For a couple of years, AI in software delivery mostly meant one thing: a coding assistant in the editor. That is no longer the world we are in. Agents now plan work, write and review code, generate tests, prepare releases and raise pull requests on their own — and plenty of enterprises are wiring several of them together across the delivery lifecycle. The board question has moved with them, from “should we use AI in delivery?” to something harder: can we trust what it delivers?
Adoption is racing ahead of production
Look across this year’s industry research and one pattern repeats: nearly everyone has adopted agents somewhere, but only a small fraction — one in ten, give or take, depending on the survey — actually runs them in production, where the value lives. Meanwhile Gartner has warned that more than 40% of agentic AI projects could be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
Our reading: this is not a technology gap. Given clear specifications and firm architectural boundaries, agents can genuinely compress delivery from months to weeks. The gap is in the operating model — pipelines are accelerating faster than the planning, quality, release and governance structures around them. An agent that produces ten times a developer’s output also produces ten times the unreviewed change, the untested assumptions, the audit surface. Unless the delivery system is redesigned to keep up, the maths eventually catches up with you.
The South African picture: regulators are watching now
This has stopped being an abstract, offshore debate. The South African Reserve Bank and the Financial Sector Conduct Authority have published their first cross-sector snapshot of AI adoption in the financial industry, and the direction is unmistakable: banks and insurers are moving AI out of pilots and into core operations, with close to half of surveyed banks planning materially bigger AI budgets this year. The same work flags that governance frameworks remain uneven — many institutions are stretching existing risk structures rather than building dedicated AI oversight.
Now add POPIA, which already constrains automated decisions about individuals, and the draft National AI Policy moving towards approval. For South African enterprises the message is fairly simple: if AI touches the delivery of customer-facing software, someone will eventually ask you to show how it is controlled. What the agent did, why, on whose authority, and what checked it before it reached production.
What good looks like in an agentic delivery organisation
The organisations getting real value from agents share a recognisable pattern. Five things, most of them unglamorous.
1. Specification before generation
Agents work from explicit, reviewable specifications — not hopeful prompts. The quality of what ships gets decided before a line of code exists.
2. Boundaries agents cannot cross
Clear architectural and permission boundaries define where agents may act on their own and where a human decision is mandatory. Anything touching money, personal information or production configuration sits behind a person. No exceptions.
3. Human accountability at the release gate
People stay accountable for what reaches production. Review gets redesigned for agent-scale output — risk-based, automated where safe — instead of quietly abandoned.
4. Delivery intelligence, not anecdotes
Leaders see agent and team performance in one connected view: cycle time, rework, defect escape, cost, risk. What you cannot measure, you cannot govern. Or prove.
5. Traceability by default
Every agent action is logged and attributable, so audit and compliance questions get answered from the system of record rather than reconstructed from memory.
Where to start
Not with more tools, tempting as that is. The organisations that will get the most out of agentic delivery over the next eighteen months are the ones that first understand their current delivery system — where work slows down, where quality risk concentrates, where an agent would genuinely remove a constraint rather than amplify a mess. Then introduce agents deliberately: one governed workflow at a time, measured before and after, with the controls above in place from day one.
Software itself is not the competitive advantage. The ability to turn business priorities into reliable software — now with agents in the loop — is. The enterprises that treat this as a delivery-capability change, rather than a tool purchase, will still be running their agents in production while the 40% of cancelled projects get written off.
Wondering where your organisation stands? Review your AI delivery readiness or discuss your delivery priorities with us.
%402x.svg){kind=icon}
%402x.svg){kind=icon}
