1. Direct AI usage: tokens, model choice and request patterns
The simplest place to start is the per-call cost. Work out: tokens per call × calls per month ÷ 1,000 × cost per 1,000 tokens for the model you plan to use. Don’t forget both prompt and response tokens, and that embedding requests are charged differently to text generation.
Watch for patterns that multiply costs: parallel calls, retries on failure, unnecessarily long prompts, and workflows that call the model for every event (for example a HubSpot or Pardot form submission). A short checklist:
- measure average tokens per request (prompt + response)
- count expected calls per day and multiply to monthly volume
- apply the model’s cost-per-1k-tokens to get a monthly spend estimate
2. Human review, exception handling and throughput
Models make mistakes and your workflow needs a plan for when they do. If you add a human-in-the-loop (quality review, approval, or corrections) estimate reviewer time per item and multiply by hourly cost; 2–5 minutes per review quickly dominates token costs for low-value items. Platforms differ only in where you surface those exceptions: a Salesforce case queue, a Marketo moderation list, or a HubSpot property that flags records.
Also budget for exception handling work: building decision gates (trust scores, confidence thresholds), routing rules, and simple fallbacks when the model fails. These are one-off build costs plus ongoing time to triage false positives and tune thresholds — treat them as recurring operational overhead, not a single development expense.
3. Monitoring, maintenance and deciding whether to keep the AI step
Running AI in production needs monitoring: logging calls, tracking error and fallback rates, and periodically reviewing prompt performance to prevent drift. Hidden ongoing costs include storage for logs, costs of observability tools (or time spent in spreadsheets), API key rotation, and prompt rework as data or policies change.
To judge whether the workflow is worth keeping, track three simple metrics over a month: time saved for the team, unrecoverable failures (customer impact), and total cost per processed item (token + human + monitoring). If time saved × value per hour < total cost per item, pause and redesign. If you’d like a short, practical audit of these numbers in your stack (HubSpot, Salesforce, Marketo or Pardot), Optira can help run the calculations and tests.