Agentic AI Replaces RPA for SA E-commerce Automation

A scripted bot can place an order, copy an invoice number, and update a spreadsheet. But the moment a product name changes, a supplier sends a “quick note” as a PDF, or a customer asks for a swap because a gift arrived late, the bot stalls. That’s the practical reason so many teams are quietly frustrated with robotic process automation (RPA) in 2025.

RPA isn’t “bad”. It’s just the wrong centrepiece for modern e-commerce and digital services in South Africa, where customer journeys run across WhatsApp, email, apps, marketplaces, call centres, and back-office platforms that rarely behave exactly the same way twice. What’s replacing it is agentic AI: systems that can interpret messy inputs, plan actions, call tools safely, and deliver outcomes rather than mimic clicks.

This post is part of our “How AI Is Powering E-commerce and Digital Services in South Africa” series, and it’s focused on one very specific shift: moving from task automation to outcome automation—without losing control, auditability, or compliance.

RPA didn’t fail — it got outgrown

RPA’s core limitation is simple: it automates steps, not intent. Scripted bots do exactly what they’re told, which sounds comforting until you realise how often e-commerce operations change week to week.

Deon van Niekerk (CTO at Ovations Technologies) put it bluntly: RPA as a strategic automation paradigm is dead—mainly because the enterprise has outgrown what deterministic, UI-driven scripts can handle. The reality in retail and digital services is constant variation: promotions, partial shipments, supplier substitutions, backorders, fraud checks, and customer exceptions.

Why RPA becomes a cost centre in fast-moving retail

For South African e-commerce teams, RPA pain usually shows up in three places:

• Brittleness: A UI label changes from “Submit” to “Place Order”, and your bot breaks.
• No judgement: A bot can’t interpret an angry email thread or a scanned ID document without a pile of add-ons.
• Linear scaling: More processes means more bots—each with its own maintenance burden, credentials, exceptions, and failure modes.

RPA still has a place (we’ll get to that), but it stops being “strategic” when the business is trying to automate end-to-end journeys like refunds with risk checks, delivery exception handling, or customer support triage across channels.

A good litmus test: if your process includes negotiation, ambiguity, or unstructured content, RPA will disappoint you.

Agentic AI: outcome automation for e-commerce and digital services

Agentic AI flips the model. Instead of scripting every step, you define the outcome—“resolve this delivery exception” or “publish a compliant product listing”—and the agent plans and executes the sequence using approved tools and policies.

In the enterprise framing from the source article, the modern stack looks like this:

• AI agents as the “brain” (reasoning, planning, interpreting context)
• Model Context Protocol (MCP) as the standard connector to tools and systems
• Headless BPM/workflow engines as the “nervous system” (state, governance, audit, routing)

That combination matters in retail because you don’t want a free-roaming chatbot. You want governed autonomy: agents that can act, but only within your rules.

What agentic AI does that RPA can’t

Agentic AI handles the realities that dominate e-commerce operations:

1. Unstructured inputs: PDFs, email threads, chat logs, voice call transcripts
2. Context switching: customer history, order history, delivery status, inventory constraints
3. Exception handling: “If out of stock, offer substitutes under R200, otherwise escalate”
4. Multi-step coordination: contacting supplier, updating customer, issuing voucher, logging case

I’ve found the biggest mindset shift is this: you stop automating screens and start automating decisions—then you log every decision and keep humans in the loop where risk is high.

Where this shift hits hardest in South African e-commerce

If you run an online store, marketplace operation, or digital service in South Africa, you’re probably dealing with a familiar mix: load shedding disruptions, courier variability, supplier fragmentation, and customers who expect immediate support on mobile-first channels. Agentic AI is particularly strong in workflows where time-to-resolution drives revenue.

1) Customer support: from “ticket handling” to “issue resolution”

Answer first: Agentic AI reduces support costs by resolving more issues without hand-offs.

A practical agent flow might look like:

1. Detect intent from inbound WhatsApp/email (“wrong size”, “late delivery”, “missing item”)
2. Pull order details, delivery scans, return policy, and customer history
3. Propose a resolution (exchange, refund, reship, voucher) based on policy and risk score
4. Execute steps via tools (create return label, book courier pickup, update ERP, notify customer)
5. Escalate to a human only when confidence is low or risk is high

This matters because speed is retention. When your competitor can resolve a problem in one interaction, a 3-day ticket ping-pong is expensive.

2) Product content creation that stays compliant and consistent

Answer first: Agentic AI can produce product copy at scale and enforce brand and legal rules.

South African retailers often manage thousands of SKUs across categories with different compliance needs (health claims, electronics specs, warranty wording, pricing disclosures). A governed agent can:

• Extract specs from supplier sheets and PDFs
• Generate title, bullets, long description, and FAQs in your brand tone
• Check claims against approved policy documents
• Flag missing attributes needed for search and filters
• Route exceptions (e.g., “supplement claims”) to a human reviewer

This is where the “agent + policy grounding” concept from the RSS content becomes very real: your policies become executable guardrails, not a PDF no one reads.

