# Agentic data analysis: signals that become decisions. Agentic data analysis does not produce charts to archive: it finds signals in already available data, connects them to a decision and states where the model is not worth building. ## What agentic data analysis is It is a service that verifies where company data can reduce delays, waste, errors or risk. If the signal is missing, the useful outcome is knowing which project not to fund. ## Dashboards look back. Signals decide now. 1. Available data 2. Business question 3. Verified signal 4. Possible decision 5. Measurable or estimable value 6. Targeted new data collection ## Outputs built for decisions ### Executive Summary Main result, recommended decision, value at stake, limits and actions for the next 30, 90 or 180 days. ### Technical report Data used, controls, methods, metrics, reproducibility and evidence that the model beats a minimum benchmark. ### Action plan Low-risk pilot, responsibilities, timing, measures to observe and criteria to extend, change or stop. ### Data collection plan Which data to collect next, why, with what priority and which decision it would strengthen. ## Forms of value ### Recovered value Customers, orders, lots or bookings that can be saved before value is lost. ### Avoided cost Predictive projects not to fund when current data does not contain the needed signal. ### Organisational efficiency Resources reallocated to time slots, products, checks or processes that truly matter. ### Customer promise More credible deliveries, availability, timing and communication based on better estimates. ### Data governance Less generic data collection, more closely tied to concrete decisions. ## When the signal is not only in data Many AI opportunities emerge where data, documents, processes and operating decisions meet. To browse possible applications, examples and need signals: [AI applications atlas for companies](ai-applications-atlas.html). To choose between data analysis, consulting, training or technical software: https://ar-tik.com/en/ai-business-faq.md ## First question: which decision must improve? - Recurring decision: order, plan, contact, check. - Visible or suspected cost: waste, delays, returns, failures, penalties. - Available data: transactions, sensors, orders, tickets, records. - Possible action: call, check, process change, different priority. ## Anonymised examples of signals and decisions Each story explains the business problem, which operating data enters the analysis, which signal emerges and which decision can be supported. These are not standard promises; they show how public, anonymised or realistic datasets, without recognisable client projects, can become verifiable action. ### 1. Hospitality: Understand which bookings are likely to fall through A hotel can read risk at booking time: on more than 119,000 bookings, the system catches more than eight cancellations out of ten. The story is simple: management usually sees cancellations when the loss has already happened. Agentic data analysis looks earlier, using signals already present before the stay, such as lead time, payment terms and customer history. For a hotel, residence or hospitality group, booking-engine data becomes a commercial priority list. Fragile bookings can be confirmed, contacted or managed with different conditions. - Useful signal: The system separates solid bookings from bookings that deserve preventive action. - Possible decision: Confirm, contact or protect the most exposed bookings first. - Useful data next: Call-back outcome, recovered value and customer response. - Limit to state: The analysis does not eliminate cancellations; it helps decide where to act in time. ### 2. Food delivery: Discover where a lost order really begins When the kitchen does not confirm the order as ready, the risk of losing it rises to 35.7%. At first the problem looks like final delivery: an order does not arrive, the customer complains, the restaurant loses trust. The analysis shows that the signal appears earlier, inside the kitchen workflow. For a delivery platform or restaurant chain, the question changes: not “which rider is late?”, but “which order is leaving the process before it can be delivered?”. - Useful signal: A missing intermediate confirmation becomes an operational warning. - Possible decision: Trigger a prompt, reassignment or customer communication immediately. - Useful data next: Recorded cause, order recovery and service failure cost. - Limit to state: The model works when intermediate order states are recorded well. ### 3. Last mile: Give customers a more credible delivery window Average delivery-time error moves from about 41 minutes to about 17 minutes. For urban logistics companies, the issue is not only delivering faster. It is promising a realistic arrival time, so customers wait less, support receives fewer calls and fleet coordination improves. The analysis starts from orders and historical timing, but does not stop at the average. It finds recurring conditions that make a delivery slower or faster and turns them into a more useful forecast. - Useful signal: A more reliable arrival window