Most AI projects don't fail on the model. They fail on the three things around it: the state of the data, whether the process is stable enough to automate, and whether anyone in the business will act on the prediction. Get those right and the rest is engineering.
In short: how AI is transforming industries comes down to three repeatable moves: forecasting demand and failure, automating inspection and paperwork, and personalising what each customer sees. Sixteen sectors are covered below, each with the applications working now, who's shipping them, where we've built them ourselves, and a readiness score across data maturity, process repeatability, regulatory load and integration cost. In most sectors, payback lands in six to eighteen months.
Before judging how AI is reshaping business, agree on the words. Artificial intelligence is any software doing humanlike work: learning, planning, problem-solving. That's a wide net, so narrow it to the three families that do most of the commercial lifting.
Machine learning finds patterns in historical data to predict what you don't yet know: next month's demand, this machine's next failure, this loan's likelihood of default.
Computer vision interprets images and video, grading a weld, spotting a crop disease, tracking a basket through a shop.
Natural language processing (NLP) reads and generates text and speech. It sits behind chatbots, document processing and search.
Two more recur below. Predictive maintenance means servicing a machine when the data says failure is near, not on a fixed calendar. Lidar is a laser sensor that builds a 3D map of a vehicle's surroundings. Our technology guides cover each in depth.
Plenty of businesses are still wary of AI, and of what it means for the future of work. So rather than argue the abstraction, this guide walks through what AI is doing, sector by sector, and where we've done it.
The adoption numbers have shifted sharply. McKinsey's State of AI 2025 survey found 88% of organisations now use AI in at least one business function, up from 78% a year earlier. But the revealing figure sits underneath: only about a third have actually scaled AI across the enterprise. The rest are stuck in pilot purgatory, running experiments that never reach production.
That gap is the whole story, and it matches what we see. Across the software and data projects we've delivered at Imaginary Cloud, the models are rarely the hard part. Take Environmental Intellect: we built the machine-learning platform behind their energy product, and the pattern was textbook. The model was tractable; the effort, and the value, sat in the data pipeline and in getting predictions into the customer's workflow. Ei's clients now save over $200,000 in contractor costs. The lesson repeats on every engagement below: buy or build the model, and you're a third of the way there.
Three observations run through every section.
The pilot that works has a named owner in operations. Projects sponsored only from IT stall at handover, because nobody in the business has agreed to change how they work.
Data readiness sets the timeline, not model choice. Complete, consistently labelled records mean a first useful model in weeks. Records scattered across spreadsheets and legacy systems mean most of the budget goes on the pipeline first, which is the bulk of what a digital transformation programme actually involves.
Narrow beats ambitious. One decision, one dataset, one measurable outcome reaches production. A platform-wide "AI strategy" usually doesn't.
Rather than list industries alphabetically and call it analysis, we score each on the four factors that decide whether a project reaches production. It's the grid we reach for when a client asks whether their sector is ready.
Where the ratings come from matters, because they carry weight below. They're our own judgement, drawn from projects we've delivered across these sectors, not a published benchmark. The two-thirds/one-third budget split and the six-to-eighteen-month payback window rest on the same basis. Treat the grid as a sequencing tool, not a measurement, and expect two companies in one sector to differ more than the sectors do.
.webp)
Read it as a sequencing tool. High data maturity and repeatability with low regulatory load (retail, travel, logistics, technology) pays back fastest. High regulatory load with high integration cost (healthcare, energy, government, transport) is where the engineering is sound but the approval path sets the timeline.
Each section names the applications working today, flags who's shipping them, points to where we've delivered, and calls out the one factor that decides how fast a project reaches production. Where it matters, it also says when not to reach for AI.
By forecasting production, farmers improve yields and plan ahead; AI also tracks water and electricity use to cut waste. Computer vision detects crop disease early and stops it spreading.
Binding factor: data maturity. Yield models need several seasons of field-level records to be worth anything. Growers holding only farm-level totals should instrument the fields first, train a model second.
Who's shipping it (buy signal): Cropin, Blue River Technology, FarmWise, mature enough to buy; build only the yield model on your own data. We've worked in agritech ourselves, rebuilding GrainFox's product experience, a reminder that adoption, not algorithms, is often the real constraint on the ground.
Where enough project history exists, AI estimates build periods and costs and predicts a new project's budget and timeline within a usable margin. Analysing that history also exposes how a team is really performing, and flags issues before they land.
Binding factor: data maturity. Most contractors hold history in documents, not structured records, so phase one is data recovery, not modelling. If your last ten projects live in PDFs and email threads, don't start with a model.
