Most AI budget conversations start and end with the tool subscription. That's the wrong place to start — and it's why so many implementations run over budget, underdeliver, and get quietly shelved.
This post breaks down every layer of cost in a production AI workflow: the platform fees you can find on pricing pages, and the hidden costs that don't appear until you're three months in. If you're a South African business weighing up an AI build, read this before you sign anything.
The Four Cost Layers Nobody Puts on One Page
Running an AI-powered workflow has four distinct cost layers. Most vendors quote one of them.
Layer 1: Platform and orchestration fees
This is the most visible cost — the monthly subscription to the tool that connects your systems and runs your automations. The main players in this space are Make (formerly Integromat), Zapier, and n8n, and their pricing structures differ meaningfully.
Make's Core plan sits at around $9/month for 10,000 operations, with a Pro plan at roughly $16/month. Zapier's Professional plan runs $49/month for 2,000 tasks. At higher volumes the gap widens sharply: according to a detailed 2026 pricing comparison, running a complex 10-step workflow across 10,000 records a month costs around $79/month on Make versus $799/month on Zapier for the same logic. n8n offers a self-hosted Community Edition that is free — you only pay for the server you run it on — with cloud plans starting at around $20–$24/month for 2,500 workflow executions.
For South African businesses, the rand/dollar exchange rate is a real variable here. A $69/month Zapier Team plan (roughly R1,300 at current rates) sounds manageable until you start processing serious volumes and need to upgrade tiers.
The principle: Match the tool to your volume and technical capacity. For most SA SMEs doing under 10,000 automations a month, Make or n8n Cloud is the pragmatic choice on cost. Zapier makes sense when your team is non-technical and needs to self-manage.
Layer 2: AI model (API) costs
If your workflow includes any AI processing — document extraction, email triage, summarisation, classification — you're paying for model tokens on top of your orchestration tool.
Current OpenAI pricing (verified October 2026): GPT-4o runs at $2.50 per million input tokens and $10.00 per million output tokens. GPT-4.1, the newer workhorse model, comes in at $2.00 input and $8.00 output per million tokens. For lighter tasks, GPT-4o mini drops to $0.15 input and $0.60 output per million tokens — a meaningful saving for high-volume, lower-complexity jobs like email classification.
To put this in SA business terms: imagine a Cape Town property management firm processing 5,000 inbound emails a month through an AI triage workflow. Using GPT-4o mini, and assuming roughly 500 tokens per email, that's 2.5 million tokens — around $0.375 for input plus $0.75 for output, call it $1.25 a month in model costs. Even at GPT-4o rates, you're looking at $18.75 a month. The model cost itself is rarely the expensive part.
Where token costs do add up is in poorly designed prompts, unnecessarily verbose system instructions, or workflows that call expensive models for tasks a cheaper one could handle. Prompt engineering isn't glamorous, but it directly affects your monthly API bill.
Layer 3: Setup and integration
This is where most budgets get ambushed.
A focused single-workflow automation with simple integrations typically costs $1,500–$4,000 to build properly. Multi-step workflows connecting several tools, with conditional logic and error handling, generally land in the $7,000–$12,000 range. According to a 2026 analysis of AI automation projects, well-scoped builds at this level tend to pay for themselves within two to six months.
But setup is not just configuration time. Integration with your existing systems — your accounting platform, CRM, ERP, or industry-specific software — often takes more engineering hours than the workflow logic itself. South African businesses frequently run a mix of locally prevalent tools (Sage, Xero, SYSPRO) alongside international platforms, and the connectors aren't always pre-built. When they're not, you're paying for custom API work.
One frequently underestimated setup cost is data preparation. According to Deloitte's 2025 study of 300 mid-market AI deployments, data preparation accounted for 38% of total first-year implementation costs on average — the largest single category. As one Unravel Data executive put it when speaking to TechTarget, data infrastructure and readiness can consume roughly $60 out of every $100 spent on AI projects.
What does this look like in practice? Consider a Durban logistics firm that wants to automate invoice matching against purchase orders. The workflow logic itself is straightforward. But if their historical invoice data lives across three email inboxes, two spreadsheets, and a legacy system with inconsistent supplier naming conventions, the data cleanup phase could take weeks before a single automation runs. That time costs money whether you're paying a consultant or burning your own staff hours.
