Before comparing model-driven AI and data-driven AI, it is important to separate the two very different sources of intelligence. Model-driven AI is powered by large pre-trained models, reasoning patterns, domain logic, symbolic constraints, simulators, and reusable knowledge captured in the model or architecture. It is highly important and powerful now because it can generalise, explain, plan, code, summarise, and interact with users in natural language.
But many business problems cannot be solved by model intelligence alone. In forecasting, fraud detection, energy optimisation, clinical decision support, financial risk, customer personalisation, industrial monitoring, and many other use cases, the system must be grounded in real operational data. Without that grounding, a model-driven approach may appear confident while producing hallucinations, generic answers, or decisions that are impossible to verify.
Data-driven AI is often more precise for specific use cases because it learns from measured behaviour, transactions, sensor signals, outcomes, and domain-specific feedback.
In practice, the strongest solutions are frequently hybrid: model-driven AI provides reasoning and interaction, while data-driven AI provides accuracy, calibration, and evidence.
We sat down with Taras Firman to discuss why strong models still need strong data, and what that means for businesses building AI systems today.
Background & experience:
- More than a decade of experience spanning computer vision engineering, mathematical modelling, data science, and consulting.
- Focuses on Explainable AI, addressing the critical challenge of building trust in AI models for both academic and business environments.
- Mathematician by training with a Ph.D. in Mathematical Sciences and extensive cross-industry experience across retail, logistics, banking, healthcare, and bioinformatics.
1. Can you define model-driven AI in simple terms?
Taras Firman: Model-driven AI starts from an explicit model or structured representation of the world. This can be a large language model, a foundation model, a physics model, a rules engine, a simulation model, or a planning system. The system has some built-in structure that helps it reason beyond the raw data.
For example, an LLM can explain a financial report, a physics-based model can simulate power flows in an energy network, and a clinical knowledge graph can represent relationships between symptoms, diagnoses, medications, and contraindications. These are model-driven because the system relies heavily on a structured representation or a pre-trained model to make sense of the task.
2. How would you describe data-driven AI?
Taras: Data-driven AI starts from data. It learns patterns from transactions, sensor signals, claims, medical images, customer interactions, clickstreams, production logs, market prices, or historical outcomes. What makes data-driven AI powerful is its specificity — it can learn from data specific to a particular company, region, device, customer segment, hospital, factory, or portfolio.
A data-driven fraud model not only knows what fraud means in general. It learns which combinations of device fingerprint, transaction velocity, merchant category, geography, account age, and previous behaviour are suspicious in that exact environment. That is why it can be much more precise for operational decisions.
3. How important is model-driven AI for the future of enterprise systems?
Taras: Model-driven AI is absolutely one of the most important forces in enterprise technology right now. It gives us a new interface to software, a new way to reason over documents, a new way to generate code, and a new way to automate complex workflows. Large language models and foundation models are powerful because they already contain vast amounts of general knowledge and can adapt to many tasks without being trained from scratch.
But we need to be careful with the conclusion. Model-driven AI is not magic, and it is not a universal replacement for data-driven AI. A business problem is usually not just a question like “write an answer” or “summarise this document.” Very often, the business problem is: predict what will happen in our exact market, identify our exact customer risk, optimise our exact grid, detect anomalies in our exact machines, or recommend the next best action for our exact user. For those problems, general model knowledge is not enough. You need data.
4. Is it fair to say that model-driven AI is more general, while data-driven AI is more specific?
Taras: That is a useful simplification. Model-driven AI is often better at generalisation, language, explanation, planning, and abstract reasoning. Data-driven AI is often better at measured prediction, scoring, ranking, optimisation, anomaly detection, and calibration for a specific business process.
The mistake is treating one as a replacement for the other. Model-driven AI can help you understand a problem, create hypotheses, orchestrate tools, and communicate results. Data-driven AI helps you validate whether the hypothesis is true in the data and whether the decision works in production.
5. Why do many businesses believe that model-driven AI can solve most problems on its own?
Taras: Because the demos are impressive. You ask a model a complex question, and it produces a fluent answer. It can write a strategy, generate SQL, draft legal text, explain code, and simulate a conversation with an expert. This creates a feeling that the model “understands” the whole business.
But fluency is not the same as accuracy. A model can produce a confident, well-written answer even when it lacks the underlying data. Without access to current inventory, customer history, or risk exposure, it will either stay generic or fill in the gaps. This is where hallucinations appear — because it is completing a pattern with insufficient evidence.
6. Can you give a simple example of hallucination caused by missing data?
Taras: Imagine a customer asks, “Why was my insurance claim delayed?” A model-driven assistant can explain common reasons for claim delays: missing documents, validation issues, manual review, payment checks, or compliance flags. That sounds helpful, but it may not apply to this customer.
