What do companies pay for, and what can AI take over?
According to Kaufman Lab data as of August 2026, of the 140 tasks that businesses pay for in job listings and on freelance marketplaces, only 10 are done by AI with no human in the loop. 89 are done with a human reviewing the result, and 41 stay with people. The ten have one thing in common: they are monitoring, classification or digest tasks, where a metric verifies the result.
- Sample
- 84 jobs: 32 from Russian and 52 from Western job listings, in marketing and sales and in the back office · 140 tasks copied verbatim from the listings · prices from 6 freelance marketplaces · 36 cards with an API availability check · 13 AI agents, snapshot of August 19, 2026
- So what
- The back office yields as many automation candidates as marketing: entering primary accounting documents, bank reconciliation, payroll checks, intake of requests. The price to beat is a freelance marketplace rate, not a salary: ad setup from 500 RUB (about $6), a reply to a review 500 RUB, a grant application 37,000 RUB (about $430).
Published Sep 2026 · data as of Aug 19, 2026 · collected by 13 AI agents: 7 market researchers, 6 card writers with API checks · reviewed by a person
How to find a task for AI in your own work: six steps
- List tasks, not job titles. You automate a single task, not a whole job.
- Mark the ones a number can verify. Monitoring, classification, digest: a metric accepts the result, no reading needed.
- Start with those. A task goes to the machine not when a model can do it, but when its result can be checked without reading it.
- Don't skip the back office. Primary documents, bank reconciliation and reporting produced as many candidates as marketing, with fewer competitors.
- Price against the marketplace. Compare with freelancer rates on marketplaces, not with an employee's salary: these are different markets and different money.
- Keep a human on review. Where the result is checked by reading, the machine prepares a draft, not a decision.
Where to start, by situation
- Your own work — start with one task you do every day.
- Client services — pick work where the pain has a price in rubles: reconciliation, reporting, primary documents.
- Team — measure the share of tasks that leave each person's desk, not “replacing a person”.
Below is the registry itself: 84 jobs, 140 tasks and a score for each on becoming a microservice.
What we found: 140 tasks from job listings and what AI does with them
What a number can verify goes fully to the machine. Ten level A tasks: tender monitoring, request routing, search rank monitoring, reputation monitoring, contact enrichment, checking an email campaign before sending, meeting minutes, transcription, subtitles, translation. They share one pattern (monitoring, classification, digest), and each has a metric that accepts the result without anyone reading it.
The back office is as promising as marketing. The finance and operations clusters produced as many candidates with priority 36–48 as marketing did: entering primary documents into 1C (Russia's standard accounting software), bank reconciliation, management reporting, payroll checks, intake and routing of requests. Yet the Russian market has only a handful of ready products for them, and the cost of an error is measured in rubles.
The marketplace sets the price, not the salary. Setting up a campaign in Yandex Direct, the ad platform of Russia's largest search engine, costs from 500 RUB (about $6), a review reply 500 RUB, CRM integration 1,000–4,500 RUB (about $12–53), a turnkey chatbot 4,000–6,000 RUB (about $47–70), a grant application 37,000 RUB (about $430). A service that does a task for a fixed fee competes with these numbers, not with a median salary.
What we tested: five hypotheses set before data collection
Three decisions were made before the build. The catalog unit is a task, not a job: “running social media” is not a unit, “draft a content plan for the month” is. The automation bar is strict: a task gets level A or B only if a microservice with a checkable result can be built for it from a script, an API and a prompt. “A person with ChatGPT can handle it” is level C. Depth beats coverage: cards are written only for the top of the ranking, scored as demand × price × automatability, each factor from 1 to 5.
