What is the first thing to automate in business and how to calculate it?

Briefly

• It is not enough to introduce AI or a robot, you still need to get an effect from it. In the 2025 McKinsey survey, 88% of companies use AI, and only 39% see the impact on profits.
• The main reason for the failures, according to the RAND interview, is a poorly formulated task, not a weak technology. Therefore, you need to start with a list of routine and measurements, and not with the choice of a tool.
• The ICE, RICE, WSJF methods and the "value — complexity" matrix solve one problem: to compare the benefits and costs on the same scale. For a business owner, a simplified version is enough.
• We offer our simplified formula (this is the author's SystemsLab scheme, not the standard): (frequency + lost hours + money and errors) * Staff readiness * complexity. Whoever scored more points goes to the pilot first.
• It takes two weeks to collect data and select one or two tasks for the pilot. The order of actions by day is below.

Automation is the transfer of repetitive work to a program: an accounting system, a robot, a bot, or an AI agent. Let's look at why projects don't pay off, what prioritization techniques are available, and how to get your task rating in two weeks. The data is current as of September 2026.

Why do many automation and AI projects not pay off?

Most often because they chose the wrong task or could not measure the result. Figures about failures are often retold inaccurately, so we look at the primary sources.

MIT NANDA: 95% of organizations without measurable returns

The MIT NANDA report "The GenAI Divide" (July 2025) says: despite $30-40 billion of corporate investments in generative AI, 95% of organizations do not receive measurable returns from it. Generative AI are models that create text, images, or code on demand, such as ChatGPT or GigaChat.

This figure is often retold as "95% of AI projects fail", but this is inaccurate. The report is marked as preliminary (v0.1, "Preliminary Findings"). It is based on a review of more than 300 public initiatives, interviews with representatives of 52 organizations and 153 questionnaires of managers collected at industry conferences. The authors themselves call their figures " correct in direction ", and not accurate. The correct wording sounds like this: in the MIT NANDA sample, 95% of organizations do not see a measurable effect of generative AI on profits.

According to one estimate in the report, about half of generative AI budgets go into sales and marketing, although, according to the authors, back-office automation often pays off better. The back office is an internal work that the client does not see: reconciliation, documents, accounting. The most noticeable savings were given by the rejection of a part of outsourcing and agency services, rather than reducing its own staff.

McKinsey: almost everyone uses it, less than half see the effect

According to McKinsey's "The State of AI in 2025" (November 2025), 88% of respondents regularly use AI in at least one company function. But at least some impact of AI on operating profit (EBIT) is noted by only 39%, and for most of them AI accounts for less than 5% of this profit. Companies-"leaders", where the effect is noticeable, about 6%. What distinguishes them is that they often rebuild the workflows themselves, and not just add a new tool to them.

Disclaimer: 1,993 people from 105 countries and companies of various sizes were interviewed, 38% of them work in organizations with revenues of more than $ 1 billion. Therefore, it is impossible to transfer these shares to Russian small businesses directly. The general conclusion is transferred: the instrument itself does not bring profit.

Gartner: Forecasts, not measurements

Gartner, in a press release dated June 25, 2025, predicts that more than 40% of agent-based AI projects will be canceled by the end of 2027. The reasons are rising costs, unclear business value and weak risk control. Agent—based AI are systems that not only answer a question, but perform a chain of actions themselves: create an order, write to a client, change an entry in the database.

A year earlier, on July 29, 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after the pilot by the end of 2025. Both numbers are predictions for specific types of AI, not measurements. We have not found verified data on whether the forecast about 30% has come true as of September 2026, and it is generally incorrect to transfer these percentages to automation.

RAND: Why projects fail

In the RAND report "The Root Causes of Failure for Artificial Intelligence Projects" (2024) there is a well-known phrase: according to some estimates cited by RAND, more than 80% of AI projects fail. These are other people's estimates from the press, RAND himself did not measure the share of failures. RAND's own work consists of interviews with experienced data scientists and engineers, and its value lies elsewhere: the authors identified five root causes of failures.

1. The task is misunderstood or poorly formulated. This is the most common reason.
2. There is no necessary data.
3. The team got carried away with a fashionable technology, not a user problem.
4. There is not enough infrastructure, that is, servers, integrations and processes around the model.
5. The task is too difficult for AI.

Three reasons out of five are eliminated even before the start, if you honestly evaluate the task.

This is not just about AI

Routine Robotization (RPA) is a "robot" program that repeats human actions in interfaces: copy data from one window to another, fill out forms. According to the experience of EY ("Get ready for robots", 2016), up to 30-50% of the first RPA projects fail, and it's not about technology, but about typical implementation errors.