A data-driven system would check the actual claim record, timeline, missing forms, payment status, adjuster notes, and workflow events. Then it can say: “The claim is delayed because the repair invoice is missing, and the last reminder was sent on Monday.” That answer is grounded, precise, and actionable. Without the data, the model may only produce a confident guess.
7. Where is model-driven AI genuinely powerful, then?
Taras: It is powerful in tasks where general reasoning, language understanding, and knowledge transfer matter. Examples include document analysis, code generation, research assistance, knowledge management, and human-computer interaction.
In enterprise workflows, model-driven AI can act as a reasoning and orchestration layer. It can translate natural language into structured queries, call tools, summarise results, explain trade-offs, draft decisions, and guide users through complex processes. This is very valuable. The point is not that model-driven AI is weak. The point is that it must be connected to evidence.
8. Let us talk about energy. How would this difference appear in the energy domain?
Taras: Energy is a perfect domain for this comparison. A model-driven system can understand grid topology, power-flow constraints, asset types, safety rules, regulatory language, and maintenance procedures. It can help engineers ask questions like: “What assets are most critical in this substation?” or “Explain the operational risk if this transformer fails.”
But energy optimisation cannot rely only on a general model.
Load forecasting, renewable generation forecasting, predictive maintenance, energy trading, demand response, and grid balancing require data. Operating a smart grid means you need smart meter readings, weather data, historical consumption, equipment telemetry, and real-time grid conditions. If you ask a model to forecast tomorrow's load without this data, the answer may look reasonable but will not be operationally reliable.
A data-driven model can learn, for example, that a specific region experiences a morning peak on cold weekdays, that solar output drops sharply under certain cloud patterns, or that a turbine vibration signature often precedes a fault. This level of precision is achieved through data, not general intelligence alone.
9. Can model-driven AI still help in energy forecasting?
Taras: Yes, but usually as part of a bigger system. It can help select features, explain forecast drivers, generate what-if scenarios, document assumptions, and make the forecast understandable to operators. It can also combine technical documentation with real-time analytics, so engineers do not need to manually search across many systems.
The actual forecast must be based on time-series data, weather feeds, grid events, and historical demand. Model-driven AI enhances usability and reasoning. Data-driven AI delivers numerical accuracy.
10. What about financial services?
Taras: Data-driven AI is unavoidable in the finance industry. A model-driven AI assistant can interpret regulations, summarise contracts, explain portfolio exposure, generate client reports, or help analysts reason about market narratives. That is useful and powerful.
But credit scoring, fraud detection, anti-money laundering, risk modelling, algorithmic trading, churn prediction, and collections prioritisation depend on data. A credit decision must use borrower history, income signals, repayment behaviour, macroeconomic variables, and portfolio performance. A fraud system needs real-time transaction behaviour, device data, merchant data, and historical fraud labels.
A general AI model knows what fraud is and explains common indicators. But it doesn't know which specific patterns are dangerous in your bank today. Data-driven AI learns those patterns from your actual data and produces scores, thresholds, and alerts that can be measured against real outcomes.
11. Can a model-driven system make financial decisions directly?
Taras: It can support financial decisions, but it should not make high-stakes decisions based solely on language reasoning. For example, it can summarise why a loan application requires review, explain the meaning of a risk score, or generate a compliance-friendly explanation. But the underlying risk score should be derived from validated, data-driven models and established business rules.
In the finance industry, precision, auditability, calibration, and monitoring are essential. A persuasive model that is not validated against historical outcomes can lead to financial losses, regulatory issues, or unfair decisions.
12. Can you give an industrial or manufacturing example?
Taras: In the manufacturing industry, model-driven AI can explain maintenance manuals, help technicians troubleshoot, generate work instructions, and reason about production constraints. It can also help translate operator questions into queries over plant systems.
But predictive maintenance, quality inspection, defect prediction, process optimisation, and yield improvement are strongly data-driven. You need vibration signals, temperature, pressure, machine states, cycle times, images, quality labels, maintenance logs, and downtime records. A model can explain why overheating is dangerous, but a data-driven system can detect that Machine 17 is drifting from its normal vibration pattern and is likely to fail within a specific window.
13. What about retail and supply chain?
Taras: Retail industry is full of examples. A model-driven system can help category managers analyse reports, create marketing copy, explain assortment strategy, or answer questions about product performance. But demand forecasting, inventory optimisation, dynamic pricing, recommendation engines, promotion planning, and route optimisation depend on data.
A model may understand that holidays increase demand, but a data-driven forecast can quantify how much demand changes for a specific SKU, in a specific store, under a specific promotion, with specific weather and local events. That precision is what prevents stockouts, overstock, and margin loss.
14. What is the biggest business risk of relying only on model-driven AI?
Taras: The biggest risk is confident hallucination because the system can produce a convincing narrative that hides uncertainty. This is risky for businesses as people may convert a fluent answer into an operational decision.
There is also a second risk: generic automation dressed up as intelligent automation. A company may deploy an assistant that sounds intelligent but does not improve KPIs because it is not connected to the data that drives the business. It can answer questions, but it cannot optimise outcomes.
