AI agents
Work with documents and context, prepare recommendations, check outputs and hand complex decisions to a person.
Your business.
Connected.
Bring processes, data and team workflows together. AI agents, integrations and custom software, built around your work.
Sergeev AgentsWorkflow architecture
Documents
CRM
Team
Data
Inputs matchedInvoice linked to order
A person decidesReview and approve
Illustrative connected system
When work depends on spreadsheets, messages and personal follow-up — it is time to connect the workflow.
For leaders and teams working across several tools.
Every task has
a path.
Connect inputs, rules and human decisions. Choose a workflow to explore an example.
From demand to purchase
Stock, incoming orders and constraints — in one purchasing decision.
Demand
Demand, stock and incoming deliveries.
Data checks
Lead times, pack sizes, budget and possible options.
Human decision
The buyer reviews the evidence and approves the plan.
Purchase plan
A recorded decision and a clear next step.
Automation prepares the evidence.
A person makes the decision.
Illustrative workspace
Replenishment recommendation
Buyer approval required
Check the lead time and approve the quantity.
Source availableDecision recorded in history
From request to resolution
Facts, rules and actions — with operator approval.
Request
A customer request and available history.
Fact checks
Terms, documents and permitted actions.
Approval
The operator reviews the evidence and authorises the action.
Outcome check
The outcome is checked; uncertainty goes to a person.
Automation prepares the evidence.
A person makes the decision.
Illustrative workspace
Resolution proposal
Operator approval required
Review the terms and confirm the permitted action.
Source availableDecision recorded in history
From documents to a decision
Orders, invoices and payments — with the source of every discrepancy.
Documents
Contracts, orders, invoices, deliveries and payments.
Matching
Partial transactions, versions and duplicates.
Human review
A reviewer inspects the source and records a decision.
Final report
A report with evidence and review history.
Automation prepares the evidence.
A person makes the decision.
Illustrative workspace
Document reconciliation
Discrepancy review required
Open the source document and review the discrepancy.
Source availableDecision recorded in history
Illustrative workflow with synthetic data. No live systems are connected.
Explore the logic
with an example.
Choose a task and run a review. This illustrative workspace compares synthetic data, finds a discrepancy and explains where a person needs to decide.
Runs in your browser
Invoice and order do not match
Source data
Order PO-024 / invoice INV-024
| Source | Qty | Price | Total |
|---|---|---|---|
| Order | 12 | 8 000 ₽ | 96 000 ₽ |
| Invoice | 12 | 8 400 ₽ | 100 800 ₽ |
Rule: invoice quantity and unit price must match the order.
REVIEW SIGNALIllustration · not a forecast
Review steps
- 1Read
- 2Compare
- 3Explain
Explore the example: sources → rule → decision evidence.
₽4,800invoice discrepancy
HUMAN DECISION REQUIRED
Check the price change with the supplier. Hold payment processing until approval.
Local analysis using preset rules. No actions are performed in business systems.
A refund request needs review
Source data
Order SO-031 / request CS-031
| Fact | Value |
|---|---|
| Order payment | 18 000 ₽ |
| Requested refund | 20 000 ₽ |
| Reason verified | Document missing |
Example rule: refund must not exceed payment; the reason must be verified.
REVIEW SIGNALIllustration · not a forecast
Review steps
- 1Read
- 2Compare
- 3Explain
Explore the example: sources → rule → decision evidence.
₽2,000above the payment amount
HUMAN DECISION REQUIRED
Clarify the amount and request supporting evidence. Prepare a reply; an operator approves sending and refunding.
Local analysis using preset rules. No actions are performed in business systems.
Stock runs out before delivery
Source data
SKU-014 stock / delivery plan
| Measure | Value |
|---|---|
| Stock | 24 units |
| Daily usage | 8 units |
| Lead time | 5 days |
| Pack size | 12 units |
Simplified example: constant usage, no safety stock or other deliveries.
REVIEW SIGNALIllustration · not a forecast
Review steps
- 1Read
- 2Compare
- 3Explain
Explore the example: sources → rule → decision evidence.
2 dayswithout stock before delivery
HUMAN DECISION REQUIRED
Check expedited delivery of 24 units or an alternative supplier. A regular 5-day order cannot close the gap.