3) Returns and refunds: fewer mistakes, less fraud

Answer first: Agentic AI improves refund speed while adding smarter checks.

Refund operations are a high-volume, high-dispute area. Agentic AI can:

• Verify return eligibility based on product category and timeframe
• Identify risk signals (repeat claims, delivery confirmation conflicts)
• Choose the correct refund method (original payment, voucher, store credit)
• Trigger communications and update records consistently

RPA can “click through” the refund screen. It can’t reliably decide what to do when the parcel is scanned as delivered but the customer claims it wasn’t.

4) Back-office automation that actually scales

Answer first: Agents scale by reusing reasoning and tools, not by cloning brittle scripts.

Think of reconciliations, supplier onboarding, invoice matching, and stock discrepancy investigations. With agentic AI, you’re not building 50 bots; you’re building a smaller set of capabilities (tools + rules + workflows) that agents can reuse across processes.

MCP and workflow engines: the part people skip (and regret)

A lot of “AI agent” projects fail because teams treat them like chatbots with permissions. The RSS article highlights two enterprise ingredients that stop this from becoming chaos: MCP interfaces and headless BPM engines.

MCP as “USB-C for AI” (and why it matters in retail)

Answer first: MCP standardises how agents discover and use tools, which reduces fragile integrations.

In an e-commerce environment, tools include:

• Order management systems
• ERP and inventory services
• CRM and support platforms
• Courier tracking portals/APIs
• Payment gateways and fraud tooling

Instead of writing one-off “glue code” for each integration and agent, MCP-style tool interfaces allow the agent to use approved actions in a structured way—read order, create return, issue voucher, update address—with predictable inputs and outputs.

Workflow engines: how you keep auditability and control

Answer first: A workflow engine provides state, logs, approvals, and exception routing.

Retail processes often run longer than one session: returns can take days, supplier queries weeks, delivery investigations multiple hand-offs. A headless workflow engine gives you:

• State management (where is this case in the journey?)
• Audit trails (who/what changed what, and why?)
• Approvals (human sign-off for high-value refunds)
• SLAs (time-based escalations)

If agentic AI is going to power customer-facing outcomes, governance isn’t optional. It’s the product.

Don’t rip out RPA — shrink it to where it belongs

RPA still works well when the world is stable and deterministic. Even the RSS source acknowledges it won’t vanish overnight. My stance: keep RPA as a tactical tool, not the operating model.

Use RPA when:

• The UI is stable and rarely changes
• The task is repetitive and rules-based
• There’s no API and you can’t justify building one
• Reliability matters more than “understanding”

But for anything involving customer language, policy interpretation, cross-channel context, or exceptions, you’ll get better outcomes by placing agentic AI at the centre and letting RPA handle the last-mile UI only if necessary.

A practical migration plan for SA retailers and digital services

Answer first: Start with one end-to-end journey, add governance early, and measure outcomes.

Here’s a rollout approach that tends to work in the real world:

Step 1: Pick one workflow with clear pain and clear metrics

Good candidates:

• Delivery exceptions (late/lost/damaged)
• Refund approvals
• Customer support triage
• Product listing enrichment

Define success in numbers (examples):

• Reduce average time-to-resolution from 48 hours to 6 hours
• Increase first-contact resolution from 35% to 55%
• Cut manual content production time per SKU from 20 minutes to 5 minutes

Step 2: Turn your policies into “agent rules”

Convert return policies, fraud thresholds, brand tone guidelines, and escalation rules into structured, testable artefacts. If it’s only in someone’s head, the agent can’t apply it consistently.

Step 3: Add a workflow engine before you add more agents

This is where most teams save themselves later. Get state, logs, approvals, and exception queues working early.

Step 4: Connect tools the safe way

Expose only the actions you’re comfortable with (issue voucher up to R500, refund requires approval above R2,000, etc.). Treat tool permissions like financial controls.

Step 5: Run a “human-in-the-loop” period

Let agents propose actions first, then execute with approvals. Tighten prompts, rules, and tool constraints based on real failure cases.

What to do next

Gartner’s widely cited trajectory (referenced in the RSS content) is that AI agents are moving toward autonomous, goal-driven digital workers, with a meaningful share of enterprise decisions expected to be made autonomously by 2028. Whether your business buys that forecast or not, the direction is obvious: customers are already behaving like they expect instant, personalised outcomes.

If you’re in South African e-commerce or digital services, the question isn’t “should we automate?” It’s whether your automation strategy is stuck on brittle scripts while your customers, channels, and operations get messier each quarter.

If you want help scoping a first agentic AI workflow (support, content, returns, or back-office), start by documenting one journey end-to-end: inputs, tools, policies, exceptions, and hand-offs. Once that’s clear, the right architecture almost designs itself.

What’s the one process in your operation that breaks every time the business changes—and would be worth fixing for good?