AI
AI agents for business
In business, AI shows up in financial reports where agents run processes inside the business systems the company already uses, not where employees were handed chat assistants. An AI agent is a program that performs process steps on its own inside those systems and passes only disputed cases to a human. An exception is an operation that failed a rule check: a supplier invoice does not match the order, the delivery address is wrong, or there is no stock.
Companies first hand to agents repetitive processes that involve cross-checking data between systems and have clearly defined exceptions. Start with one department, chosen by operation volume, cost, and the number of neighbouring departments involved; most often this is finance. Then record baseline metrics and run the agent in shadow mode: for the first two weeks it makes decisions in parallel with a human without the right to act, then on a schedule it gradually earns the right to execute part of the cases. This article maps those processes across nine departments and adds adjustments for a company operating in Uzbekistan.
Why a chat-assistant subscription does not show up in reports
McKinsey's survey, The state of AI in 2026, conducted in May-June 2026 among 1,719 respondents, records a gap between perception and numbers: 8 in 10 respondents say AI improved their personal productivity, while 37% credit AI with a contribution to earnings before interest and taxes (EBIT), almost the same share as a year earlier.
The cause of the gap is the nature of the tool. Vas, CEO of Varick Agents, a company that sells and implements AI agents itself, puts it this way: a chat assistant speeds up a person by 10-20% when they think to ask. If 1,200 of 12,000 employees work 15% faster for a third of the working day, that is about 24 minutes a day for each of those 1,200 people. Minutes do not turn into a smaller payroll: the work and the headcount stay the same, so reports do not show it.
Varick reports that background agents that pass to people only exceptions deliver 60-90% efficiency gains. Vas gives an illustration: in a 40-person accounts payable department where agents take on 75% of the work, that is 30 positions fewer. We recommend deciding what to do with freed-up positions before launch. The difference between a chat assistant and an agent working inside business systems is covered in a separate article.
Which processes to hand to AI agents first: a map across nine departments
The figures below are Vas's observations from Varick's large clients, mostly American; they are a vendor's observations, not an industry study. The selection logic applies to any company; your own numbers must be measured before launch, because a business of 100-500 people will have different ones.
| Department | Process | What the agent does | What stays with a human | What Varick sees |
|---|---|---|---|---|
| Finance | Exceptions in accounts payable (invoice does not match order or receipt); month-end close (bank reconciliations, accruals, intercompany settlements in the background); posting customer payments and working accounts receivable | Cross-checks documents, runs reconciliations and accruals in the background, posts payments, sends reminders for overdue receivables | Handles exceptions where amounts or details did not match | Before: 600-800 exceptions per month handled manually; month-end close 18-22 days. After: 7-9 days |
| Sales ops | CRM updates; preparing data for the weekly forecast meeting; approving a discount or price | Updates deal records, gathers data, starts the approval chain | Negotiates, handles non-standard terms | Before: sales reps spend 38% of their time on CRM; discount approval takes 23 days |
| HR | Cross-checking HR data against payroll before the run; routine questions; onboarding from offer letter to first day | Checks data before payroll runs, answers questions about leave and benefits, walks a new hire through the checklist | Handles non-standard HR situations | Up to 80% of routine employee questions can be resolved without a person; most payroll corrections stem from errors before the calculation runs |
| Legal | First edit of standard contracts; intake of requests; tracking renewal dates and auto-renewals | Edits NDAs, framework agreements, and supplier contracts per internal editing rules; accepts a request; writes out renewal and notice dates | Edits non-standard contracts, negotiates | Before: 5 or more days for the first edit. After: same day |
| Operations | Investigating a supply disruption (delay, supplier problem); order exceptions (customer on credit hold, wrong address, no stock) | Checks credit limit, address, stock, runs the disruption investigation | Manages the emergency, finds a replacement | Before: a week for disruption investigation. After: a few hours |
