The phrase «we’ll save 30% of the time» shows up in every other presentation about enterprise AI. In our assessments we ask to see the math – and in most cases there is no math. There’s a feeling. Then a budget gets signed off on the strength of that feeling, six months later the CFO asks where the promised millions are, and there’s nothing to say. The project may well be working fine – the effect was simply calculated in a way that makes it impossible to verify. Below are the three mistakes we see most often, and the model we use ourselves.
Mistake one: percentages instead of hours
«We’ll free up 30% of the department’s time» sounds weighty and means nothing. Thirty percent of what? Of a workday that’s half meetings? Of the time spent on an operation the agent only partly takes over? A percentage can’t be verified – which means it can’t be defended in front of a finance person.
The honest path is more boring: time the specific operation. Take one operation, measure the minutes, multiply by frequency and by the cost of an hour. An example from our assessments: logging call outcomes into the CRM. A manager makes 20 calls a day and spends 6 minutes on the card after each one: 2 hours a day. A department of 8 managers: 352 hours a month.
6 min × 20 calls = 2 h/day → 8 managers × 22 days = 352 h/month × 2,800 ₸ ≈ 985k ₸ per month
Now the honest correction. The agent parses the call and fills in the card itself, but the manager spends a minute and a half to two minutes checking it. You save not 6 minutes but roughly four. The result: not 985 but about 690k ₸ per month. Less than in the presentation. But this figure survives any scrutiny: every multiplier can be produced on demand.
Mistake two: counting only the time saved
Hours saved are the first term of the effect, and often not the main one. The second is the lift in the outcome of the process itself. The agent replies to a lead in five minutes instead of four hours – conversion to a first meeting rises. A claim for the fund is checked before submission – fewer line items get struck from payment. A claim to a contractor goes out on time – the penalty doesn’t expire.
Let’s continue the sales department example. The department closes 30 deals a month with an average margin of 400k ₸. The speed of first contact, and the fact that no lead is lost in manual entry, give a conservative plus 10% to conversion – three extra deals, 1.2M ₸ of margin per month. More than all the time savings from the first term. How this works in a live scenario, we showed on the page for the sales department solution.
In our assessments we see this constantly: in half of the scenarios the lift in outcome brings more money than the hours saved – and it’s precisely that lift that most often isn’t counted at all. In healthcare the second term is overturned fund denials, in procurement it’s risks caught before signing, in support it’s retained customers.
Mistake three: the effect in a vacuum
The third mistake is to count the benefit and forget the costs. From the effect you have to subtract the cost of the subscription, staff time on control points, and the transition period: for the first one or two months the agent learns on your examples, people get used to it, and the savings don’t plateau right away. If the model doesn’t include this, it will shatter against the very first quarterly report.
And a sober note about the hours themselves: an hour saved turns into money only when it’s directed somewhere – toward higher output, cutting overtime, dropping a planned hire. If the freed-up time dissolves over the course of the day, the calculation stays pretty and the effect stays on paper. That’s why every scenario needs a metric owner: a person who decides where the freed-up 352 hours go, and answers for them.
An honest model: two terms, one subtraction
Effect = (hours × cost of an hour) + Δ process outcome − control and subscription costs
Let’s assemble our example in full. Time saved: 690k ₸. Lift in outcome: 1.2M ₸. Checking cards and reviewing the agent’s errors: minus 150k ₸. Net effect: about 1.75M ₸ per month. A platform subscription from 12M ₸ a year is a million a month. The annual subscription is covered by the accumulated effect in about 7 months – that’s the very 6–9 month range we quote in the demo.
| Term | How to measure it | Where the data comes from |
|---|---|---|
| Time saved | minutes per operation × frequency × hourly rate | time studies, exports from the CRM and telephony |
| Lift in outcome | shift in conversion, denials, penalties in money | current process statistics over 3–6 months |
| Subtractions | subscription, time on control points | contract, measurement on the prototype |
Every row is verifiable. That’s the main property of an honest calculation: it doesn’t have to be big, it has to withstand questions.
Mini-scenario: two calculations for one department
A company is evaluating the automation of customer support. The naive calculation from the presentation: «operators spend 60% of their time on routine questions, we’ll cut costs by 60%» – with a payroll of 6M ₸ a month, that promises 3.6M in savings. The CFO didn’t accept that calculation, and rightly so.
The honest recalculation based on time studies: routine requests are 55% of the flow, the agent closes them without a human, but 15% of dialogues go to escalation and spot quality control. The savings: about 1.9M ₸ per month. Plus the second term: replies at night and on weekends, which previously didn’t exist at all – fewer lost requests. The final figure is almost half of the naive one, but it became the basis of the contract and matched reality six months later. We broke down similar mechanics in the customer support solution.
How to check a calculation in ten minutes
- Every number traces back to a measurement or an export, not to «that’s how it’s usually counted».
- Both terms are present: hours and process outcome. If the second is missing, the calculation is incomplete.
- Subtractions are present: subscription, control, transition period.
- A metric owner is named – the person responsible for turning hours into money.
If the model passes all four points, you can show it without fear to either the CFO or the board of directors. You can test the terms in practice within a week: we set up a prototype on your data for free, and a measurement on it is more honest than any presentation.
Frequently asked questions
What cost of an hour should I use in the calculation?
The full one: salary with taxes plus the workplace, divided by actual working hours. For office roles in Kazakhstan, in our assessments we usually take 2,500–4,000 ₸ per hour. A bare salary understates the effect, and «revenue per employee» overstates it.
How do I estimate the lift in outcome before implementation?
Take the current process statistics – conversion, denial rate, total penalties – and assume a conservative shift of 10–15%. Then test the hypothesis on a prototype: a week, your data, free. After the run, the model contains measurements rather than assumptions.
What do I do if the effect doesn’t cover the subscription?
Don’t implement that scenario. Seriously: a negative result of the calculation is also a result, it saves 12M ₸ a year. Usually it means the process was poorly chosen – a rare operation, a cheap hour, no second term. Take an adjacent process and recalculate: we published the selection checklist in the article on choosing the first process.