Where is the effect still there? Stanford AI Index 2026 provides data on productivity gains of 14-15% in customer support and 26% in software development. These tasks have one thing in common: the work is structured, and the result is easy to measure.

What is known about Russian business?

Many people use neural networks, but few are integrating AI into processes yet. According to Rosstat, 4.8% of organizations use AI in processing the HSE IIEZ (March 2026): 14.9% among companies with more than 500 people and 4.1% among companies with up to 100 people. The survey covers large and medium-sized organizations, small businesses are not included in it.

For small businesses, the picture is different. According to a survey by PSB and "Opora Russia", which "Delovaya Rossiya" retells with reference to "Izvestia" (August 2026), the share of entrepreneurs using neural networks has grown from 17% to 53% over the year. There is no contradiction here: this survey takes into account any use of the neural network, including the generation of texts in a chatbot, and not the built-in automation of the process.

Back in 2024, NAFI found out that 63% of small and medium-sized businesses see the main benefit of AI in saving time. More than half of those who agree that AI saves time, call 3-10 hours a week.

We have not found any verified fresh Russian data on how many hours employees spend on routine, so we will have to measure the hours at home.

What techniques help you choose what to automate first?

The most famous are ICE, RICE, WSJF and the "value — complexity" matrix. They all do the same thing: they translate " I want" and " it seems" into numbers that can be compared with each other.

ICE: Fast, but subjective

ICE is an assessment of three factors: Impact, Confidence, and Ease. The technique is associated with the name of Sean Ellis. According to the description of ProductPlan, each factor is put on a scale from 1 to 10 and usually multiplied. The weaknesses of the methodology are known: estimates are almost entirely subjective, and ease pulls up " quick wins", even if they are of little use.

RICE: takes into account how many people it concerns

RICE was described by Sean McBride of Intercom in an article published on January 5, 2018. The formula: coverage × influence × confidence - labor costs. The impact is assessed on a scale 3 / 2 / 1 / 0,5 / 0,25, Confidence is in percentages (100, 80 or 50), labor costs are in man-months. The main benefit of RICE is that confidence is a separate multiplier here: a bold but untested idea honestly loses points.

WSJF: how much do you lose until the task is done

WSJF (weighted priority "short tasks first") is used in the Scaled Agile Framework methodology. The formula is: the cost of the delay is the size of the work. The cost of delay is what the business loses for each period until the task is solved. It consists of value for the client and business, urgency and risk reduction. The SAFe page quotes Don Reinertsen: if you measure one thing, measure the cost of the delay.

The "value — complexity" matrix

This is a 2x2 table: benefits on one axis, complexity on the other. Tasks are divided into four quadrants: "quick wins" (lots of benefits, little difficulty), "big projects", "little things" and " traps". The matrix does not have a single author, it is a working tool from the practice of product management. It is visual, but two tasks in the same quadrant cannot be compared with its help.

Methodology	What counts	A strong side	The weak side
ICE	Influence × confidence × ease	Five minutes per task	A lot of subjectivity, likes "quick wins"
RICE	Coverage × impact × confidence - Labor costs	Penalizes untested ideas	It is necessary to estimate labor costs in man-months
WSJF	The cost of delay is the size of the work	He thinks about the money that is lost every month	The cost of the delay is difficult to estimate without data
The matrix is 2×2	Value vs. Complexity	Clearly, it is clear to the whole team	Does not rank tasks within the quadrant

These techniques came from IT product management, and there is a lot of excess in them for the head of a fitness club or workshop. But from each one you can take a useful idea: from RICE — coverage and confidence, from WSJF — money that is lost every month, from the matrix — visibility. This is how the formula below works.

And if the data is already in the systems?

Process-mining is an analysis of event logs from accounting systems, which shows how the process actually goes, and not as it is described in the regulations. This is how the manifesto of the IEEE process mining working group defines it (there is an official translation into Russian). A small company usually has a simplified version: uploads from the accounting system with the number of documents, edits, and refunds per day.

How to calculate which task to automate first?

Evaluate each routine task according to five criteria from 1 to 5 and calculate the score using a simple formula:

1. Frequency (H). How often the task is repeated. The more often, the more benefits of automation.
2. Lost hours (N). How many working hours per month are spent on a task for all employees together. Count by measurement, not by feeling.
3. Money and mistakes (E). What errors, delays and missed sales cost: incorrect checkout reconciliation, forgotten subscription renewal, marriage due to an error in the task.
4. Staff readiness (D). Whether people are ready to work in a new way and whether there is a person who will be responsible for the result. This is a multiplier: without readiness, even a good idea loses most of the points.
5. Complexity (C). How difficult is it to do: whether integrations are needed, in what state the data is, whether the accounting system has a software interface. This is the divisor.