for each delivery. - Possible decision: Update customer messages, operational priorities and fleet planning. - Useful data next: Complaints, avoided calls and manual interventions by operations. - Limit to state: It does not promise faster delivery; it promises more credible estimates. ### 4. Energy: Forecast tomorrow’s demand more reliably The forecast reduces error by 77% against the reference rule. An energy operator or large consumer must decide in advance how much energy to buy, hedge or reserve. If the forecast is too cautious, resources are locked up; if it is too low, the company faces correction costs. The analysis reads hourly consumption history and builds an expected profile for the next day. The output is not a chart to archive, but decision support for energy planning. - Useful signal: An expected hourly profile that is more reliable than the comparison rule. - Possible decision: Buy, hedge or plan capacity with less defensive margin. - Useful data next: Prices, imbalance costs and procurement rules. - Limit to state: Economic savings must be calculated with the real contract numbers. ### 5. Restaurants: Prepare for the week that is coming Revenue forecasting improves by 24% versus the “same as last week” rule. A restaurant decides every week how much fresh stock to buy and how many people to schedule. If the decision is based only on intuition, weak days create waste and strong days create service pressure. The analysis starts from revenue history and recognises the real rhythm of the venue. The forecast becomes a practical tool for kitchen, dining room and purchasing before demand arrives. - Useful signal: A future revenue estimate that is stronger than the empirical rule. - Possible decision: Use the forecast beside purchasing, preparation and shift decisions. - Useful data next: Real waste, missed sales and staffing cost. - Limit to state: Value appears only if the forecast changes operating decisions. ### 6. Food retail: Save perishable lots before they become waste The riskiest lots waste almost three times as much as the safest ones. In a supermarket or food supply chain, waste does not appear all at once. It starts with small signals: packaging, handling, cold chain, arrival timing and sales priority. The analysis reads these signals when a lot enters the process and creates a risk ranking. The point is not to predict every loss, but to decide which lots to check, rotate or discount early. - Useful signal: A list of lots that deserve attention before deterioration is visible. - Possible decision: Focus checks, rotations and preventive markdowns on the most exposed lots. - Useful data next: Value saved, waste reason and margin after intervention. - Limit to state: Not all waste is predictable; the goal is to use preventive action better. ### 7. Banking: Recognise customers who are about to leave The system recognises about three at-risk customers out of four. A bank can notice a customer leaving when the account is already lost, or it can read earlier signs that the relationship is cooling. The analysis separates generic risk from the commercial lever that can be acted on. The useful story is not “this customer will leave”, but “this customer shows inactivity signals and can be reactivated with a targeted action”. That difference matters when building credible campaigns. - Useful signal: A contact priority based on behaviour and churn risk. - Possible decision: Launch targeted reactivation campaigns, not the same message for everyone. - Useful data next: Behaviour history, contacts made and retained value. - Limit to state: Recognising risk today does not always mean predicting it far in advance. ### 8. Quick service restaurants: See which menu items and time slots really carry the business A few dayparts and a few menu items generate almost three quarters of revenue. In a quick service chain, the problem is not only selling more. It is understanding where revenue really comes from: which time slots need staff, which products deserve stock, which items occupy space without carrying weight. Descriptive analysis becomes an operating story: the menu is not all equal and the day does not weigh all the same. This helps decide shifts, stock and promotions with less impression and more evidence. - Useful signal: A map of the products and moments that sustain the economics. - Possible decision: Realign staff, purchasing, promotions and menu review. - Useful data next: Margin by item, preparation time and stock-outs. - Limit to state: This is not a forecast; it is an operating priority to complete with margin data. ### 9. Winery: Protect the most promising wine lots early With lab data, the system recognises 87% of high-tier lots. A winery already collects chemical measures during production. Often those data remain technical, separated from commercial choices and lot destination decisions. The analysis shows that these signals can help identify premium-potential lots early. It does not replace the winemaker’s judgement; it helps protect value before blending decisions disperse it. - Useful signal: An early classification of lots that deserve more