Local analysis using preset rules. No actions are performed in business systems.
Tools that fit
your work.
Integrations
Connect tools with agreed permissions, approvals and outcome checks.
Custom software
Workspaces, task queues and interfaces that give every role the context, status and next step.
Different tools.
One workflow.
Connect systems around how your team works. Data moves forward with context and rules.
CRM
Requests and customer history.
This illustrates possible connections. Integrations depend on APIs, access rights and the agreed scope.
Select a node to explore its role
See the logic
inside the system.
Our own reference systems show how data, actions and team decisions fit together.
All four are demonstrations on synthetic data. These are not client deployments.
Documents and paymentsEvery discrepancy has a source.
Illustrative system
InvoiceINV-024OrderPO-024
INV-024PO-024Review
INV-025PO-025Matched
INV-026PO-026Matched
Discrepancy → source → decision
Matches contracts, orders, invoices, deliveries and payments. Handles partial transactions, versions and duplicates. A reviewer sees the evidence beside the source document and records the decision.
Matching rules and supported documents are defined before implementation. Suitability is tested against agreed examples.
Customer resolutionAn action ends with a check.
Illustrative system
FactsApprovalAction
The action’s outcome is checked separately
Connects facts and rules with operator approval and an action in a sandbox system. Shows the next step and outcome. Uncertain outcomes are routed to manual review.
Before live connections, we define permitted actions, approvals and situations that require a person’s decision.
Demand and inventoryA recommendation with clear reasoning.
Illustrative system
Illustrative demandReplenishment plan
DemandHistory
StockChecked
ReplenishApprove
Connects demand, stock and incoming deliveries with replenishment options. Accounts for lead times, pack sizes and budget, explains recommendations and records the buyer’s approved decision.
Data and rules are agreed for your workflow. Recommendation accuracy and suitability are validated separately; financial outcomes are not guaranteed.
Agent testingInspect failures before rollout.
Illustrative system
Grounded response
Access check
Action approval
Failing versionBlocked by a check
Synthetic scenarios and test agents
Compares versions against defined scenarios and blocks a version that fails the checks. The team can inspect passed scenarios and issues that need fixing.
This demonstration was tested with synthetic agents. A real local model and its quality still require validation.
Two more prototypes in development
PR assistant
Meeting notes, assignments, tasks and drafts in one context. Tested on synthetic data; external integrations are not yet connected.
Bakery operations
Demand, purchasing, production, quality, delivery and documents. Scenarios are modelled on synthetic data; live connections are still in development.
Start with one
workflow.
Define the first stage, the required outcome and acceptance checks. Pricing and timing follow the agreed scope.
- 1
Understand
Map the current workflow: people, data, tools and handoffs where context gets lost.
- 2
Design
Agree scope, rules, access rights and criteria for checking the result.
- 3
Build and test
Build the system. Test normal paths, failures and approval points.
- 4
Roll out in stages
Agree live connections, team handover and ongoing support.
The first stage produces an agreed scope, tested scenarios and a decision on what comes next.
Discuss the first stage
Fedor SergeevSergeev Agents
Behind the system,
a person.
I’m Fedor Sergeev. I build AI agents, integrations and tools for business workflows. The team should understand how the system works and where their decision is needed.
In my writing, I explain the approach, calculations and limitations through practical examples.
Personal Telegram channelThe approach, in detail.
All articlesBefore we start.
Do we need to replace our existing tools?
We start with your existing tools and available connection methods. The system is defined after mapping the workflow. Integration depends on each service’s API, access rights and limits.
Does every project need AI?
Tools follow the task. Rules, integrations or a clear interface may be sufficient. Where AI is useful, we define how outputs are checked and when a person takes over.
How are pricing and timing determined?
After agreeing the first stage: data, integrations, user journeys and acceptance checks. There is no universal price or timeline; scope depends on your workflow.
How will our data be handled?
Before connecting anything, we agree the data, access rights, processing environment and permitted actions. Demonstrations use synthetic data. Handling of real data is defined for each project.
What happens after launch?
The scope, timing and format of support are agreed separately. This may include operational checks, team support and changes as the workflow evolves.
Tell me how work happens today:
the tools, the people and the repetitive handoffs.
What needs
connecting?
Discuss your workflow on TelegramStart with your current workflow and the result you need.