Layer 4: Ongoing maintenance and human oversight
This is the cost that disappears from vendor pitches entirely.
Production AI workflows are not set-and-forget. According to analysis published by Apptio, production AI requires continuous evaluation, monitoring, security, auditability, and human oversight. AI systems trained on historical data degrade over time as your business changes — a supplier changes their invoice format, a product category gets renamed, an API gets deprecated. Industry benchmarks put ongoing maintenance costs at 15–30% of the original build cost per year.
Human oversight doesn't disappear either. As one AI operations lead quoted in TechTarget put it: "AI is not a one-time purchase; it's a living system that requires continuous human oversight and operational investment." In practice this means someone in your business needs to review flagged exceptions, spot-check outputs on high-stakes decisions, and manage the escalation path when the system encounters something it wasn't trained to handle.
For SA businesses, this has a practical implication: budget for 2–4 hours of internal review time per week for the first six months of any new AI workflow. That's not inefficiency — that's quality control on a system that processes your data at speed.
What Does a Realistic First-Year Budget Look Like?
Here's a worked example. A mid-sized Johannesburg accounting firm wants to automate two workflows: client document intake (extracting data from uploaded documents) and invoice approval routing.
| Cost item | Estimate (ZAR) |
|---|---|
| Orchestration platform (Make Pro, 12 months) | ~R3,600 |
| AI API costs (GPT-4o mini, estimated volume) | ~R2,400 |
| Setup and integration (2 workflows, existing systems) | R45,000–R90,000 |
| Data preparation and cleanup | R15,000–R40,000 |
| Ongoing maintenance retainer (post-launch) | Scoped per provider |
| First-year total (conservative) | ~R66,000–R136,000 |
Those ranges are wide because scope drives cost more than any other variable. Two firms could run the same workflow at radically different cost depending on how clean their data is, how many systems need connecting, and whether their existing tools have working APIs.
The ROI case has to sit alongside this. If those two automated workflows save two staff members 15 hours a week each — a conservative estimate for document-heavy processes — you're recovering roughly 120 staff-hours a month. At a blended internal cost of R350/hour, that's R42,000/month in recovered capacity. The payback period on a R90,000 setup cost is under three months. That maths works. But it only works if the build is scoped and executed properly.
The Mistake Most Businesses Make
According to ITWeb's October 2026 reporting on AI ROI, only about a quarter of enterprises are converting their AI investments into significant financial returns, and 58% report steadily increasing returns. That means a meaningful portion of businesses investing in AI are not seeing clear payback.
The pattern behind failed implementations is consistent. An MIT Media Lab study reported in Fortune found that 95% of corporate generative AI pilots delivered no measurable P&L impact. S&P Global Market Intelligence found that 42% of companies abandoned most of their AI initiatives in 2025 — up from 17% the year before. These aren't projects that ran out of money. They're projects that ran before anyone had done the unsexy groundwork: mapping the actual process, cleaning the data, defining what "success" means in a number, and designing the human oversight layer.
Dell Technologies' Q2 2024 Innovation Catalyst Research found that 49% of South African respondents were creating a dedicated budget for AI projects. Budget is necessary — but it's not sufficient. Budget without a scoping process is how you end up paying for a demo-quality build that falls apart in production.
Before You Spend Anything: Three Questions
1. What specific, measurable outcome does this workflow produce? Not "improve efficiency" — that's not a number. Hours saved per week, invoices processed per day, response time in minutes. If you can't define the output metric, you can't evaluate whether the build worked.
2. Is your data ready? Ask yourself: could a new employee find and understand the information this workflow needs? If the answer is no, budget a data preparation phase before you budget a build phase. It will cost you either way — better to know upfront.
3. Who owns this after launch? AI workflows need a named internal owner responsible for exception management, output review, and escalation. If that person doesn't exist in your org chart yet, add them before you add the workflow.
What to Do Next
If you're trying to get clarity on what an AI workflow would actually cost your business — not a generic range, but a scoped estimate based on your systems, your data, and your specific process — that's exactly what our AI audit is designed to produce.
For R4,500, we map your highest-value automation opportunity, assess your data and integration readiness, and give you a plain-language cost and ROI model before any build work starts. It's the due diligence step that most businesses skip — and the one that separates implementations that earn their keep from the ones that get shelved.
Explore what we build on our services page, or browse more breakdowns like this in our Insights section.