| Procurement | Onboarding a new supplier; three-way match of invoice, purchase order, and goods receipt | Collects documents, details, checks; matches three documents, catches duplicates and prices above the ordered amount | Negotiates with the supplier | Before: 26 days for supplier onboarding. After: 4 days |
| IT | Granting and revoking access; routing IT tickets | Grants access to a new hire on day one, revokes it on the last day, routes a ticket to the right team | Handles non-standard security requests | Before: one in three tickets goes to the wrong team |
| Marketing | Enriching an inbound lead (adding client company data); reconciling media invoices with orders and actual impressions | Enriches the lead, scores it, passes it to the right sales rep; reconciles orders with impressions | Runs campaigns, resolves disputes with platforms | After: an inbound lead reaches the right rep in under 5 minutes. Before: 8-12% of media budget is lost to discrepancies |
| Support | Escalation and summarising a ticket | Routes a ticket to the right team immediately, writes a brief summary so it is not sent back for details | Handles complex technical cases | Varick has no numbers here; the goal is first-pass escalation without being sent back for details |
All these agents run in the background inside existing systems: ERP (the company's accounting and operations system), CRM (the system for tracking customers and deals), HR modules, email, and working chat channels. Exceptions go to people; everything else is completed without a person stepping in.
It is useful to compare your own measurements with industry reports. Ardent Partners, in State of ePayables 2025, records the average exception rate in accounts payable at 18.4%, 11.1% at top companies, and the full cost of processing one invoice at $9.84. The report was sponsored by Bottomline, so the figures should be checked against your own data. Salesforce, in State of Sales 2026, reports that sales reps spend 40% of their working week on selling; 60% goes to data entry, preparation, and internal tasks. Note that Salesforce sells CRM itself. APQC notes that top companies close the month in about 5 days, slower ones in roughly 10. Varick's 7-9 days after agents fall within that range, while 18-22 days before agents is probably large groups with multiple legal entities. That is a guess, not a fact.
What processes suitable for an AI agent have in common
In all nine cases the same pattern repeats. Human actions are well documented: there is a procedure or an established routine. Data lives in systems, not in people's heads. Check rules are known: an invoice must match the order on amount and details, access is granted by a role list, a contract must contain certain clauses. Exceptions are recognised by a clear signal: the amount does not match, there is no stock, a date was missed.
Where input is predictable, that part of the agent is written in code: rule-based matching, routing, sending notifications. A language model inside the agent is needed where input is unpredictable: a new invoice layout from a supplier that changed its accounting system, posting a supplier invoice to the right account based on previous similar invoices where no single rule applies. Vas writes that in most agents the model occupies a small share; the rest is code on top of existing systems.
Why start with a department, not a single process
The finance department meets its deadline, the procurement department meets its own, and the supplier invoice still takes three weeks. Time is lost in handoffs between departments: a document sits in someone else's queue waiting for someone to notice the order number does not match. A single process inside one team does not touch those handoffs, and the whole company is too wide for a first step. A department is the working unit where both processes and their handoffs to neighbours are visible: finance agents read from and write to procurement, payroll, and other departments, and the whole cycle shortens. We recommend that within the chosen department you start with the process that has the most handoffs between people - the reach axis measures the same thing at department level.
How to choose the first department on three axes
Before scoring, in each department select the processes that fit an agent (the four signs from the section on what such processes have in common) and that run frequently. Then score the department on those selected processes. The scale below is Majento's proposal; Varick only names three criteria without thresholds. The CFO does the scoring together with department heads.
Volume: the number of runs of selected processes per month: how many times the process goes from start to finish (an invoice, a request, a ticket - each counts as one run). Scale: under 100, 1 point; 100 to 299, 2; 300 to 999, 3; 1,000 to 2,999, 4; 3,000 and above, 5. Sources for the numbers: accounting system reports, CRM, support ticket history, e-invoice exports.
