South African organisations are navigating an AI environment that’s evolving faster than any implementation roadmap can accommodate, says Andrew Rushworth, Practice Head of Microsoft Power Platform at Braintree.
There’s a particular kind of pressure that comes from a technology that refuses to stand still long enough to be fully understood. Enterprise AI is currently that technology. If anything, it’s setting new benchmarks in speed of change. What was considered a capable implementation six months ago has to potentially be rebuilt today, and companies taking the lead in the AI race aren’t necessarily the ones moving the fastest. They are the ones that are moving carefully and with a clear understanding of what they’re deploying and its measurable value over time.
Despite global enterprise AI spending projected to reach $665 billion, three out of four AI deployments are failing to reach their projected ROI. As a recent MIT study found, 95% of generative AI pilots fail but this isn’t because of the tech, it’s because of underlying, foundational weaknesses that are impacting how companies pull at the value of AI. At their core, AI projects need stability and strategy to step into their business potential, and they need companies to pay attention to the areas where the technology can most deliver longevity and that ever-elusive value.
At the ECS Conference that took place in Germany in June, this value came down to two themes that have emerged as key to enterprise AI: agent security and workflow orchestration.
Agent security starts with asking what companies are feeding the machine. The free tools employees reach for by default were not built to protect the data that’s entered into them and unless appropriate controls are in place, sensitive information may be stored, retained or used in ways the organisation cannot govern. It has become imperative that companies secure the data that flows into the agent, not just the agent itself because you don’t want to give away your information or intellectual property.
This exposure runs both ways, because if a public-facing company asks a customer for an ID number, they have to be certain the intelligence behind that request is secured. If you’re using AI agents to collate and collect data, you have to protect that data in line with regulatory requirements and data sovereignty expectations. Your AI has to be secure on the backend and your systems have to remain closed within a walled garden to protect your data and that of your customers.
On the second theme, workflow orchestration, there is an interesting change taking place in the enterprise. Static agents capable of responding to questions are being steadily replaced by orchestrated agents capable of taking a task, breaking it into its relevant parts, and working through them systematically. Users can present a problem to the agent and it is capable of spinning off different tasks, creating plans, fixing the issue, and recompiling the data. AI is capable of taking a more iterative approach to problems and this can fundamentally change how companies use the technology, and where they put it.
Underneath this capability sits the Model Context Protocol (MCP) standard that integrates with the AI models. Organisations and software vendors implement MCP servers that expose platform-specific capabilities while using the same open protocol. For instance, the MCP for Power Platform exposes tools, actions, schemas, API’s and metadata which allows it to interact with the Power Platform environment.
These are increasingly gaining momentum as more are being released across different systems, particularly within the Microsoft environment where MCPs for Finance & Operations, Customer Engagement and Dataverse are living, holding the knowledge required to operate effectively and deliver deeper value. Before MCP, connecting AI agents to enterprise systems required extensive and complex integrations and this played no small role in AI pilot failures. Now, with MCP standardisation, companies can move through AI agent pilots to production far more efficiently.
The jump from customised integration to MCP is an example of how this technology does not sit still. Anthropic introduced the technology in the latter part of 2024 and then donated it to the Linux Foundation’s Agentic AI Foundation in December 2025 and by which time it had exceeded 97 million monthly SDK downloads. The industry is in its learning phase and it may take another six months to a year before the market settles down enough to build with confidence. This is the paradox facing every organisation – the pressure to adopt now against a technology that will look different by the time the project ships.
For South African companies sitting behind the curve is not the disadvantage it looks like. The churn playing out in Europe can be skipped as companies adopt the next generations that have learned from what has gone before. What matters now, and what is strategically smart going forward, is to prioritise education first, then policy, then security and then the technology which is, in an environment rewriting itself in real time, the closest thing there is to a roadmap.
