All services · AI for one work task

One workflow · verifiable output

AI management assistant using company data

For business owners and managers who need to understand sales, team tasks, and stock levels. The assistant answers different questions about sales, recorded employee actions, and warehouses within an agreed scope. You can ask follow-up questions and choose another period. The assistant works with recorded data; the manager makes the decisions. Covert employee monitoring is outside this scope.

Denis Skripnik — a blind specialist in working with AI agents. I examine the task, configure the workflow, and work with you to verify the result against agreed rules.

AI agent setup is already selected in the brief. Name the scenario: “AI management assistant using company data”. Describe the files or systems you use, the result you need, and any exceptions. Start with a description of an anonymized sample.

Input: business records and metric definitions

• Agreed CRM, accounting, or warehouse sources, or their exports: deals, recorded actions, orders, and stock, using only necessary fields.
• Definitions of metrics, statuses, periods, time zones, and units. We define sales, activity, and available stock, including returns and reservations.
• Examples of different management questions, user roles, permitted fields, and query volume. We assess systems and access using samples rather than promise support for every CRM or 1C installation.

Output: an answer you can verify

• A concise answer and, where needed, a table: record and export links or identifiers tied to specific statements.
• For numbers, a reproducible calculation: contributing rows, formula, filters, period, units, and source timestamp.
• The assistant flags sources that are unavailable, outdated, or inconsistent with other records. When evidence is insufficient, the assistant asks for clarification or says it cannot answer rather than invent a value.

Example: deals without a next action and available stock

An illustration, not a client case study. A manager asks: “Which open deals have no next action, and is the required product in stock?” A sample CRM export shows deal D-17 requiring 8 units of product A, with no next task recorded. A stock export as of 09:00 lists 12 units of A, including 5 reserved.

Answer: “D-17 has no next action recorded. Available stock is 12 − 5 = 7 units of A against demand of 8: a shortfall of 1 unit.” The answer links to the deal and stock row, timestamps the export, and states the calculation rule. This is not a shipment promise. A missing record does not prove an employee did no work; mismatched product codes stop the comparison.

The manager can then ask about sales for another period or recorded team activity. The assistant uses the same permitted sources and calculation rules.

What the implementation includes

• Agreement on systems, access roles, data types, volume, and supported questions; assessment of existing reports before development.
• Reading selected sources, checking permissions, linking records by agreed keys, rule-based calculations, and source-linked answers.
• A test set covering different questions and ambiguities: a missing record, stale stock, different units, an undefined metric, and a request outside the permitted role.

Minimum access and customer responsibilities

• Preferably a dedicated account with read-only access to selected entities and fields, restricted by role. Agreed exports are an alternative; answers are then limited by their contents and dates.
• The customer assigns a data owner, explains fields and metrics, confirms processing rights, and reviews test answers. Access secrets are not submitted through the brief.

Acceptance and assistant limits

In the customer’s environment, we test several question types, filters and calculations, links between statements and sources, freshness, and refusal when data or permissions are missing. Error tolerances and volume are agreed before work starts. Handover includes instructions and training for the manager or responsible employee.

Read-only: no arbitrary SQL, CRM or accounting changes, employee performance judgments, or automated decisions. Correlation is not presented as the cause of a sales decline; a CRM gap is not proof of inactivity. New systems, roles, and volumes require separate agreement.

Sources, access, and your responsibilities

• An anonymized export is enough to start: remove information that could identify a person, business secrets, and unnecessary fields. Do not put passwords or keys in the brief.
• For implementation, we agree the minimum necessary permissions, preferably a dedicated account with read access only to selected sources. An export instead of production-system access is an acceptable alternative.
• You provide lawfully available samples, rules, an authorized environment, and a designated reviewer who resolves disputed data. We agree storage, deletion, and permitted model-provider data transfers in advance.

Acceptance, instructions, and training

At acceptance, we run the chosen workflow in the customer’s environment. We agree criteria, a test set with ordinary and disputed examples, output format, and error reporting before implementation. Handover provides operating and review instructions and training for the responsible employee; their format and scope are specified in the proposal.

Iteration count, correction of nonconforming results, requirement changes, and support after acceptance are agreed separately before starting. Unlimited revisions and ongoing support are not implied.

Timing, price, and external expenses

The price and exact duration are confirmed after reviewing a sample and before work starts. They depend on data quality, formats, volume, environment, and review requirements. Existing service price guidance is not a fixed quote for this scenario.

If a sample is insufficient to define the scope reliably, a paid discovery or bounded pilot can be agreed separately, with its own deliverable and criteria. It does not automatically commit either party to full implementation.

Server costs, paid services, and model usage are separate from implementation fees. Before starting, we agree what is necessary, who pays for accounts, and spending and processing limits. Payment for the work follows scope and quote confirmation only.