Cost: the department's annual payroll as a share of the most expensive department being evaluated: 80% or more of the maximum, 5; 60% to 79%, 4; 40% to 59%, 3; 20% to 39%, 2; under 20%, 1. The CFO usually knows the figure.
Reach: the number of other departments whose processes exchange data or documents with this one: 1, one point; 2, two; 3, three; 4, four; 5 or more, five.
In Varick's client work, about 40% of projects fall in finance, 25% in sales operations, then HR, procurement, or operations, and at the end legal and marketing.
Teaching example with hypothetical numbers. A fictional distributor in Tashkent, 300 employees, four departments in focus:
| Department | Runs per month | Volume, score | Annual payroll (illustrative) | Cost, score | Reach, score | Total |
|---|---|---|---|---|---|---|
| Finance, 15 people, touches procurement, sales, HR, logistics, marketing | 3,200 | 5 | UZS 900 million | 5 | 5 | 15 |
| Sales, 20 people, touches marketing, logistics, finance | 1,500 | 4 | UZS 1,000 million | 5 | 3 | 12 |
| Procurement, 8 people, touches finance, logistics, sales | 400 | 3 | UZS 400 million | 3 | 3 | 9 |
| HR, 5 people, touches finance, IT | 300 | 3 | UZS 250 million | 2 | 2 | 7 |
Conclusion: start with finance. It has the most operations and the widest reach, and on cost it is nearly equal to sales, so the total is higher. How to put a price on a task before starting work is described in a separate article.
What changes when moving the process map to a company in Uzbekistan
In a company in Uzbekistan, three-way matching is built on e-invoices. Under Cabinet of Ministers Resolution No. 489 of 14 August 2020, invoices are issued in electronic form, and VAT is deductible only against electronic invoices; for goods subject to mandatory digital labelling, an e-invoice without the digital labelling code cannot be issued. Documents are issued through electronic document operators and the taxpayer's portal. For an agent this is a plus: data is already electronic, and matching an invoice, contract, or order against a delivery note or acceptance certificate is easier than reading paper scans. The matching agent also checks the digital labelling code.
American forms and requirements do not apply to a company in Uzbekistan. In Varick's material, the US W-9 form comes up in supplier onboarding, the I-9 in employee onboarding, and SOX in audit. A company in Uzbekistan works with different documents: counterparty details and tax registration number, counterparty verification, HR documents per the Labour Code, auditor and bank requirements. Systems differ too: for example, 1C, a local CRM, working chats in Telegram. The principle is the same: the agent works inside those systems, but specific matching rules and document lists are written from scratch.
Personal data
Law ZRU-1125 of 26 March 2026, in force from 27 March 2026, narrowed mandatory local storage. Before March 2026, all citizens' personal data had to be stored on servers in Uzbekistan. Now only biometric and genetic data, plus telecom operators' subscriber data, must be stored locally. Other personal data can be stored and processed abroad under one of three grounds: the country is on the Cabinet of Ministers' list of adequate protection; the operator adopted standard contractual clauses or binding corporate rules as required by the competent authority; the operator complies with international standards from the approved list. For an HR agent and support, before launch, choose the ground and check the current lists. A personal data database that must be stored in Uzbekistan is registered in the State Register of Personal Data Databases; if an HR agent touches biometrics (for example, Face ID at the gate), that must be accounted for before launch. This is not legal advice.
How to launch the first AI agent without disrupting operations
In the first week, the process owner involves security (access rights), legal (personal data and contracts), and finance (baseline metrics). Then metrics are agreed in writing across three groups. The first group covers process cost and time: cost of processing one invoice, days to close the month, number of exceptions. The second group is revenue net of model expenses; for internal departments this metric can be skipped. The third group is risk, for example the error rate or the value of disputed operations. Record current values before building.