attention. - Possible decision: Prioritise tastings, ageing choices and commercial destination for promising lots. - Useful data next: Final destination, realised value and qualitative judgement. - Limit to state: The model supports technical judgement; it does not replace it. ### 10. Industrial maintenance: Use sensors to recognise a failure in progress With available sensors, the analysis recognises about 84% of observed failures. In a factory, a failure is not only a technical event: it stops people, orders and production capacity. Many machines already have sensors, but signals remain scattered or are read too late. The analysis creates a warning when machine behaviour resembles failure situations already seen. It is useful when it immediately triggers a work order, inspection or field check. - Useful signal: An operational alert when the machine shows patterns compatible with a failure. - Possible decision: Connect the warning to maintenance, escalation and avoided downtime checks. - Useful data next: Intervention time, downtime cost and parts used. - Limit to state: Recognising a failure in progress is not the same as predicting it weeks ahead. ### 11. Fashion returns: Avoid a model when the right data is missing In catalogue data, the best variable explains less than 2% of returns. A fashion e-commerce business may want to predict which items will be returned. The temptation is to use convenient data: category, price, colour and product page information. The analysis shows that those data are not enough. This is a good managerial finding: it avoids a fragile investment and points to the information that matters, such as fit, customer history and return reason. - Useful signal: A clear negative verdict: the signal does not live in the product catalogue. - Possible decision: Do not fund the predictive model before changing data collection. - Useful data next: Fit, return reason, measurements and customer history. - Limit to state: The “no” does not close the problem; it shows which data would make it addressable. ### 12. Logistics: Know when a model cannot predict delay With planning data only, the best model remains close to a random choice. A logistics operator wants to know before departure which deliveries will be late. But if it uses only planning data, it is looking at an incomplete picture: the events that happen during the trip are missing. The analysis avoids forcing a weak forecast. The better decision is to collect travel events, stops, unloading, weather and operating anomalies before building a more ambitious model. - Useful signal: No useful signal in the data available before departure. - Possible decision: Stop the model and design data collection around real transport events. - Useful data next: Stops, unloading, traffic, weather, exceptions and delay cost. - Limit to state: A more complex algorithm cannot create information the process does not record. ### 13. Compliance: Put checks first where risk is higher With the same number of checks, the ranking catches more severe cases. A control body or compliance function always has more cases to check than it can handle immediately. The question is not to run infinite checks, but to choose the right order. The analysis uses inspection history to build a priority list. Checks remain human, but the agenda is ordered to increase the chance of finding the most serious cases first. - Useful signal: A risk ranking for scheduling checks and follow-up. - Possible decision: Order inspections, audits or internal controls without increasing budget. - Useful data next: Check outcome, recurrence, severity and time back to compliance. - Limit to state: The model does not decide sanctions; it helps order priorities. ### 14. Manufacturing: See when a machine consumes without producing enough value Operating profiles show an 11.4-point efficiency gap. In production, average consumption often hides different stories. The same machine can work in more or less efficient ways, but raw energy data does not immediately explain why. The analysis groups machine behaviours and shows which profiles deserve comparison. Before buying new sensors or equipment, the company can ask which operating conditions separate efficient work from waste. - Useful signal: A map of operating profiles, not just average consumption. - Possible decision: Compare best and worst profiles and start a waste-reduction pilot. - Useful data next: Energy cost, machine hours, production and operating settings. - Limit to state: Euro value should be estimated only when consumption and production are linked. ## Frequently asked questions ### Does agentic data analysis replace Business Intelligence? No. BI monitors known indicators; agentic analysis diagnoses causes, searches for hidden signals and connects results to decisions. ### Is a perfect data warehouse required? No. The first value can be verifying whether existing data is fit for purpose, what its limits are and which data to collect next. ### What if there is no signal? The method states the negative verdict and indicates which investment to avoid or which data collection to start before funding a model.