Companies rarely suffer from a shortage of data. They suffer from a shortage of agreement about which facts matter, who owns them and what decision should change when a signal moves. Sensors, transactions, customer contacts and logistics events accumulate faster than an organization can turn them into coordinated action.
Forbes Russia described the transition on March 25, 2025, citing an estimated 402.74 million terabytes generated worldwide each day. It reported that the share of executives who said they had built a data culture rose from 28.3% in 2019 to 42.6%, while a Gartner forecast expected 65% of organizations to be data-oriented by 2026.
For businesses in Russia, the lesson was not to collect everything. It was to build an operating system in which definitions, quality, access, accountability and human judgment reinforce one another. Data becomes capital only when it repeatedly improves a real decision.
A data strategy starts with an economic decision
A useful program begins with a business question: which customer to serve, which machine to maintain, how much inventory to hold or where to place capital. The question establishes the required speed, accuracy and history. Without it, an architecture team can deliver an impressive platform that nobody needs during an operating meeting.
Management should state the baseline cost of the decision and the value of improvement. A demand forecast may reduce waste, working capital and missed sales, but those benefits occur in different accounts. Finance, operations and technology need one agreed calculation before investment begins.
The first release should be narrow enough to learn and important enough to matter. A pilot on an artificial problem produces polite adoption but little evidence. A bounded production process with a named owner reveals missing fields, incentives and exceptions while the cost of correction remains manageable.
A data lake without purpose becomes a liability
The earlier data-lake model encouraged organizations to store information first and decide on its use later. Storage became cheaper, so postponing selection appeared rational. Over time, undocumented copies, incompatible formats and unclear ownership turned the lake into a data swamp whose contents were expensive to trust.
Retention has a carrying cost beyond disks. Teams must secure, catalogue, reconcile and answer legal questions about every sensitive dataset. Old information can create breach exposure while producing no decision value. A retention rule should connect each class of data to an operational, regulatory or analytical purpose.
Deletion is therefore part of governance. An owner approves the lifecycle, archives evidence that must remain and removes redundant copies through a controlled process. The organization preserves lineage so that a reported number can still be traced without keeping every intermediate file forever.
Quality is defined by the cost of a wrong answer
Forbes cited research in which 85% of senior executives recognized data as important, yet only a third were satisfied with its quality and respondents considered 40% of accumulated information unusable. The gap shows why volume is a weak measure of maturity.
Quality requirements vary by decision. A marketing experiment may tolerate an incomplete demographic field; a safety limit or financial reserve cannot tolerate the wrong unit. Controls should reflect consequence, with the strongest validation applied where an error can stop production, misstate accounts or harm a person.
A useful quality record names the field, definition, source, owner, expected range and last successful check. It also records exceptions rather than silently correcting them. Repeated exceptions show whether the process, interface or incentive is creating bad information upstream.
Common units prevent expensive illusions
The source illustrated how departments might report cost in millions of rubles per tonne and thousands of rubles per barrel. Combining both without conversion can distort material reserves and production plans. A dashboard can be mathematically precise while comparing unlike quantities.
Canonical units, currencies, time zones and calendars should be machine-readable. Conversion rules need version control because exchange rates, product factors and accounting policies change. A user must see both the normalized value and enough source context to challenge it.
Reconciliation belongs near the operational boundary. Waiting for a central team to clean every discrepancy at month end turns data management into archaeology. Operators should receive immediate feedback when a measurement violates a contract, while stewardship teams resolve structural conflicts across systems.
One customer needs one negotiated meaning
Data silos are not only technical. A corporate bank may define a customer as a legal group while a small-business unit treats each borrower separately. Both definitions can be valid locally, but exchanging records without context creates duplication and contradictory risk views.
A shared semantic layer does not force every department to abandon its model. It describes relationships among local meanings and establishes the enterprise definition used for consolidated decisions. Changes require representatives from the affected processes, not only database administrators.
Data products need service contracts: what is supplied, at what frequency, with which quality and whom to contact after failure. Decentralized ownership can speed domain work, but interoperability remains a central obligation. Autonomy without standards merely distributes confusion.