How we counted: 13 AI agents and three decisions made before the build
Seven agents worked in parallel. They collected Russian and Western job listings in two segments, marketing with sales and back office, plus price lists from six freelance marketplaces; the seventh looked for new Western jobs that have not yet become separate roles in Russia. Collection rules: tasks are copied from listings verbatim, no number is used without a source URL, pre-tax and take-home pay are labeled and never mixed, and only horizontal roles where information work makes up at least half the job are included. The tasks from different listings and languages were then merged into 140 units, each scored on three factors, and six second-wave agents wrote cards for the top 36, spot-checking the key APIs.
| Lab practice | What it gave this study | Canonical name · source |
|---|---|---|
| The catalog unit is a task, not a job | 140 comparable units instead of 84 incomparable roles | Task-based view of employmentAutor, Levy & Murnane, QJE, 2003 |
| A–D bar by how verifiable the result is | Levels A and B go only to tasks with an acceptance metric | Task-level suitability for machine learningBrynjolfsson, Mitchell & Rock, AEA P&P, 2018 |
| Tasks copied verbatim from listings | The market's own wording, not the lab's hypotheses | Content analysis of job listingsKrippendorff, “Content Analysis”, 2004 |
| Priority as the product of three scores | The top of the ranking gets a card, the rest stays in the registry | Multi-criteria evaluationKeeney & Raiffa, “Decisions with Multiple Objectives”, 1976 |
| API checked against documentation | 36 cards, each with a fact such as “available, free, on request” or “not available” | Primary source verificationthe lab's own rule |
Results: four views of the data
What to trust: limits of the registry
- Sample
- 84 roles from listings that agents found through search, not an API export from job sites. Roles where information work is less than half the job, and offline tasks, were filtered out at intake, so level D is empty by design.
- Variance
- Salaries come from aggregators and live listings, labeled “pre-tax”, “take-home” or “unknown”; they cannot be added into one total. Demand and price scores are expert estimates based on visible demand, not a count of listings.
- Does not generalize
- The registry covers horizontal roles in small and medium businesses: marketing, sales, finance, operations, documents. Engineering, medical and manufacturing jobs were not collected.
- Not measured
- 36 of the 140 tasks have cards. The API check is a spot check, 1–3 facts per card; before building a specific service, its whole stack needs a full review. A service actually built and sold outweighs any estimate and would change the scoring.
- Scale 1–5
- Demand: the number of roles that include the task and how often it appears in listings. Price: marketplace rates and share of salary. Automatability: whether input, processing, output and the check can be described; 5 means fully, 1 means speed-up only.
What to do with it: start with what a number can verify
Start with tasks where a metric accepts the result without anyone reading it: monitoring, classification, digest. Set the price against marketplace rates, not against a salary. The back office has fewer competitors, and its pain has a price in rubles: primary documents, bank reconciliation, payroll, requests.
A task goes to the machine not when a model can do it, but when its result can be checked without reading it.
Common questions about automating tasks with AI
What is a task registry for AI microservices?
A task registry is a catalog of 140 tasks copied verbatim from job listings and freelance marketplace price lists in Russia and the West, scored on three factors: demand, price, automatability. Each task gets a level from A, where AI does it with no human, to C, where AI only speeds it up. Compiled by Kaufman Lab with data as of August 19, 2026; the top 36 tasks are broken down into cards with an API availability check.
Which tasks does AI already do without a human?
In the registry, 10 of 140: government tender monitoring matched to a company profile, routing of incoming requests, search rank monitoring, review and mention monitoring, enriching and cleaning a contact list, checking an email campaign before sending, meeting minutes, transcription, subtitles and translation. They have one thing in common: a metric accepts the result instead of a person reading it, which is why the human leaves the loop here first.
What business work can't AI automate?
In the registry, 41 of 140 tasks stayed at level C: AI speeds up the work, a human does the core. These are negotiations, sales conversations, strategy, decisions about money, process interviews, publishing a reply to a negative review, submitting applications and defending them before a committee. Tasks with no information core at all did not enter the registry: physical work, personal relationships, accountability that comes with a signature.
How much does it cost to replace an employee with an AI agent?