Who's shipping it: Caidio, Kwant.ai, AirWorks. At the giga-project end, we're delivering AI and software for NEOM, where the data foundations get built alongside the buildings.
The highest-value application is the one nobody photographs: taking repetitive administration off teachers, from marking and filing to progress reports. The data already sits in the learning management system. Enrolment planning comes second: institutions plan intake against admissions and outcomes history they already hold.
Binding factor: regulatory load, as student data protection. Anything touching individual records needs a lawful basis and an audit trail before a model sees the data.
Who's shipping it: Carnegie Learning, Cognii, Century Tech. We designed the product experience for ELSA, an AI language-learning app that now ranks among the top AI tools in its category, a reminder that in edtech the model is only as useful as the interface wrapped around it.
In energy, AI estimates the emissions a system produces over a period, forecasts short-term demand accurately, and, powered by computer vision and analytics, detects component defects before a failure in service costs real money.
Binding factor: integration cost. Grid and plant systems are operational technology with long change cycles, so getting a prediction into the control room is usually harder than producing it.
Who's shipping it, and where we've built it: we built the machine-learning platform for Environmental Intellect, whose clients save over $200,000 in contractor costs. Alongside vendors like Google DeepMind, Anova and Siemens, it's a working example of the sector's central trade-off: strong models, hard integration.
Finance has moved fastest of all; many processes are now substantially automated. AI supports human judgement in credit, weighing borrower and contract characteristics, and in portfolio management, choosing uncorrelated positions and flagging movements worth acting on.
Binding factor: regulatory load. A credit model that can't explain its decision isn't deployable. The EU AI Act still classifies creditworthiness assessment of individuals as high-risk (Annex III, 5(b)), but the timeline moved: the Digital Omnibus, in force since 27 July 2026, pushed high-risk obligations for standalone Annex III systems from August 2026 to 2 December 2027. That's a reprieve, not a repeal, and UK firms sit under a different regime again (no equivalent statute; an ICO-led, pro-innovation stance). Either way, do the explainability work in your first sprint; it's good practice long before it's law.
Who's shipping it, and where we've built it: we've delivered financial products for clients that include BNP Paribas, EY and Sage, and built platforms such as FundSpace and Confinze. Vendors worth watching: Socure, Scienaptic AI, AlphaSense.
The clearest public-sector application is socioeconomic modelling: employment, productivity, trade, housing, health and education access, income distribution, all interconnected, all needing computational help. The decision that changes is which initiative gets funded, and at what scale. The second is infrastructure operation: transit, siting hospitals and schools, utilities, network load-balancing, each a scheduling or allocation decision with years of history behind it.
Binding factor: integration cost, compounded by procurement. Public systems are long-lived and rarely built to accept an external prediction, so the route in sets the timeline more than the model does.
Who's shipping it, and where we've built it: this is home ground. We built Eurofound's socio-economic analytical (MPI) platform, the exact "model interconnected distributions to inform funding" problem described above, and delivered a front end covering 280+ policy initiatives in six weeks. Platforms in use elsewhere: IBM Watson, Amazon Machine Learning, Google Vertex AI.
Across healthtech, AI runs from diagnosis to drug development. The hard part of diagnosis is that many signs point to several conditions at once, so ruling conditions out is slow, and only the most experienced clinicians reliably handle the most complicated cases.
Binding factor: regulatory load, the heaviest in the grid. Diagnostic software is a regulated medical device in most markets: the FDA maintains a public list of AI-enabled devices it has cleared. The model rarely delays the project; validation and clinical governance do. When not to reach for AI: if you can't resource the validation path, don't start the model.
Who's shipping it, and where we've built it: we built Aurora Analytica's AI-powered clinical-trials decision engine, and have delivered healthtech for Thermo Fisher, VestaConnect and Jinga Life. Vendors worth watching: Komodo Health, Corti, Ezra.
AI helps decide which suppliers and vehicles to use, which routes to take, when to dispatch, how to pack, where to site distribution centres and how to sort. An operation that large is broken into sub-problems and solved piece by piece.
Binding factor: integration cost, moderate rather than severe. The data is good and decisions repeat daily, so the work is wiring an optimiser into the transport management system, not building the optimiser. That's why logistics pays back faster than its complexity suggests.
Who's shipping it, and where we've built it: we've delivered software at national postal and logistics scale for CTT. Vendors worth watching: Uptake, Symbotic, HAVI.