What access to give an AI agent
IT sets up a separate account for the agent with minimal permissions and an action log. For reference: SailPoint's survey, Horizons of Identity Security 2026-2027, among 340 leaders in security, IT, and risk, shows that 79% of organisations have already deployed AI agents in production, and only 2% use access-control tools built for agents. SailPoint sells such tools, so treat the figure as a market signal. How to set up access rights in corporate systems is covered here.
Mapping how the process actually runs
Vas describes a client, a global distributor that had handed chat assistants to 12,000 employees a year earlier. The client's month-end close procedure listed 11 steps. System logs and conversations with people revealed 20 steps in the system and 4 in email. Breaking down those 20 steps performed by the finance team led, per Varick, to the following: 7 steps removed, 4 handled by code, 6 by agents, 3 left to a human because auditors require it. The fate of the email steps is not specified. After 4 months, month-end close took 8 days instead of 20, and exceptions dropped below 50 per month from 700. The cost of processing one invoice turned out to be $31, although the company had assumed $5. That is why the first step of a launch is recording baseline metrics. How to build such a map is described in an article on steps before deploying agents.
Shadow mode: the agent first only proposes decisions, then gradually executes
Varick reports about 4 weeks to build one agent, then a 12-week schedule counted from the start of shadow mode. For the first two weeks the agent runs at full volume without the right to act, and the team reviews every discrepancy with the human; from week three the agent executes the cases where the team has coded all the checks and is at least 99% sure of the action, usually about half of operations; by week twelve about 90% runs through the agent, the rest stays with people. For reference: at a Varick client, the board-level result was expected at 16 weeks, and the summary was delivered at 4 months. Before letting the agent execute, we recommend additionally checking how often its shadow-mode decisions matched human decisions.
Lean on employees who already work with AI
In client companies, Varick finds 5-10% of employees already using chat assistants on their own. Train them in a day, rank them by time saved, and they build agents for the next department.
What to do with already-purchased subscriptions and stalled pilots
Chat-assistant subscriptions remain: employees work faster with them. From the usage report, identify the most active users (Varick sees about 10%) and make them future agent builders, see the section on employees. Stalled pilots get measured: record what the pilot did, how much time it saved, where it failed. Typical reasons for stalling that Varick sees: built inside one team, did not map the process, did not set a metric before building, the demo worked on a smooth scenario and broke on the first exception. Vas's example: a pilot saves 8 minutes in a process that takes 22 days; it closes one step, and reports do not show it. Pilots that cannot be measured are closed; the rest are included in the chosen department, their process is mapped as described above, and only then they go into shadow mode.
Questions and answers
Will this approach work for a company of 50-100 people
The method applies, but Varick has no data on companies of that size. Measure your own volume and start with the department that has the most selected processes.
What if accounting is outsourced
We would build it this way: the agent works at the junction, receives e-invoices from the electronic document operator, matches them against the order and delivery note, and passes a ready package to the outsourcer. Exceptions still go to people, but the volume arriving at the outsourcer shrinks.
Can employee data be sent to a foreign AI service
Employee data can be sent to a foreign AI service under one of the three grounds in the Personal data section. Biometric and genetic data must be stored in Uzbekistan by law; do not send them to a foreign service. Check the lists on the launch date.
How long until the first result
About 4 weeks to build an agent, then the 12-week shadow-mode schedule (details in the launch section). Varick's client reference point is 16 weeks to results for the board. This describes what Vas sees at Varick's clients, not a Majento promise.
Do you need to replace 1C or CRM
We would do it this way: the agent works inside existing systems, reads data, writes results, passes exceptions. If the system allows an API connection or the agent operates through the program's screen like a person, that is enough. If not, access is resolved first, then the agent is built.
Who is responsible for an AI agent's error
The human who owns the process is responsible for the outcome. The agent operates within granted permissions; the human approves rules, handles exceptions, and answers for the result. That is why shadow mode is mandatory: for the first weeks a person checks every decision.
Majento breaks down department processes and builds agents on top of existing systems. AI implementation and business process automation. Telegram t.me/shimaoz, email hello@majento.ai.