Artificial intelligence
AI creates business impact when it is embedded in a specific decision or process: a recommendation, communication, consultation, forecast, control or automation. A use case should be selected based on the value of the task, data availability, cost of error, integration feasibility and measurable impact on P&L or operating results.
Author: Maksim Shelukhanov · Published and updated: 25 August 2026
A practical artificial intelligence strategy starts with decisions that affect customers, revenue, cost and speed of execution rather than with technology.
Recommendations, next-best offers, outfit or bundle selection and personalised communication help increase conversion, average order value, frequency and retention. Impact should be measured against a control group and margin, rather than by the number of recommendations shown.
An assistant searches policies, catalogues and the knowledge base, suggests the next step and reduces consultation time. The employee remains accountable for the decision, especially in complex, costly and sensitive situations.
AI helps identify intent, prepare a response, route a request and monitor quality. The goal is faster, more accurate resolution with a clear route to a human, rather than the highest possible automation rate.
Models can support demand planning, replenishment, pricing, retention and risk assessment. Their value depends on whether the real decision changes and the business metric improves compared with the existing process.
Document classification, data validation, draft preparation and variance control free up team capacity. Priority goes to high-volume operations with a clear quality standard, measurable effort and a manageable cost of error.
A successful AI implementation should begin with a narrow task where value can be proven before processes and platforms are rebuilt at scale.
How often the task occurs, how much time, revenue or quality the current process loses, and who owns the outcome. A frequent operation with clear economics usually takes priority over a rare but impressive use case.
The use case needs accessible, lawfully used and sufficiently reliable data, together with a signal for the correct answer. Before the pilot, check completeness, freshness, labelling and whether the model output can be linked to the next action.
Define the acceptable error, cases that require human review and a safe failure mode. The greater the consequences, the more important human oversight, transparent limits and the ability to reverse an incorrect action quickly become.
Even an accurate answer is useless without a place in the interface, CRM, contact centre or management cycle. Define the user, decision point, required systems and action following the recommendation in advance.
Before launch, set the baseline, control group or comparable period, and leading and outcome metrics. This separates the real contribution of AI from seasonality, marketing pressure and other business changes.
Business AI agents differ in their degree of autonomy, access to systems and the human role in the process, rather than in their label.
Prepares information, a draft response or a recommendation, while the decision and action remain with an employee. This is a strong first format for knowledge, analytics and complex consultations where context and control matter.
Plans several steps and performs authorised actions in specified systems. It needs narrowly scoped permissions, limits, logging, outcome checks and a clear escalation path.
Executes predefined rules without independently choosing a goal. It is often cheaper, more stable and easier to explain than AI, so first determine whether the task truly requires a probabilistic decision.
A pilot tests the full loop—data, decision, user action, economics and the process’s ability to operate reliably—rather than a model demonstration.
Record the current metric, process cost, target user and expected mechanism of impact. Frame the hypothesis so that a limited test can disprove it.
The team checks sources, access constraints and example quality, then builds a minimum prototype around a real use case. Manual preparation is acceptable when it helps validate value quickly.
Set accuracy, speed, business metric, budget and stopping criteria in advance. A limited audience reveals real errors and user behaviour without excessive risk.
Embed the solution into the working interface, access rights, procedures and team accountability. Employee training and a new operating procedure often matter more than another incremental model improvement.
Once impact is proven, expand segments and volume while monitoring quality, cost, data drift and business outcomes. Scale in stages with a rollback option.
The financial model includes incremental revenue and margin, lower service costs, and faster and better decisions. Deduct the costs of data, integrations, computing, controls, support and process change. TCO matters more than prototype cost: a cheap pilot can become an expensive system. For a portfolio of initiatives, assess expected impact, confidence, time to value and resource requirements.
Every use case needs a business owner, data-access rules, privacy requirements, quality criteria and human escalation. Decisions and actions are logged, critical responses are reviewed, and permissions follow least privilege. Data ageing, quality variance across segments and process dependency on a vendor are monitored separately.
Each quantitative result belongs to a specific company and is not transferred between cases.
At Kenguru, AI personalisation of product offers and an outfit builder generated RUB 13 million in additional monthly revenue. The solution operated within a broader CRM, omnichannel sales and customer-base management ecosystem.
At ORTEKA, an employee AI assistant, knowledge base and personalisation formed part of the transformation programme. RUB 165 million was the impact of the full portfolio of UX, traffic, mobile app and CRM initiatives, rather than a single AI mechanism.
A case demonstrates the link between a task, process, decisions and measurement, rather than a universal return from the technology. Before applying the approach elsewhere, reassess the customer journey, data, constraints and economics.
AI cannot fix missing ownership, conflicting rules or fragmented execution; it can only reproduce their consequences faster.
If customer, product, order and financial results are defined differently, the first requirement is data strategy and management analytics. Otherwise, training, measurement and control will rely on incompatible signals.
When similar cases are handled inconsistently and there is no owner or quality standard, define the target process first. Only then is it clear which part should be supported by an assistant, agent or conventional automation.
If the use case affects the product, channels, roles, KPIs and technology architecture, manage it as part of digital transformation or an omnichannel model, rather than as an isolated experiment.
Brief answers to questions that arise before the first pilot and when moving to scale.
List 10–20 recurring decisions and operations, then assess their value, frequency, data, risk and measurability. Choose one narrow task with an owner and a fast outcome signal for the first test.
What is needed is an aligned portfolio of use cases, shared data and risk rules, technology principles and economics. It should form part of the business and product strategy.
When the solution improves a preselected business or operating metric, meets quality and risk requirements, is adopted by users and has clear economics for scaling.
Searching for technology without a problem, weak data, no owner, a pilot outside the working process and measurement limited to technical accuracy. Address these issues before increasing the budget.
Describe the decision or process that needs improvement. I can help define the value of the use case, data requirements, pilot design, outcome criteria and the initiative’s place among business priorities.
Contact
Roles with responsibility for P&L, growth, commercial strategy, operations and transformation.