AI use cases by industry: what actually works in logistics, retail, hospitality and manufacturing

Search for AI use cases and you'll find one of two things: lists of a hundred ideas with no context, or industry decks full of numbers nobody can verify. Neither helps you decide. What you need is more boring and more useful: what is actually being deployed today in a business like yours, on what data, what changes in day-to-day operations, and what has to be in place before you start.

This article walks through five sectors — logistics, retail, hospitality, manufacturing and professional services — with cases we've seen work in mid-sized Spanish companies. No promised improvement percentages: the return depends on your volume, your margin and how tidy your data is, and anyone quoting you a generic figure before looking at your operation is selling you a slide.

What counts as an AI use case, and what doesn't?

A use case is a specific decision or task, with an owner, a frequency and a known current cost, that AI does better, faster or cheaper. It has four mandatory ingredients:

• A decision that changes. How much to order, who to call, what to review first, what to answer.
• An input that already exists. Not "we'd have to start recording it", but something sitting today in the ERP, the CRM, a sensor or an inbox.
• A person accountable. Whoever uses the output and answers for it when it gets things wrong.
• A measurable current alternative. What it costs to do it manually today, in hours or in errors.

If any of the four is missing, it isn't a use case: it's an idea. Ideas are free and they don't reach production. That's exactly why so many projects stall at the demo stage, something we cover in why AI pilots fail.

The best first use case isn't the most impressive one. It's the one that leaves the plumbing installed — access, cleanup, integration — that makes the next five cheaper.

Which AI use cases work in logistics and transport?

Logistics has an unusual advantage: almost everything is already digital and timestamped. Delivery notes, routes, loading times and incidents leave a trail you can use without building anything new.

Demand forecasting by SKU and warehouse. The classic case, and the one that moves the most weight. It doesn't replace the buyer: it gives them a starting point for 4,000 SKUs instead of the 40 they have time to look at. The bar to beat is your current forecast, not random guessing; we explain it in predictive analytics in practice.

Automated reading of transport documents. Signed delivery notes, CMRs, carrier invoices. Field extraction, matching against the order, and an alert only when there's a discrepancy. One of the lowest-risk cases: if extraction fails, the document falls back into the usual manual queue.

Incident triage. Automatically classifying claims and delays by likelihood of real cost, so the team starts with what hurts. It's pure scoring, built on the incident history you already have.

Assisted route optimisation. Careful here: much of this is classical operations research, not AI, and it works very well. A vendor labelling it artificial intelligence doesn't make it better or worse; it only inflates the invoice if they charge you for the label.

What about retail and distribution?

In retail the problem is rarely a lack of data — it's that the data is split across POS, e-commerce, the warehouse and a couple of spreadsheets only one person understands.

Replenishment and assortment per store. What to send to each location based on its own pattern, not the chain average. It needs ticket-level sales and a product master that doesn't hold three names for the same item.

Shrinkage and till anomaly detection. Odd patterns in returns, voids or discounts. AI isn't accusing anyone here: it flags what to review, and a person with judgement does the reviewing.

Assisted customer service. Order status, returns, availability. It works when it's wired into the real system; otherwise it's an expensive FAQ. Where to draw the automation line is covered in chatbots for business.

Product content enrichment. Generating descriptions, attributes and tags from technical sheets. An unglamorous case with a fast return when you have thousands of SKUs and a marketing team of three.

Visual checks on shelf or in the warehouse. Out-of-stocks, placement, counting. Heavily dependent on camera installation quality; we break it down in computer vision in business.

What does AI do in hospitality and hotels?

Hotels have sharp seasonality, many channels and a lot of free text. That combination opens up specific cases.

Occupancy forecasting and rate support. Models fed with your own history, an events calendar and booking pace. The decision stays human: the model supplies the curve, revenue management supplies the commercial judgement.

Shift planning against forecast demand. Kitchen, housekeeping and front desk sized by expected occupancy rather than by habit. One of the cases that shows up soonest in the P&L, because payroll is the big line.

Systematic reading of reviews and surveys. Thousands of comments in several languages turned into recurring themes by property and period. It stops being the manager's anecdotal read and becomes a comparable indicator.

Answering frequent pre-arrival questions. Opening hours, parking, transfers, pets. It takes the repetitive share off the front desk and frees time for what genuinely needs a person.

Which AI use cases exist in manufacturing?

Manufacturing is where data is richest and hardest at the same time: sensors generating plenty of volume but rarely linked to the quality outcome or to the actual downtime event.

