Compare AI automation costs by converting the full cost of the same workflow into rand, including usage, hosting and human review. A cheap dollar rate can become an expensive operating choice if it requires longer prompts, repeated attempts or more checking. Build a scenario worksheet, then show where the preferred option changes rather than promising a saving.
The comparison below is a proposed planning method. Every example value is hypothetical, and the fictional provider rates are not market quotations.
1. Define the completed work before comparing providers
Use one accepted business outcome as the comparison unit. For example, compare the cost of preparing a customer enquiry for a staff member to review, rather than the cost of one API request. The workflow might classify the enquiry, extract reference details and prepare a draft reply. It does not send that reply without the agreed human approval.
Write down what counts as complete: required fields present, supporting information available, a usable draft and an exception flag where necessary. Keep rejected outputs and abandoned attempts in the cost numerator, but do not count them as successful outcomes.
Give each proposal the same input set, output requirements and review responsibilities. Otherwise, one quote may cover extraction only while another includes integration, monitoring and support.
For broader terminology, the AI automation glossary provides context. For this decision, the important boundary is the scope of workflow automation: which steps software prepares, which systems it touches and where people retain judgement.
2. Translate workflow volume into billable usage
Estimate billable usage from the entire processing path, not just the customer's message. Include instructions, attached content, retrieved records, tool definitions, generated responses and repeat attempts. Separate input and output tokens because they may have different rates.
A proposed usage log should record a business item ID, attempt ID, model configuration, billable usage, tool charges, status and review time. Keep the business item ID stable across attempts so retries cannot inflate the apparent number of completed items.
Create separate rows for ordinary enquiries, long attachments and exceptions. A single average can hide the expensive tail. Treat caching discounts as a separate assumption until the workload demonstrates eligible reuse; do not assume every repeated instruction receives a discount.
The Source: OpenAI API changelog records separate capability releases on 29 September 2026, including computer use in the Agents API and GPT-6.1 Sol. Such announcements do not establish the usage of your workflow. Ask suppliers to identify the exact model, endpoint, processing tier and any preview, access or regional conditions on which their quote depends.
3. Convert dollar exposure without hiding exchange assumptions
Convert each dollar-denominated subtotal using a visible budget exchange rate and a separate conversion-fee assumption. Do not convert rand hosting or local labour costs again.
Use this proposed formula:
Rand dollar exposure = USD subtotal × ZAR per USD × (1 + conversion fee percentage)
Where a supplier provides an all-in settlement rate, use that rate and omit any fee already included. Otherwise, the worksheet could charge the same conversion cost twice. Record who supplied the rate, its date and whether it is a planning assumption or an actual settlement rate.
For a hypothetical sensitivity range, use R17, R18.50 and R21 per dollar. These are scenario inputs, not exchange-rate forecasts. First change the exchange rate alone, then combine it with higher usage to show a stress case.
Keep tax and cash-flow treatment outside unsupported assumptions. Ask the finance owner or accountant to confirm invoice treatment, applicable taxes, recoverability and payment fees. Present cash payable and comparable operating cost separately where their treatment differs.
4. Add hosting, maintenance and human capacity
Include the infrastructure and people needed to keep the workflow usable. List orchestration, database, storage, backups, monitoring, support, licence charges and maintenance. Mark each amount as included, separately quoted or unresolved.
Architecture can change the fixed-cost profile. The Source: n8n queue-mode documentation describes a main instance, Redis, workers and a database. These components need explicit provision in a queue-based quote. Do not treat a worker server as the whole hosting bill. The documentation also limits viewing running workers to Enterprise arrangements, so a proposal relying on that interface should confirm the relevant plan.
The n8n AI agents resource is useful when scoping an orchestrated workflow. The custom AI agents resource supports a different question: whether the proposed processing path needs custom behaviour. Neither choice removes operating costs.
Model human review as items reviewed multiplied by minutes per item and a finance-approved hourly cost. Also check whether reviewers can cover busy periods. Available capacity and accounting cost are different constraints.
5. Build scenarios around uncertainty, not optimism
Use low, expected and stress columns with explicit assumptions. The expected column is a planning estimate, not a measured result. Change volume, input length, retries, review share, review duration and exchange rate independently before combining them.
For a proposed evaluation, give both options the same representative records and have a reviewer score completeness, unsupported statements and correction effort. Include difficult inputs deliberately. Record the configuration used so a later change does not silently invalidate the comparison.
The Source: Claude Sonnet 5.5 announcement, dated 28 September 2026, reports vendor findings about lower per-task cost and faster output relative to its predecessor. Those findings do not establish savings for this workflow. Use them as a reason to evaluate an option, not as a percentage to subtract from a quote. Confirm API access, configuration and relevant safeguards before relying on a named model.
Proposed acceptance rules should require both acceptable outputs and adequate review capacity. A lower forecast cost cannot compensate for unusable drafts or unresolved security requirements.
6. Use a reusable rand cost-comparison worksheet
Copy this worksheet once for each option and keep all assumptions visible. Use separate sheets for different workflows if their usage or review requirements differ substantially.
