First calculate the whole process, not the model
The most expensive automation mistake is counting only the subscription. If a service costs 20,000 ₽, that does not mean the process costs 20,000 ₽: an employee may review every result, the integration may fail, and support may eat two days a month.
The mistake I would remove first: Counting only the subscription.
What to prepare and what result to expect
- Outcome: You get the cost of one successful operation and a payback period — not a pretty estimate based only on the plan.
- Describe run frequency, input, result, exception share, employee time, and required quality.
- Keep source data and access rights separate from the output so you can verify what full AI automation cost work actually did.
- Run a pilot of 100–300 operations and compare the manual baseline, cost, errors, and employee time.
Break price into line items and validate with a pilot
- Break down the costs
Create a table: model, subscriptions, operations, integrations, setup, support, training, human review, fixes, and risks.
Проверьте: The calculation includes the process owner’s cost and exceptions.
Если не сработало: Do not set an unknown line item to zero.
- Calculate the model
Model cost = calls × ((input tokens / 1,000,000 × input rate) + (output tokens / 1,000,000 × output rate)). Check rates and limits before calculating.
Проверьте: The calculation uses real volume, not a max example from ads.
Если не сработало: Measure the actual average on a test sample.
- Add the cost of error
Account for refunds, reprocessing, downtime, manual review, and fix time. Assign different weights to different error types.
Проверьте: An expensive failure is not hidden inside an average accuracy percentage.
Если не сработало: Separate critical and non-critical errors.
- Compare with the manual process
Payback period = setup / monthly net savings. In the pilot, compare cost per successful operation, not the number of runs.
Проверьте: The expand-or-stop decision is tied to thresholds.
Если не сработало: Reduce volume or change the process, but do not force the result.
Three tabs in Google Sheets
Create `Assumptions`, `Runs`, and `Monthly`. Fields: `workflow_id`, `unit`, `monthly_volume`, `manual_minutes`, `hourly_rate`, `error_rate`, `error_cost`, `input_tokens`, `output_tokens`, `calls_per_unit`, `retry_rate`, `subscription_cost`, `infrastructure_cost`, `review_minutes`, `maintenance_hours`, and `setup_cost`.
Calculate three scenarios: cautious, base, and bad. Full cost = subscriptions + API + infrastructure + monitoring + manual review + support + error fixes. Cost per successful operation = all costs / operations that passed review.
AI cost = services + review + support + errors
Manual process cost = volume * minutes / 60 * rate + fixes
Payback = setup cost / net monthly savingsDecide by thresholds, not demos
Set conditions in advance: quality no worse than the manual baseline, critical errors = 0, manual review within the limit, cost per successful operation below current. In events store `automation_started`, `completed`, `failed`, `retry`, `human_review_requested`, and `human_review_approved` with `workflow_id`, `run_id`, `duration_ms`, `estimated_cost`, and `result_status`.
An example where the subscription misleads
Say 500 documents a month take 12 minutes each manually, and an employee hour costs 1,800 ₽. Manual baseline — 180,000 ₽. AI services cost 33,000 ₽, review — 30,000 ₽, infrastructure — 5,000 ₽, support — 12,000 ₽. Full result — 80,000 ₽, not 33,000 ₽. Net savings — 100,000 ₽ before setup.
If setup costs 360,000 ₽, simple payback is 3.6 months. But if 15% of operations need full rework, savings change. So in `Runs` record `result_status`, `review_minutes`, `retry_count`, and `correction_cost` for every run.
What to include in the cost of one operation
| Criterion | Question | Good sign |
|---|---|---|
| Input | What exactly enters full AI automation cost? | Describe run frequency, input, result, exception share, employee time, and required quality. |
| Action | What is the system allowed to do on its own? | Only pre-listed actions, with no access to the whole account |
| Verification | How do you know the result can be accepted? | Run a pilot of 100–300 operations and compare the manual baseline, cost, errors, and employee time. |
| Failure | Where does an unclear case go? | Mark unknown costs as a separate line and update the calculation after real load. |
What should change after setup
You get the cost of one successful operation and a payback period — not a pretty estimate based only on the plan.
Where automation suddenly costs more than people
Counting only the subscription.
Forgetting integrations, retries, and human control.
Automating a rare task where support costs more than the savings.
Calling a run that needed full manual rework a “success.”
When the calculation needs a financial model
You need a specialist if cost depends on several teams, complex infrastructure, or financial/legal error risk.
What to ask a vendor about price
How do you know automation pays back?
Compare monthly savings with full cost: setup, services, people, errors, and support.
Do you need to buy an expensive plan right away?
No. First limit volume, set caps, and check actual load.
Should you include employee salary?
Yes, if their time goes into review, fixes, and supporting the scenario.
Can you count only time savings?
As a first signal, yes — but separately check quality and business outcome.