Manufacturing pairs complex processes with thin margins, so small gains become significant savings, most of them in quality control and automation. Computer vision detects imperfections reliably and diverts affected products for manual review. And predictive maintenance monitors machine data in real time, servicing a machine before it fails rather than on a fixed calendar. The US Department of Energy's guidance sets out the cost case for condition-based servicing.
Binding factor: integration cost. Shop-floor systems and PLCs were never designed to accept a model's output, so the link between prediction and work order is where projects stall. For teams modernising the whole line, this is Industry 4.0 territory.
Who's shipping it, and where we've built it: we built TrustPortal, an operations-automation platform, and Nextbitt, for asset maintenance. Vendors worth watching: Rockwell Automation, Nvidia, and Environmental Intellect, above.
In proptech, automate valuations and sellers list more readily; build digital duplicates and buyers narrow their preferences before viewing, which means fewer viewings and more parallel deals. Document processing and immersive models round it out.
Binding factor: data maturity. Automated-valuation quality tracks the completeness of local transaction records, which is why the same model is reliable in one market and useless in the next.
Who's shipping it, and where we've built it: we've delivered proptech for RE/MAX and Invisible Homes. Vendors worth watching: Compass, Zillow, Redfin.
Retail has more going on than store optimisation: the obvious plays are e-commerce and personalised recommendations, now joined by virtual try-on, AR and voice or visual search. Physical retail is changing too, though not as first advertised: cashier-free formats proved expensive to run. Amazon removed Just Walk Out from its US Fresh grocery stores in 2024, replacing it with scan-as-you-go Dash Carts, and now sells the technology to third-party venues rather than running it in its own big-box grocery. The lesson: checkout-free ROI is format-dependent, strongest in small, high-throughput stores.
Binding factor: barely a constraint. Clean transaction data, daily repetition, light regulation outside personal data. That's why retail is the fastest sector to a first result.
Who's shipping it, and where we've built it: we've digitalised retail operations, including UNUA's fashion business. Vendors worth watching: Intel, Bloomreach, Pinterest.
Every company protects confidential information, and AI helps prevent unauthorised access. Standard filtering catches most malicious email; the attempts that get through are the ones needing another layer. Bots are a large share of internet traffic, so detecting suspicious patterns keeps systems running. The UK's National Cyber Security Centre publishes guidance on both defensive and offensive uses of AI.
Binding factor: process repeatability, in an unusual sense, because your attacker changes the process on purpose. Security models decay faster than any others here and need retraining as a standing cost, not a project line.
Who's shipping it: Cynet, Trellix, Qualys.
Analysing match video is standard. Object detection locates players and the ball frame by frame; pose estimation tracks how a body moves. Together they surface patterns across games, feeding forecasts, scouting and recruitment.
Binding factor: data maturity, usually bought rather than built. Most clubs licence tracking data, so the competitive question is what you do with a feed your rivals also have.
Who's shipping it, and where we've built it: we've delivered sports and fitness platforms including obé and Portugal's national Desporto Escolar programme. Vendors worth watching: Run Motion, Wingfield, Hawk-Eye.
AI came out of IT, so it finds most of its applications there, and in telecoms. It draws users in through chatbots and personalised products, and usage analysis shows where an app needs work.
Binding factor: none of the four, which is precisely the risk. Clean data, repeatable processes, light regulation, low integration cost: technology firms ship AI features faster than they decide whether a feature is worth shipping. Here the discipline has to come from product, not engineering, and it's where our teams spend most of their time, from Sedna in maritime comms to Advania (AI-assisted IT refurbishment) and Elephants Don't Forget. Much of it runs through our AI-enabled custom development work.
Vendors worth watching: DeepL, UiPath, NIQ.
Autonomous driving splits into perception, control and navigation. A control algorithm activates pedals and steering from a 3D representation built by computer vision, lidar and maps, with GPS routing. Some implementations skip the human-engineered step and predict control actions straight from video. Autonomy promises longer-running fleets, lower costs, and, as behaviour becomes more predictable, fewer accidents. Regulators already publish safety data, including the US NHTSA's standing general order on crash reporting for automated driving systems.
Binding factor: regulatory load. Perception and control are engineering problems with known approaches. Approval to operate on public roads is not, and it sets the timeline.
Who's shipping it, and where we've built it: we've delivered software and design for the airline TAP. Vendors worth watching: Tesla, Zensors, Uber.
The most valuable application is the recommendation engine, because the decision it changes is commercial: which itinerary, hotel or fare a customer sees first. An online travel agency already holds the history a recommender needs, which makes it the fastest project in the sector to a measurable result. Route optimisation, security screening and automated first-line responses follow the same pattern.