Technology cannot overrule the operating culture
A spare-parts forecast creates no saving if a production manager continues ordering by instinct. Resistance may reflect habit, but it may also reveal constraints the model missed: supplier reliability, maintenance windows or the local cost of a stockout. Adoption requires a dialogue between evidence and experience.
Leaders should ask for the forecast and record why they override it. Overrides become training evidence and expose recurring blind spots. Punishing every disagreement encourages employees to follow a model even when conditions have changed, which replaces informed judgment with compliance theatre.
Visible success builds culture faster than slogans. When a team reduces downtime or releases working capital, management should show the decision, evidence and measured result. Training then uses familiar cases, and employees understand how better records affect their own work.
Privacy belongs in product design
Behavioral and biometric signals can identify people even when obvious names are removed. Combining a payment fragment, location, device and movement pattern may recreate an identity. Data minimization is therefore more reliable than assuming that a dataset is harmless because one column was deleted.
Every use case needs a lawful purpose, access boundary and retention period. Sensitive attributes should be separated, encrypted and logged. Teams should test whether a less granular feature can deliver the business outcome before collecting information that customers cannot replace after a breach.
Artificial-intelligence agents increase both productivity and attack speed. Automated access must use least privilege, short-lived credentials and reviewable actions. An agent should not inherit broad employee permissions merely because it performs work on that employee's behalf.
Digital systems must survive analog events
A route optimizer may understand traffic and weather yet fail during a drivers' strike, border closure or sudden fuel shortage. Historical data contains few examples of rare structural breaks. The best statistical answer inside the old rules can become operationally impossible.
Resilience requires scenarios, alternative inputs and manual authority. Teams should rehearse loss of a feed, cloud region, supplier or communication channel. A drill measures how long the business can operate, which decisions degrade first and what minimum dataset must remain locally available.
Fallback is not a permanent rejection of automation. It buys time while people diagnose the exception and update the model. After the event, the organization records what signal was missing and whether a new rule improves response without overfitting to one crisis.
Specialists should be valued against the delay they remove
Data engineers, architects, analysts, security experts and domain stewards are scarce because a working system needs all of them. A vacancy cap based only on comparable salaries ignores the value lost while a project waits or an unreliable process continues.
The business case should estimate the bottleneck. If one specialist advances a high-value release by three months, the economic contribution may exceed compensation many times. The same logic prevents indiscriminate hiring: a role without access, sponsorship or a defined decision cannot create the expected return.
Retention depends on professional conditions as well as pay. Skilled people need clear ownership, usable tools, access to domain experts and permission to improve root causes. Repeatedly asking them to repair manual spreadsheets after upstream errors drives away the capability the company intended to build.
Governance needs decision rights, not a committee maze
A council is useful when it settles cross-business definitions and risk. It becomes harmful when every field change waits for a monthly meeting. Rights should be distributed: domain owners approve ordinary changes within standards, while enterprise stewards resolve conflicts with material downstream impact.
Escalation thresholds must be explicit. A local defect can be corrected by its product team; a change affecting financial reporting, personal information or multiple customer channels receives wider review. The record should explain the decision and its effective date so consumers can adapt.
Funding follows ownership. If a critical dataset serves five divisions but belongs to none, maintenance will lose every budget contest. A named executive sponsor finances reliability, while usage-based evidence shows which consumers benefit and should share the long-term cost.
Metrics should connect adoption to financial outcomes
Counting dashboards, records or model releases rewards activity. Management needs a chain from technical health to process behavior and economic result. For inventory, that chain may include forecast error, planner adoption, stock availability, write-offs and released working capital.
Attribution should remain conservative. Sales can change because of price, season or competition as well as analytics. Controlled tests, matched groups and pre-agreed baselines reduce exaggerated claims. Finance should validate benefits with the same discipline used for other capital projects.
Some returns are risk reductions rather than immediate profit. Faster breach detection, traceable regulatory reporting and tested recovery have option value. Scenario analysis can quantify the exposure reduced without pretending that a prevented event produced ordinary revenue.
