediately and costs little to fix. Mis-classify an inbound enquiry and someone reroutes it in ten seconds. Get a customer's credit limit wrong and you have a very different sort of afternoon. Start where the blast radius is small, build the trust, then move inward. This is also the first place your POPIA obligations bite: Section 71 restricts decisions taken solely by automated processing where they have a legal or similarly significant effect on a person, and where such a decision is permitted the person must be able to make representations and be given sufficient information about the logic behind it. The simplest way to stay clear of that is to keep a person in the loop on those decisions.
5. The data is already somewhere you can reach. If the input arrives as a photograph of a delivery note sent over WhatsApp, or lives only in a spreadsheet on one laptop, the integration is the project and the AI is a footnote. That does not make it a bad project. It makes it a different, larger one, and you should know that before you sign for it rather than three weeks in.
The candidates that look good and are not
Three shapes come up in nearly every scoping conversation, and all three tend to disappoint.
The monthly report. High visible effort, low frequency, and usually a job where the difficulty is deciding what the numbers mean rather than assembling them. Fails test one, often fails test two. If the report is genuinely painful, the underlying problem is usually that everyone is looking at different numbers, which is a data problem wearing a reporting costume.
The job where the rules are contested. Two people in the business would do it differently and both think they are right. An AI workflow will not resolve that disagreement; it will encode whichever version was in the room when the prompt was written, and then everyone will distrust the output. Settle the rule first.
The customer-facing chatbot. Still the most requested first project and still, in our experience, the worst one. It is the most visible thing AI can do, which is exactly why it attracts the budget, and it puts your least-controlled output in front of your most valuable audience on day one.
AI does not fix a bad process, it accelerates it
The most useful finding in the 2025 DORA research on AI-assisted software development generalises well beyond software. Its central conclusion was that AI acts as an amplifier: it magnifies the strengths of organisations that already work well, and magnifies the dysfunctions of ones that do not.
Applied to workflows, that means automating a process nobody has looked at in four years mostly buys you a faster version of a process nobody has looked at in four years. Before you automate a job, spend an hour asking why it exists, whether all four approval steps are still necessary, and whether half the work would disappear if the form upstream asked one better question. We have watched clients cancel an automation project during that hour and come out ahead.
None of this is a reason to delay. It is a reason to spend the first week on the process and the next three on the build, rather than the reverse.
Start with one, and count the hours honestly
Pick a single workflow. Write down, before you build anything, how many times it runs a week, how long it takes now, and how you will know in six weeks whether that changed. If you cannot answer those three questions, you are not ready to build, and you will not be able to tell afterwards whether it worked. That discipline is the whole of measuring AI ROI without lying to yourself, compressed.
Then put it where the work already happens. The single largest cause of a working AI feature going unused is asking people to go somewhere new to use it. Connecting it into the systems your team already opens every morning is not a nice-to-have; it is the difference between a saving and a screenshot. That plumbing is what AI integration actually is, and it is usually where the majority of the effort lands.
One workflow, well chosen, running every day, with the hours written down before and after. That is a better position than most businesses in this country are in, and it is a smaller piece of work than the market wants you to believe. If you want help picking which one, that is the first conversation in any sensible AI consulting engagement.
FAQ
What is an AI workflow? An AI workflow is a fixed, predefined sequence of steps in which an AI model handles a specific language-shaped task, such as reading a document or classifying a message, while ordinary software handles the rest and a person reviews the exceptions. The path is written by you, so the cost, runtime and behaviour are predictable.
What is the difference between AI workflow automation and traditional automation? Traditional automation handles structured, rule-based steps: if this field equals that value, do this. AI workflow automation adds the ability to handle unstructured input, such as reading an emailed invoice or interpreting a free-text complaint, which previously required a person. The surrounding logic is still ordinary software.
Which business workflows are best suited to AI? Ones that run frequently, involve reading or writing text, could be explained to a competent new starter in a page of instructions, are cheap to get wrong, and use data already held in a system you can connect to. Document extraction, inbound message classification, meeting and call summarisation, and first-draft standard replies are the usual candidates.
How do I know if an AI workflow is worth the money? Before building, record how often the task runs, how long it currently takes, and what measurable thing should change. Six weeks after launch, measure the same three things. Without a before figure you cannot tell the difference between a saving and a feeling.
Should I automate a process before improving it? No. Published research consistently finds that AI amplifies whatever is already there, so automating an inefficient process produces a faster inefficient process. Spend an hour on why each step exists before you spend a month automating it.
References
- Bizcommunity / Lula: Funding constraints slowing AI adoption among manufacturing SMEs (News24 x Lula Small Business Survey, 1,088 SME owners, January 2026). https://www.bizcommunity.com/article/funding-constraints-slowing-ai-adoption-among-manufacturing-smes-519710a
- Google Cloud / DORA: 2025 DORA Report: State of AI-assisted Software Development (AI as amplifier; nearly 5,000 respondents). https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report
- Anthropic: Building effective agents (workflow patterns, and the case for the simplest solution that works). https://www.anthropic.com/news/building-effective-agents
- Gallup: AI in the Workplace: What Separates Adopters and Holdouts (fit with existing workflow as a leading predictor of use). https://www.gallup.com/workplace/704252/workplace-separates-adopters-holdouts.aspx
- South African Government / popia.co.za: Section 71: Automated decision-making. https://popia.co.za/section-71-automated-decision-making/
Written by JP, Sixees Labs. Last reviewed August 2026.