Visual quality inspection. In-line defect detection with a camera. When the defect is visible and repeatable, it's one of the clearest returns available. When the defect is spotted by a veteran operator through sound, it's a different conversation.

Predictive maintenance. Anticipating failure from machine signals. It demands something many plants don't have: a properly recorded, properly dated failure history. Without it there's no label to learn from and the project settles into monitoring — useful, but not prediction.

Process parameter tuning. Recommending setpoints from previous batches and their quality outcome. Delicate ground: always with hard safety limits and the operator deciding.

Production scheduling. Sequencing orders around changeovers, availability and delivery commitments. A mix of optimisation and forecasting.

A quick comparison of where each sector tends to start:

| Sector | Usual first case | Essential data | Typical risk | |---|---|---|---| | Logistics | Demand forecasting | Sales history by SKU and date | Inconsistent product master | | Retail | Per-store replenishment | Ticket-level sales | Data split across POS and e-commerce | | Hospitality | Occupancy forecasting | Booking history by channel and date | Seasonality with thin history | | Manufacturing | Visual inspection | Labelled defect images | Lighting and camera position | | Professional services | Document extraction | Digitised, organised documents | Confidentiality and legal basis |

And in law firms and professional services?

Accounting firms, law firms, engineering practices and consultancies share a pattern: the product is qualified people's time, and a share of that time goes into mechanical work.

Document extraction and classification. Invoices, payroll, deeds, contracts. Pulling fields, organising files and flagging what's missing. The case with the most immediate mileage, detailed in AI for accounting firms.

Assisted contract review. Locating relevant clauses, comparing against a template, flagging omissions. It never replaces professional judgement: it shortens the reading and prevents the silly oversight.

Search across the firm's own knowledge. Asking in plain language about files, reports and internal precedent. It needs serious permission control, because not everyone should see everything.

Draft preparation. Repetitive filings, progress reports, meeting summaries. With mandatory human review and a record of what was generated and who validated it.

In this sector the personal-data conversation isn't a formality: if end-client data is involved, legal basis, purpose and retention have to be settled before anything is touched, and you need to decide whether the flow can work on treated data instead of originals.

How do you pick the first use case out of all these?

Ranking the list matters more than lengthening it. The criteria we use have four axes and can be applied in an afternoon with a sheet of paper:

1. Frequency. How many times a month it happens. A case that occurs five times a year won't pay for the integration, however expensive each occurrence is. 2. Current cost. Hours of qualified people, errors you pay for, delays that cost customers. 3. Data availability. Is it in a system, with history and a known outcome, or do you have to start recording it? 4. Error tolerance. What happens if it's wrong one time in twenty. If the answer is "nothing serious, there's a review step", go ahead. If it's "a fine" or "a line stoppage", that isn't your first case.

Run those four axes and an unglamorous case almost always wins — documents, forecasting, triage — while the one the board was most excited about loses. That's a good sign. The flashy case comes later, once the data path exists and the organisation has learned to work with probabilistic outputs.

One warning about getting the house in order: if axis 3 reveals that no case has its data available, the project isn't an AI project yet — it's a data strategy project. Skip that step and you'll pay twice.

Frequently asked questions

How many AI use cases should we tackle at once?

One. Two at most, if they're independent and involve different teams. Opening five fronts with the same people guarantees that none reaches production and that the year ends with five stalled pilots and zero changed processes.

Are use cases from other industries useful as a reference?

As inspiration yes, as a plan no. What transfers well across sectors is the problem family — forecasting, classification, extraction, anomalies — not the specific implementation. A hotel's occupancy forecast and a distributor's demand forecast share a technique and almost nothing else.

Do we need to replace the ERP or CRM before applying AI?

Almost never. The usual approach is to integrate with what's there through controlled reads and return the result into the same system the team already uses. Replacing an ERP is a two-year project that tends to postpone any improvement indefinitely; the sensible order is the other way round. More on this in AI process automation.

When does an agent make sense instead of ordinary automation?

When the task requires deciding the next step based on the previous result and querying several systems along the way. If the flow is fixed and known, classical automation is cheaper, more stable and easier to audit. The difference, with examples, in AI agents for business.

If you've got this far with two or three cases circling in your head and no certainty about which comes first, that's exactly what the audit sorts out: the cases in your operation ranked by effort and return, with the real state of your data alongside, and a verdict on what's viable this year and what isn't. And if you'd rather pressure-test it in half an hour before moving anything, let's talk.