Score = (H + N + D) × G - S. This is our author's formula, not a generally accepted methodology: it simplifies the ideas of RICE and WSJF for the business owner, and the scales can be adjusted to suit themselves. Maximum — 75 points, minimum — less than one. We add value because its parts complement each other. Readiness multiplies the result, and complexity divides — this is how the formula repeats the logic of RICE and WSJF: benefits in the numerator, costs and risks in the denominator.

Criteria	1 point	3 points	5 points
Frequency	Once a month and less often	Several times a week	Many times a day
Lost hours per month	Up to 5 hours	10-20 hours	More than 40 hours
Money and mistakes	Mistakes are rare and cost nothing	Noticeable losses once a month	Regular loss of money or customers
Staff readiness	Resistance, there is no responsible	Neutrally, there will be a responsible person	Employees ask for it themselves, a responsible person has been appointed
Complexity	Ready-made function or setting	Need a revision or a bundle of two systems	No data, many systems, non-standard logic

Intermediate grades are 2 and 4. The following scale is convenient for watches: 2 points — 5-10 hours, 4 points — 20-40 hours.

Money should be counted separately, even if they are included in the score with an estimate. Multiply the lost hours by the cost of the employee's hour, along with taxes, and add direct losses from errors. It turns out the amount that the task "eats" every month, that is, the same delay cost from WSJF. It is easier to compare contractor offers with this amount.

What does the evaluation table look like in practice?

This is one table where each task has five grades and a final score. Below is a template with five examples from a fitness club, a thermal spa complex and a production facility.

Important: the examples are illustrative. The scores are conditional and are set to demonstrate the formula, this is not the data of our clients. In your business, the same tasks may get completely different estimates.

Task	Where	H	P	D	D	With	Score
Reminders about the end of the subscription and freezing: the administrator calls customers manually	Fitness Club	5	4	4	4	2	26
Answers to typical questions of guests: working hours, prices, what to take with you	Thermal spa complex	5	4	2	4	2	22
Reconciliation of payments by bank cards (acquiring) with the accounting system and 1C at the end of the shift	Fitness club, thermal baths	4	3	5	3	3	12
Transfer of paper changeable reports to 1C	Production	5	5	4	2	3	9,3
An AI agent that predicts the purchase of raw materials	Production	2	2	5	2	5	3,6

H — frequency, P — lost hours, D — money and mistakes, D — staff readiness, C — complexity. Example of calculation for the first line: (5 + 4 + 4) × 4 ÷ 2 = 26.

What can be seen from the example

• Reminders and responses to guests come forward. They are frequent, understandable, and the result is easy to measure: how many extensions, how many issues are closed without an administrator. This is a structured job for which the Stanford AI Index records a noticeable increase in productivity. Often there is enough bot in the messenger (MAX, Telegram) or on the site associated with the account system. As of September 2026, the main channel is MAX and the website: according to Mediascope in the retelling of Habr, in June 2026, the monthly audience of MAX is 86.26 million people, Telegram is 75.69 million. Telegram has been officially slowed down by Roskomnadzor since February 10, 2026, so it is suitable as an additional channel, and WhatsApp has not been working without a VPN since February 11, 2026.
• Reconciliation of payments is expensive for errors, but more complicated. It is necessary to link the bank, the club's accounting system and 1C. Such a task should be done a second time when the team has already received the first result.
• Paper reports lose points due to readiness. It takes a lot of hours, but if the masters are not ready to give up paper, automation will be sabotaged. First you need a responsible and understandable input form, for example, a tablet in the shop.
• The fashionable AI forecast turns out to be the last. The benefit may be high, but the task is rare, there may not be enough data, and the complexity is maximum. Three reasons for RAND failures converge here: a lack of data, an emphasis on fashionable technology, and a task that is too difficult for AI.

How to choose the first task for automation in two weeks?

It takes 10 working days: a week to collect data, a week to evaluate and train the pilot. A pilot is a small, time—limited launch on one site to test the effect before high costs.