Compare with the marketplace price for the task, not with a salary, because that price sets the ceiling: ad setup from 500 RUB (about $6), a review reply 500 RUB, CRM integration 1,000–4,500 RUB (about $12–53), a turnkey chatbot 4,000–6,000 RUB (about $47–70), a grant application 37,000 RUB (about $430). A microservice pays off when it takes a task for a fixed fee below these numbers or handles a steady flow of work on a subscription: for example, tender monitoring at 5,000–15,000 RUB a month (about $58–175) versus the 20,000–30,000 RUB (about $235–350) share of a specialist's salary.
Which jobs is AI changing most in 2026?
The pressure already shows in prices: copywriting, where the most common salary is 50,000 RUB (about $580), lost 4.6% in a month, first-line support lost 4.4%, and filling in product listings for online marketplaces pays 35,000–45,000 RUB (about $410–525). Meanwhile the West is creating new roles: GEO specialist at $65,000–118,000 a year, Automation Specialist at $91,000–165,000 a year, GTM Engineer and RevOps. In Russian job listings these functions are still lines inside SEO and marketing jobs.
What are automation levels A, B, C and D?
A: the task is done end to end today with no human in the loop, and a metric verifies the result. B: it is done, with a human only reviewing the result. C: AI speeds up the work, the core stays with a human. D: cannot be digitized (physical work, trust, responsibility). Level A or B goes only to a task for which a microservice can be built from a script, an API and a prompt, with a describable input, processing, output and check. “A person with ChatGPT can handle it” is level C.
Are there free APIs for automating small business tasks?
From the check of 36 cards: the government procurement portal provides data free through a SOAP service with a token from Gosuslugi, the Russian state services portal, and its FTP has been closed since 2025. OData REST is a standard free feature of the 1C platform. The Yandex Direct API is free on request. Search Console and the Yandex Webmaster API are free. Apollo gives 900 credits a year. Yandex Maps has no official reviews API, only automation of the business account; 2GIS, a Russian maps and business directory service, has a Reviews API with a trial period.
- hh.ru, GorodRabot, Dream Job, rabota.ru [Russian job sites]. Russian job listings, marketing and sales, back office, snapshot of August 19, 2026. hh.ru
- Indeed, ZipRecruiter, Glassdoor, LinkedIn. Western job listings in the same segments, plus new roles, snapshot of August 19, 2026. ziprecruiter.com
- Kwork, Profi, YouDo, FL.ru, Upwork, Fiverr [freelance marketplaces; the first four are Russian]. Per-task price lists, snapshot of August 19, 2026. kwork.ru
- Government procurement portal [in Russian]. Data service with access via a Gosuslugi token; walkthrough on Habr, a Russian tech community. habr.com/ru/articles/869934
- 1C [in Russian]. Standard OData interface of the 8.3 platform. its.1c.ru
- Yandex [in Russian]. Direct API: connection and use are free. yandex.ru/support/direct
- Yandex [in Russian]. Maps API Club: a reviews API is not implemented. yandex.ru/blog/mapsapi
- 2GIS [in Russian]. Reviews API documentation. docs.2gis.com
- Apollo.io. 2026 pricing: free plan with 900 credits a year. hackingdemand.com
- Entera [in Russian]. Pricing for primary document recognition. entera.pro/cost
- Granted AI. Pricing for AI grant writing. grantedai.com/pricing
- Autor D. H., Levy F., Murnane R. J. The Skill Content of Recent Technological Change. Quarterly Journal of Economics, 118(4), 2003. doi:10.1162/003355303322552801
- Brynjolfsson E., Mitchell T., Rock D. What Can Machines Learn, and What Does It Mean for Occupations and the Economy? AEA Papers and Proceedings, 108, 2018. doi:10.1257/pandp.20181019
- Krippendorff K. Content Analysis: An Introduction to Its Methodology. Sage, 2004.