15. And what is the biggest risk of relying only on data-driven AI?
Taras: It is narrowness. A data-driven model may be very accurate for one task but fragile when the environment changes. It may not understand business context, policy, language, or user intent. It can give a score but not necessarily explain the decision in a way a human can use.
That is why model-driven AI is so important now. It gives data-driven systems a better interface, a better explanation layer, and a better orchestration mechanism.
16. So, data-driven AI is more precise, but model-driven AI is more flexible?
Taras: Very often, yes. Data-driven AI is usually stronger when the target is measurable, and the environment produces enough representative data. It is better for specific prediction and optimisation tasks.
Model-driven AI is stronger when the task requires reasoning across text, knowledge, processes, and tools. It is better for interaction, explanation, drafting, planning, and broad knowledge work. The best enterprise systems combine both strengths instead of forcing one approach to do everything.
17. How would you explain this to a CEO who thinks an LLM can replace all analytics?
Taras: I would say that an LLM can make analytics easier to access. It does not replace the need for analytics, but it can help people ask questions, generate hypotheses, interpret charts, and write summaries. The numbers still need to come from reliable data pipelines and governed metrics.
If the CEO asks, “Which customers are likely to churn next month?” the LLM should not guess. It should call a churn model trained on the company’s data, retrieve the scored customer segments, explain the top drivers, and recommend actions. The LLM is the interface and reasoning layer. The data-driven model is the precision engine.
18. What architecture do you recommend for enterprise use cases?
Taras: A practical architecture has four layers. First, a data layer: clean data, feature stores, and governed metrics. Second, there are data-driven models: forecasting, classification, and anomaly detection. Third, a model-driven layer: LLMs, agents, and planning logic. Fourth, a control layer: guardrails, monitoring, and human review where needed.
In this architecture, the LLM asks the right questions, retrieves the right evidence, reasons through the results, and communicates clearly. That is much more reliable than giving the LLM a prompt and expecting it to become the whole business brain.
19. How should teams decide whether a task is model-driven, data-driven, or hybrid?
Taras: A useful test is to ask: Does the task require general reasoning, or does it require specific evidence?
If the task is “explain this policy,” “draft a response,” “summarise this document,” or “help me plan,” model-driven AI may be enough. If the task is “predict demand,” “detect fraud,” “prioritise patients,” “optimise price,” or “identify equipment failure,” data-driven AI is essential. If the task requires both a decision and a human-friendly explanation, it is probably hybrid.
20. What about companies that do not have clean data yet?
Taras: They can still benefit from model-driven AI, but they should not confuse that with solving the core data problem. An LLM can help document data sources, generate data quality checks, accelerate SQL development, summarise data dictionaries, and help business users explore information. That is a great starting point.
But if the data is incomplete or inaccessible, the model cannot create operational truth by magic. Sooner or later, the company must invest in data quality, governance, integration, labelling, and monitoring. Otherwise, the AI will be no more than a polished interface built on unreliable foundations.
21. Does this mean data engineering becomes more important, not less?
Taras: Exactly. The more powerful the model-driven layer becomes, the more valuable and reliable the data becomes. If the AI assistant is connected to poor data, it will make poor decisions. If it is connected to high-quality data and validated models, it can scale the making of good decisions.
In other words, model-driven AI increases the return on good data infrastructure. It does not eliminate the need for it.
22. What is the key takeaway for businesses trying to adopt AI today?
Taras: Model-driven AI is highly important and powerful now, but it cannot solve every problem on its own.
Model-driven AI is highly important and powerful now, but it cannot solve every problem on its own. If a business expects it to replace all data-driven decision systems, the likely result is confident hallucination, generic advice, and weak operational performance.
Data-driven AI is often much more precise for specific use cases because it is grounded in the company's real data and measurable outcomes. The future is not model-driven versus data-driven; it is both. The future is hybrid: models that can reason and communicate, connected to data systems that can measure, predict, and validate.
In energy, finance, healthcare, customer experience, manufacturing, retail, and many other domains, the winning systems will be hybrid systems that combine model intelligence with data evidence.
Related insights
The breadth of knowledge and understanding that ELEKS has within its walls allows us to leverage that expertise to make superior deliverables for our customers. When you work with ELEKS, you are working with the top 1% of the aptitude and engineering excellence of the whole country.
Right from the start, we really liked ELEKS’ commitment and engagement. They came to us with their best people to try to understand our context, our business idea, and developed the first prototype with us. They were very professional and very customer oriented. I think, without ELEKS it probably would not have been possible to have such a successful product in such a short period of time.
ELEKS has been involved in the development of a number of our consumer-facing websites and mobile applications that allow our customers to easily track their shipments, get the information they need as well as stay in touch with us. We’ve appreciated the level of ELEKS’ expertise, responsiveness and attention to details.
