Why “automate everything” is the wrong target
A small business support team receives many repeat questions: where an order is, how to get a copy of an invoice, or when a pickup point opens. An assistant can answer these quickly from an approved source. In the same queue, however, are a disputed charge, a contractual disagreement, and a request for someone else's data. If the bot enthusiastically takes every case, an impressive automation percentage stops being a measure of savings.
The better question is: what share of cases should the assistant handle to minimize total cost and errors while preserving acceptable service? Handing a difficult case to a person is a planned outcome, not a system failure. Research calls this selective prediction: a model covers part of the workload and abstains when risk is higher. An AAAI study of such deferrals also shows why the combined human–AI team's performance must be evaluated, and why the way a deferral is communicated can affect the human decision. That study was not conducted in a Russian support center; it supplies a principle for evaluation, not a ready-made performance figure.
An illustrative calculation: two automation thresholds
Consider a **hypothetical** month and make every assumption explicit. There are 1,000 incoming cases. Fully manual handling costs RUB 180 per case, so the baseline is RUB 180,000. The assistant processes an incoming case and prepares a draft for RUB 12 whether or not its answer is sent automatically. Running the existing environment, updating the answer base and monitoring quality add RUB 18,000 per month. A case handed to a person still costs RUB 180 to handle. Additional spot checks cost RUB 90 each and cover 10% of auto-sent answers. Correcting a wrongly sent automatic answer costs an average of RUB 500. These are illustrative inputs, not market prices or the measured results of a real company.
At a cautious threshold, the assistant sends 60% of answers and hands 40% to people. Assume errors account for 3% of auto-sent answers in this scenario. The bill is RUB 12,000 for AI processing + RUB 18,000 fixed costs + RUB 72,000 for 400 manual cases + RUB 5,400 for 60 spot checks + RUB 9,000 for 18 corrections = **RUB 116,400**. The difference from fully manual handling is **RUB 63,600**, before tax and capital expenditure.
Now relax the threshold to 80% automated sends, leaving 200 manual cases. Suppose the harder set raises the error rate among automated answers to 12%: RUB 12,000 + 18,000 + 36,000 + 7,200 for 80 checks + 48,000 for 96 corrections = **RUB 121,200**. This is still below the manual baseline but RUB 4,800 more expensive than the cautious threshold. The robot has taken more cards from the queue, yet the finance team is not applauding: savings need not rise with automation coverage. Both error rates are deliberately assumed here; the business must measure them on its own cases.
Keep the cost formula next to the dashboard
At any candidate threshold, monthly cost in this simplified example equals: AI processing for all incoming cases + fixed operations + handed-off cases × manual handling cost + checked answers × audit cost + wrongly auto-sent answers × correction cost. Quality cannot be hidden in one average percentage: an incorrect opening time and an unauthorized disclosure of personal data have different consequences. The latter should not be reduced to RUB 500 at all.
Track constraints alongside currency: critical-error rates by category, response time, repeat contacts, customer satisfaction, and the maximum load on operators at peak hours. A threshold that lowers average cost but overloads the evening shift or sends more wrong refund decisions fails the test.
Design a useful handoff to a person
Do not begin with an “80% confidence” claim that the model wrote about itself. A 2026 ICML study found that calibration of verbalized LLM confidence depends on the task family and does not transfer universally between them. This does not mean every numerical signal is useless; it means thresholds must be calibrated separately on the company's task categories and the model's self-assessment must not be treated as proof of correctness.
An existing CRM or helpdesk can start with three outcomes: an automatic answer for a safe scenario, a draft requiring approval, and full handoff to a person. The router needs the case category, approved sources, order status, access rights, and a list of topics where auto-send is prohibited. Refunds, contractual commitments, personal data and contentious complaints are sensible candidates for mandatory approval at first. Pass the original request, retrieved evidence, proposed answer, and specific reason for escalation to the person, so they do not have to reconstruct the history. “I am transferring you to an expert” without the context makes the assistant a very polite forwarding desk.
No new service is necessarily required if the current ticketing system can assign queues and store status. Technical work remains: versioned knowledge sources, read permissions, decision logs, edge-case tests and a switch back to manual processing. The NIST AI Risk Management Framework calls for defined human-oversight roles, evaluation under conditions similar to deployment, and ongoing monitoring after launch. It is a general risk-management framework, not a certificate for any particular bot.
A modest pilot using real cases
Take, for example, 200 anonymized cases from the previous month, label the correct outcome and the cost of a mistake in each category. On a held-out subset, test several handoff rules and count not only the share of auto-sends but errors **among the answers sent**. Test high-impact categories separately; a large number of easy questions cannot offset a serious failure there. Then enable draft-only mode for a small team, measure actual operator time, and allow auto-send only for a narrow class of safe questions.
The pilot should produce a table of “threshold → automation share → errors → cost → peak queue”, not a single claim that “the bot closed 80%.” If actual review, correction and operating costs consume the savings, reduce coverage or keep the system as an operator copilot. Sometimes the most useful business AI is the one that promptly and clearly says: “a person should decide this.”