Rand cost-comparison worksheet
All entries are proposed inputs until supported by a quote, usage record or finance decision. Create Low, Expected and Stress columns for each option.
| Field | Entry or calculation |
|---|---|
| Scope | Workflow, accepted outcome, exclusions, owner |
| Volume N | Unique business items entering monthly |
| Accepted A | Items meeting the agreed completion criteria |
| Usage rows | Billable quantity × USD rate for each input, output, cache, tool or other category |
| USD total U | Sum usage rows, dollar subscriptions and dollar hosting |
| Exchange E | Budget rand per dollar, date and basis |
| Conversion f | Fee as decimal; exclude fees already in E |
| Local fixed L | Rand hosting, licences, monitoring and maintenance |
| Review H | Total review and correction minutes ÷ 60 × hourly rand cost |
| Setup S | Once-off implementation, integration and training cost |
| Period M | Agreed comparison period in months |
| Monthly T | U × E × (1 + f) + L + H + S ÷ M |
| Unit cost | T ÷ A; undefined when A is zero |
| Period total | M × recurring monthly cost + S |
| Finance treatment | Taxes, payment fees and recoverability confirmed separately |
| Capacity | Peak arrivals, processing limits, review hours and backlog allowance |
| Decision record | Quality evidence, unknowns, breakpoint, owner and next review date |
Proposed completion rule: do not rank an option with unresolved scope, billable units, acceptance criteria or finance treatment. Show it as incomplete rather than entering zero.
Before circulating the sheet, check units carefully. A rate per million tokens must multiply token quantity divided by one million. A monthly subscription should not be multiplied by the number of items. Keep once-off implementation separate from recurring costs even when showing its monthly allocation.
7. Calculate breakpoints and decide what would change the choice
Calculate where the total-cost lines cross rather than declaring one option universally cheaper. For a simplified hypothetical comparison, Option X has R3,500 monthly fixed cost and R2.40 variable cost per item. Option Y has R5,500 fixed cost and R1.40 per item. Assume the same accepted outcomes and review capacity.
The hypothetical crossover is:
(R5,500 − R3,500) ÷ (R2.40 − R1.40) = 2,000 items per month
Below that volume, X costs less; above it, Y costs less within these assumptions. This simplified comparison is separate from the worked example below. Real hosting costs may rise in steps, requiring a new calculation when another worker or licence becomes necessary.
Calculate a review breakpoint too. If one option's hypothetical infrastructure and API advantage is R900 monthly, an extra three review hours at R300 per hour removes it. That is not evidence that either option needs those hours. It identifies what the evaluation must measure.
The decision record should state the preferred option, unresolved conditions and the volume, exchange rate or review effort that would trigger reconsideration.
Worked walkthrough: normal work and exceptions
A hypothetical enquiry workflow receives 2,000 unique items monthly. Its fictional billable totals, including allowed retries, are 20 million input tokens at $1 per million and four million output tokens at $5 per million. The API subtotal is $40.
At a hypothetical R18.50 per dollar and a 3% conversion fee, this becomes R762.20. Add hypothetical local fixed costs of R2,500. If 200 items require four minutes of review at R300 per hour, review costs R4,000. Recurring monthly cost is therefore R7,262.20. A hypothetical R12,000 setup cost spread across 12 months adds R1,000, producing R8,262.20 for comparison. If all 2,000 items meet acceptance criteria, that is approximately R4.13 per accepted item.
For a normal record containing a clear enquiry and reference number, the proposed process prepares the classification and draft for the agreed review path.
For a missing reference number, it flags the gap. A staff member requests clarification rather than allowing repeated model attempts to guess the value. Log clarification and correction time separately from routine review.
For a duplicate submission, the proposed process flags matching identifiers for human confirmation. If a call already occurred, retain its cost, but count the accepted business outcome only once. If only 1,900 unique outcomes are accepted, the same comparison total becomes about R4.35 each. That denominator change matters even without a higher provider bill.
FAQs
What if a quote gives one dollar price per workflow run?
Ask what constitutes a run and which charges sit outside it. Confirm whether failed executions, retries, tools, storage and provider usage are included. Use the quoted run price only for its defined scope. If usage is passed through separately, create a second row rather than assuming inclusion. An unclear charging unit is an unresolved comparison input, not a reason to estimate it as zero.
Should we assume the weaker-rand scenario for every month?
Not necessarily. Use a finance-approved budgeting assumption and show a weaker-rand sensitivity alongside it. Apply exchange changes only to dollar exposure. Then calculate how much the full monthly total moves. If human review dominates, exchange changes may have less influence than review duration. If API or dollar hosting dominates, the reverse may apply. Avoid presenting either scenario as a prediction.
Can we select the cheapest option before testing review effort?
You can shortlist it, but keep the choice conditional. Use identical records, including missing and duplicate cases, to estimate accepted outcomes and correction time. Also confirm contractual scope, access conditions and operational capacity. Legal, payment and security consequences remain matters for the appropriate human owners. No model price or formatted output establishes that a workflow is safe to approve or deploy.
Turn the comparison into a scoped next step
Use the completed worksheet to ask suppliers for a like-for-like proposal with clear exclusions and charging units. A useful AI automation discussion should cover the processing path and ongoing responsibilities, not only the model choice.
If your business needs help comparing workflow proposals in rand, get in touch with Symaxx to discuss the scope, assumptions and evaluation plan. Bring the worksheet and unresolved questions so the next step addresses the actual decision.