AI agent vs automation
Automation is deterministic: the same input always produces the same action. An agent is a probabilistic decision-maker. Deterministic is better whenever it is possible; the agent earns its place only where the input cannot be structured in advance.
| Dimension | Workflow automation / RPA | AI agent |
|---|---|---|
| Logic | Fixed rules: trigger → condition → action | Model chooses actions from context |
| Inputs | Structured: fields, IDs, webhooks, spreadsheets | Unstructured: text, voice notes, PDFs, photos |
| Output | Identical every time for the same input | May vary; constrained by schemas and guardrails |
| Errors | Break loudly when a screen or field changes | Fail quietly and plausibly if ungated |
| Cost | Platform subscription; ≈US$0 per run | Per token; ≈US$0.02–0.05 per WhatsApp conversation |
| Audit | Trivial: the rule is the log | Requires a trace of every decision and tool call |
| Change | Edit the rule | Edit the prompt/tools, then re-run the evaluation |
| Governance under the CDPA | Same data-controller duties; decisions are rule-based | Same duties plus the automated-decision rule where effects are significant |
| Right for | Invoice → accounting entry; form → CRM; payment → receipt | Voice note → shift report; sentence → booking; PDF → reconciliation proposal |
When automation wins outright
- The trigger is an event with structured data — a Paynow payment webhook, a form submission, a new row in a sheet.
- The condition is expressible in one line — "amount ≥ 100 and currency = USD".
- The action is fixed — create a receipt, move the CRM stage, send the template.
Tools like n8n, Zapier, Make or your CRM's native automations do this reliably and nearly free. If you are considering an agent for a job like this, you are paying tokens to guess what a rule already knows. For the wider question of which processes to automate first, businessai.co.zw is the network's SME automation guide.
When the agent is unavoidable
When the input is a voice note, a photo of a hand-written delivery note, a customer's sentence, or a supplier's PDF that never uses the same layout twice. No rule can parse "night shift, 42 tonnes, pump 3 down" into a form; a model can, and a gated tool can file the result.
The pattern that uses both: an agent inside a workflow
The workflow owns the trigger, the routing and every deterministic step. One node calls the agent (or just the model with a strict output schema) to do the single thing rules cannot — read the unstructured input and return structured fields — and the workflow carries on deterministically. The agent's blast radius is one step; its output is validated against a schema before anything else runs; and the whole thing is auditable because the workflow is.
trigger: WhatsApp message with a voice note
step 1: transcribe (deterministic service call)
step 2: MODEL → {"tonnes": number, "equipment_down": [..], "incident": bool} ← the only agent step
step 3: validate schema; if invalid → ask supervisor one question
step 4: write shift report (deterministic)
step 5: if incident → alert safety officer (deterministic)Most production "agents" in Zimbabwean businesses should look like this. Reserve the full loop — model choosing tools across many turns — for conversations, where the sequence of actions genuinely depends on what the customer says next. The booking demo is that case.
RPA specifically
Screen-replaying bots are still automation: deterministic, and brittle when the screen changes. Agents with computer-use tools can drive screens by reading them, which handles change better but costs tokens per click and needs a sandbox. Use an API if one exists; RPA if the screen is stable; a computer-use agent only when neither is true and the volume justifies it.
Next: agent vs chatbot, or the decision framework article.