OpenAI Wants AI Agents to Do More of the Work. Businesses Still Need to Design the Work.

What happened

TechCrunch reports that OpenAI is pushing its agentic tools beyond coding and toward broader professional work through ChatGPT Work. The reported aim is to let people delegate multi-step tasks that involve the applications, files and communications systems already used in their day-to-day jobs.

This is a meaningful shift from AI as a question-and-answer interface to AI as an execution layer. An agent can be given instructions, access to approved tools and a defined outcome, then carry out steps across systems rather than simply produce a draft for a person to copy and paste.

OpenAI's own enterprise material describes a similar transition: firms are moving from asking AI for assistance to using it for execution. Its research also reports rapid growth in non-developer use of Codex-like agentic tools, while noting that longer, more complex tasks are becoming a larger share of activity.

The underlying technology should not be confused with universal autonomy. OpenAI's technical documentation describes agents as systems built from a model, instructions, tools and runtime controls such as guardrails, handoffs and structured outputs. In other words, useful agents are operational systems, not just capable models.

Why it matters

For many businesses, the most promising use cases are not dramatic replacements of whole roles. They are repeatable pieces of work that sit between people and systems: preparing a customer case from approved information, checking documents for missing fields, assembling a weekly operational report, routing requests, updating records or drafting a response for review.

That could reduce context switching and manual administration. But it also increases the importance of access management. An agent that can search shared drives, read messages or update a business system can create value only if it has the right permissions, a narrow enough task boundary and a clear route for human escalation.

The adoption gap is likely to be wider outside technical teams. Engineers are accustomed to version control, testing and structured inputs. Finance, operations, sales and customer-service teams often work across email, spreadsheets, documents and legacy systems, where processes may be poorly documented. An agent exposes those gaps quickly.

This means the key question is not whether every employee will use an agent. It is whether an organisation can identify a workflow where the outcome is measurable, the inputs are dependable and the risk of an incorrect action is manageable.

Our perspective

At SDA Systems, we see agentic AI as a workflow-design exercise before it is a model-selection exercise. Start with a process map: what triggers the task, which data sources are authoritative, what decision rules apply, which actions may be automated and where must a person approve the result.

The first production use case should be bounded and observable. For example, an agent may classify incoming requests, pull relevant policy information, prepare a structured case summary and send it to a queue for human approval. That provides operational value without immediately giving an AI system unrestricted authority to communicate externally or change core records.

Controls should be designed into the workflow rather than added after deployment. Practical controls include least-privilege access, approved data sources, action limits, audit logs, exception queues, test cases based on real work and named process owners who can update rules as the business changes.

South African organisations should apply the same discipline when working with customer, employee and supplier information. The technology conversation needs to sit alongside data protection, records management, vendor governance and the operational reality of the systems already in use. A polished agent interface does not remove these responsibilities.

The likely winners will not be the organisations that deploy the most agents. They will be the ones that redesign a small number of high-friction processes, measure the result and scale only where reliability and accountability are proven.