1. Day 1. Make a routine list. Ask the administrators, accountant, managers and foremen to write down everything they do with their hands more than twice a week. Hints: "rewriting", "checking", "reminding", "answering the same". In the evening, remove the duplicates and formulate each task in one phrase with the verb: "verify purchase payments with the cashier". Usually it turns out 15-30 points.
2. Days 2-6. Measure the time. For five working days, employees mark in a simple table how many minutes it took for each task from the list. In parallel, make an upload from the accounting system: how many documents, entries, edits and cancellations per day. This is a simplified process-mining.
3. Day 7. Count money and mistakes. Convert hours to a month and to rubles at the cost of an hour with taxes. Bring up cash discrepancies, refunds, complaints, marriage, and unpaid renewals over the last quarter.
4. Day 8. Assess readiness and difficulty. Talk to those who will be affected by the changes and find a candidate responsible for each task. Complexity check with your IT specialist or integrator: which systems are involved and whether they have a software interface for data exchange.
5. Day 9. Count the points and select one or two tasks. Fill in the table, sort it in descending order, and check the top of the list with common sense: if the task disappears by itself after six months, you can skip it. Take no more than two tasks to work so that the team does not get scattered.
6. Day 10. Formulate the pilot's task and plan. Write down, as it is now (the numbers from the measurement), what result is needed, by what metric you will check it and at what result the pilot stops. Assign the responsible person, the site, the time and date of the control measurement.

Choose the solution method after the task is formulated, not before it. Sometimes it is enough to set up an accounting system or refine 1C, sometimes you need a bot, robot or AI agent. After the pilot, go back to the table: the real numbers will show whether the solution is worth scaling up and which task is next.

What errors most often spoil the calculation?

Most often, the calculation is spoiled by estimates " by eye" and haste with the choice of a tool. Check yourself on the list before approving the pilot.

• The clock is taken from the head. The manager usually underestimates the small tasks that eat up the day for 5 minutes. Microsoft data (June 2025) shows how working hours are divided: according to Microsoft 365 telemetry, an employee is interrupted on average once every 2 minutes. Russia is not included in this sample, but the effect itself is familiar to any administrator at the reception.
• "Work about work" is confused with routine. In the Asana study (March 2023), coordination, information retrieval and statuses occupy 58% of the working day. But this is not the same as the tasks that can be given to the program. 9,615 office workers from six developed countries were interviewed, Russia is not among them.
• There is no responsible person. According to the Microsoft Work Trend Index 2026, organizational factors — culture, executive support, personnel practices — affect the effect of AI twice as much as the personal efforts of an employee. The survey was conducted in 10 countries, Russia is not among them. If no one is responsible for the task, set readiness 1.
• They take five tasks at once. The team is dispersed, and not a single task reaches a measurable result.
• There is no " to" measurement. Without the initial figures in three months, it is impossible to prove that automation has given something. And in the samples of MIT NANDA and McKinsey, most organizations just do not see a measurable effect.
• Choose a technology before the task. This is one of the five reasons for RAND failures. The wording "we want an AI agent" is not a task. The task sounds like this: " reduce the response time to the guest from 20 minutes to 2".

When should I call an integrator?

When two or more systems are involved in the task, or when an assessment of complexity from the outside is needed. A list of routine, time measurement, and readiness assessment is best done on your own: you know your business better than any contractor.

An external look is useful in the steps where technical knowledge is needed: evaluate the complexity, find a ready-made function instead of development, link the club's accounting system, bank and 1C. This is how our system integration works: first, audit and calculation, then pilot. If there are answers to guests or reminders at the top of the table, look at how we build AI agents for staff and guests. The procedure from audit to support is described on the "How we work" page.

Sources

1. MIT NANDA. The GenAI Divide: State of AI in Business 2025 (preliminary version, PDF copy) — July 2025.
2. McKinsey. The State of AI in 2025: Agents, innovation, and transformation — November 2025.
3. Gartner. Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — 25.06.2025.
4. Gartner. 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025 — 29.07.2024.
5. RAND. The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed — 2024.
6. EY. Get ready for robots — 2016.
7. Stanford HAI. AI Index Report 2026, Head of Economy — 2026.
8. ISIEZ NIU HSE. The use of AI in organizations according to Rosstat — March 2026.
9. "Business Russia" (reprint of "Izvestia"). Survey of the PSB and "Supports of Russia" on the use of neural networks in SMEs — 08/03/2026.
10. NAFI. Every third representative of SMEs uses artificial intelligence in their work — 05/20/2024.
11. ProductPlan. ICE Scoring Model — no date specified.
12. Intercom. RICE: Simple prioritization for product managers — 05.01.2018.
13. Scaled Agile Framework. WSJF — no date specified (page © 2010-2026).
14. IEEE Task Force on Process Mining. Manifesto of process mining (Russian translation) — 2012.
15. Microsoft WorkLab. Breaking down the infinite workday — 17.06.2025.
16. Asana / Business Wire. Anatomy of Work Global Index 2023 — 08.03.2023.
17. Microsoft. Work Trend Index 2026 — May 2026.
18. MAX bypassed Telegram by monthly audience (Mediascope data for June 2026) — Habr, 07/22/2026.
19. Telegram blocking in Russia (2026) — Wikipedia, verified 09/17/2026.
20. The authorities finally nailed WhatsApp in Russia — CNews, 02/11/2026.