Every December the same article appears with a new year in the title. It predicts transformation, lists the technologies that will deliver it, and is never revisited. We are writing this one in September so it has a full year to be wrong in public.
First, the honest position
We sell AI oversight and remediation. That gives us an interest in you being worried, and you should read the rest of this knowing that. It also means we spend our time inside systems built with AI assistance after something has gone wrong, which is a different vantage point from a vendor demo.
What follows separates three things that usually get blended: what has already changed and is simply not evenly distributed yet, what we expect to change through 2027, and what is being sold now that we would bet against.
Already true, unevenly distributed
Drafting is solved for most business writing. First drafts of proposals, policies, job specs and client updates are faster with assistance than without, for almost everyone. The constraint was never capability; it is whether the person reviewing the draft knows enough to catch what is wrong with it.
Search over your own documents works. Retrieval over a firm's own files — contracts, precedents, case notes, manuals — is a genuinely solved problem now, and it does not require sending anything to a public model. This is the single highest-value AI deployment we see in professional practices, and it is unglamorous enough that it rarely gets discussed.
Code assistance changed how software gets written. It also changed what gets shipped. We covered the specifics in building with AI; the short version is that assistance is excellent at producing code that runs and indifferent to whether it is safe.
What we expect through 2027
Agents will work in narrow lanes, with approval gates. The pattern that is already working: an agent operating inside one well-defined system, with a person approving anything consequential. Reconciling invoices against statements, triaging inbound enquiries, preparing a filing for review. Expect that to spread and get considerably better.
Local models will get good enough for more of the routine work. The gap between open-weight models you can run on your own hardware and frontier models is real and will remain at the hard end. But the proportion of ordinary business work that needs frontier reasoning is smaller than most people assume, and for anything touching confidential information the privacy argument usually outweighs the capability one. We build on that premise in our private AI knowledge systems.
Procurement questions will get sharper. Clients, insurers and regulators are starting to ask where your data goes, which models touch it, and who reviewed the output. Businesses that cannot answer will start losing work to businesses that can — not because of the technology, but because the question is now on the form.
The cleanup market will grow. Systems built quickly in 2025 and 2026 are reaching the point where someone has to maintain them. We expect a steady increase in businesses discovering that the thing that was built fast cannot be changed safely.
What we would bet against
This is the part most forecasts leave out, so it is the part worth holding us to.
Against: autonomous agents running end-to-end operations unsupervised. Demonstrations of this are impressive and the failure modes are quiet. An agent that is right 95% of the time across a ten-step process is right about 60% of the time end to end, and the 40% does not announce itself. Anyone selling unsupervised autonomy for consequential work in 2027 is selling ahead of the evidence.
Against: AI replacing professional judgement in regulated work. Not because the technology cannot produce a plausible answer, but because the liability does not move. An HPCSA-registered practitioner, an attorney or a professional engineer signs their name to the output. The signature is the product, and it cannot be delegated to a model.
Against: "just ask the AI" replacing having a system. A business with disorganised records and no process does not get organised by adding a model on top. It gets a faster way to produce confident answers from bad inputs.
Against: the price of frontier models collapsing to zero. It has fallen a long way and will keep falling for a given level of capability. The frontier itself stays expensive, because the frontier is defined by what is expensive.
What this means for a South African business specifically
Three things that are local rather than general.
Cross-border processing is a POPIA question. Using an overseas AI service to process personal information is a transfer of personal information outside the Republic, which Section 72 governs. This is not a reason to avoid these tools; it is a reason to know which ones you are using and on what data, and to be able to show that you thought about it.
Your professional body has a position, or will. Regulated professions are working out what AI assistance means for their codes of conduct. If you are in one, the answer to "may I use this" comes from your council, not from us or from the vendor.
The talent argument cuts both ways. AI assistance raises the floor for inexperienced builders, which is why the market is flooded with systems that look finished. It also raises the ceiling for experienced ones. The distance between those two outcomes got wider, not narrower.
What to actually do between now and 2027
Not a transformation programme. Four things, in order:
Write down where AI already touches your business. Including the tools staff signed up for themselves. You cannot govern what you have not listed — the same principle that applies to everything else you expose.
Decide what may never leave the building. Client files, patient records, matter notes, financials. That decision determines your architecture, and it is a business decision rather than a technical one.
Put a person between AI output and anyone who matters. Clients, patients, regulators, courts. One review step removes most of the risk that actually materialises.
Have anything built with AI assistance reviewed by someone who can read it. Independently of whoever built it. This is the specific gap we get called about, and it is much cheaper to close before a breach than after one.
Come back and check
We have said agents work in narrow lanes with approval gates, that local models take more of the routine load, that procurement questions get sharper and that the cleanup market grows. We have bet against unsupervised autonomy for consequential work, against AI absorbing professional liability, and against tooling substituting for process.
Some of that will be wrong. The point of writing it down in September 2026 with a date on it is that you can tell which.
If you want the version of this conversation that is about your business rather than the market, that is what our AI oversight work is for.