Binding factor: barely a constraint here either. Clean booking data and light regulation put travel alongside retail at the fast end of the grid.
Who's shipping it, and where we've built it: we helped Travel Wifi become a market leader through UX. Vendors worth watching: Skyscanner, Hopper, BagsID.
The sector sections describe what's possible. This one describes what it takes, the question a CEO, CTO or COO actually asks. As with the grid, these are our observations from delivery, not a published benchmark.
What has to be true before you start. Three to five years of usable historical data for whatever you want to predict. One process, named, with a person who owns it. A defined decision that changes when the prediction arrives. If nothing changes, the model is a dashboard, not a system.
Where the budget goes. Think of your data as water: worth nothing in a puddle, so most of the work is plumbing, meaning pipes to carry it, filters to remove what doesn't belong, and a tap the business can turn. In most engagements the split runs roughly two-thirds data engineering and integration, one-third modelling. On the Environmental Intellect build, that ratio is exactly why the value landed: the model was the easy third. Teams that budget for the model and treat the plumbing as an afterthought are the ones that run over.
How long to value. A narrow, well-scoped project on clean data produces a first result in six to twelve weeks and measurable value inside six to eighteen months. Regulated decisions in healthcare, finance, energy and government add an approval path measured in quarters, not sprints.
The failure modes, in order of frequency. Data that turns out incomplete or inconsistently labelled. A process too unstable to learn. No owner in operations, so nothing changes at handover. A model that performs in testing but has no route into the system where work happens.
Build or buy. For a common problem with a mature vendor market, such as document processing, fraud screening or demand forecasting, buy, and spend the saved effort on integration. Build where the model touches something proprietary: your pricing, your routing, your equipment, your customer data. Our applied AI and machine learning team does both.
Retail, finance, logistics and technology are furthest along: mature data, repeatable processes, low approval burden. Healthcare, energy and government have equally strong use cases but heavier regulatory loads, so deployments move at the speed of validation rather than engineering.
Three repeatable moves: forecasting demand, failure, risk and yield; automating perception and paperwork; and personalising output. Almost every application here is one of those three applied to a sector's data.
A narrow project on clean data produces a first usable result in six to twelve weeks and measurable value within six to eighteen months. Regulated decisions run in quarters. Projects that begin with data cleanup should expect the longer end.
Less in modelling than most budgets assume, more in data engineering: roughly two-thirds on pipelines, integration and monitoring, one-third on the model. Costs continue after launch, because models drift and need retraining.
Historical records covering what you want to predict, ideally three to five years, complete, consistently labelled, and accessible without a manual export. If your data is spread across spreadsheets and legacy systems, that consolidation is the project's first phase and should be budgeted as such.
Buy for common problems with a mature vendor market (document processing, fraud screening, standard forecasting) and spend the saved effort on integration. Build where the model touches something proprietary and competitive.
The pattern is task-level, not role-level. Marking, inspection, filing, first-line triage and routine screening get absorbed. Judgement, exception handling and relationships don't. The roles that change most are those where most of the working day is repeatable, documented tasks.
Data readiness, followed by ownership. It's a finding McKinsey's 2025 survey echoes at scale, with most organisations stuck in pilots rather than production. Model choice is rarely the cause.
The sector sections and the grid point the same way: the technology is not the constraint. Data maturity, process repeatability, regulatory load and integration cost decide whether a project reaches production, and they vary far more between two companies in one sector than between the sectors themselves.
So the useful next step isn't an AI strategy. It's picking one decision your business makes repeatedly, checking whether you hold clean history for it, and finding the person who would act differently if a prediction arrived. Start there.
If you'd like a second opinion on where your organisation sits on that grid, our engineering team is happy to talk it through. It's the same team behind the projects linked above, recognised in the FT 1000 and Deloitte Technology Fast 50.
Alexandra Mendes is a Senior Growth Specialist at Imaginary Cloud with 3+ years of experience writing about software development, AI, and digital transformation. After completing a frontend development course, Alexandra picked up some hands-on coding skills and now works closely with technical teams. Passionate about how new technologies shape business and society, Alexandra enjoys turning complex topics into clear, helpful content for decision-makers.
CEO @ Imaginary Cloud and co-author of the Product Design Process book. I enjoy food, wine, and Krav Maga (not necessarily in this order).
People who read this post, also found these interesting:
-%2520Survey%2520Insights.webp)