A practical executive control sheet
- business decision, owner and expected economic effect;
- critical sources, definitions and quality thresholds;
- legal purpose, access boundary and retention rule;
- model adoption, overrides and reasons for disagreement;
- technical availability and tested fallback time;
- skills bottleneck and accountable product team;
- measured process result and finance-validated benefit;
- next review date for assumptions and controls.
Scale follows a proven operating loop
After one decision works, the organization can reuse identity, quality, lineage and security components. It should not copy the model blindly. Each new process has different timing, incentives and loss from error, so reusable infrastructure must be combined with local discovery.
A portfolio view prevents dozens of pilots from competing for the same experts. Initiatives can be ranked by value, readiness, dependency and learning. Projects that establish a common customer definition or reliable product master may unlock several later applications and deserve priority beyond their direct return.
Retiring failed experiments is evidence of discipline. A pilot should have success, correction and stop conditions before launch. When the signal is weak, the team preserves what it learned and releases capacity rather than protecting sunk cost with increasingly generous assumptions.
Buy and build decisions depend on differentiation
Commodity storage, integration and monitoring can often be purchased, while unique pricing, production or risk logic may deserve internal development. The boundary should reflect strategic differentiation, switching cost, regulatory constraints and access to capable staff.
Vendor evaluation needs an exit test. The company should be able to export data, definitions, history and model evidence in usable formats. A low initial price can become expensive if interfaces, skills and contractual rights make later migration impractical.
Internal development also creates lock-in when only one engineer understands it. Documentation, testing, peer review and succession protect the asset. The goal is not maximum ownership but controlled choice across the system's life.
Continuous review keeps evidence relevant
Markets, customer behavior and operations change. A model that performed well last year may drift quietly as product mix or policy moves. Monitoring should watch input distribution, output accuracy, adoption and business outcome rather than only server uptime.
Review frequency should match decision speed and consequence. Fraud signals may need daily observation; a capital-planning model may need formal quarterly challenge. Material changes trigger revalidation even when the calendar has not reached its normal review date.
Owners should be able to suspend automated action while preserving observation. This safe state allows investigation without losing the incoming evidence. Resumption requires a recorded explanation, corrected controls and communication to affected users.
A twelve-month sequence turns principles into work
During the first quarter, management should select two material decisions, name their owners and document the present process. Teams map sources, definitions, manual adjustments and the cost of delay or error. Security and legal specialists classify sensitive fields before engineers replicate them into a new environment.
The second quarter produces a controlled minimum data product. Operators compare its output with the existing decision, record disagreements and correct upstream quality. The release remains reversible, and finance establishes the baseline against which later benefit will be measured.
In the third quarter, the organization integrates the product into normal workflow. Roles, approval thresholds, alerts and fallback procedures become part of operating instructions. Training uses actual exceptions from the pilot, while access logs and recovery drills test whether control works outside a presentation.
The fourth quarter measures economic and risk outcomes, then decides whether to scale, correct or stop. Reusable definitions and controls enter the shared platform. A successful team transfers knowledge to the next domain instead of becoming a permanent central queue for every analytical request.
This sequence is not a universal calendar. A regulated safety process may need longer validation, while a low-risk commercial experiment can move faster. Its value lies in preserving the order: decision before collection, quality before automation, controlled use before scale, and measured evidence before declaring transformation.
Data orientation is an operating discipline
The 2025 Forbes analysis showed why technology alone cannot create a data-driven company. Strategy without quality produces confident errors; quality without shared meaning preserves silos; analytics without culture remains unused; access without security destroys trust.
The discipline also changes management meetings. Participants arrive with the same definitions and can spend less time disputing whose spreadsheet is correct. They examine assumptions, ranges and consequences, distinguish a temporary signal from a structural shift, and assign the next action to a named owner. A documented decision then feeds back into the system, allowing the company to learn whether its reasoning produced the intended operational result.
A mature company connects each dataset to a decision, each decision to an owner and each investment to a measured result. It prepares for exceptions, preserves human challenge and improves the system when evidence contradicts assumptions.
The competitive advantage is not possessing more information than everyone else. It is shortening the reliable path from an event to a coordinated response. That capability grows through repeated operating loops, and it remains valuable even as platforms, models and fashionable architecture names change.
ADI News
Leave a comment