Generic advice about AI adoption is close to useless, because it usually starts in the wrong place. The deciding factor is the constraint the sector operates under, and those constraints differ enormously.
This is our reading across the twelve industries we work in. For each one, we look at what actually binds, where AI earns its cost, and where we would decline to build.
Healthcare systems live or die by the audit record
The constraint: the audit trail is the product. In healthcare, every clinical system is really two systems. One is the system clinicians use. The other is the record that proves afterwards who saw what and on whose authority. Most vendors build the first properly.
AI earns its cost in healthcare when it helps clinicians and reviewers do work they still own:
- Clinical documentation, turning a consultation into a structured note the clinician edits and signs.
- Triage support that ranks a queue by urgency with reasoning shown.
- Coding and billing.
- Retrospective review that flags cases for a second look.
Where we stop: autonomous diagnosis. The question is accountability rather than capability. More on healthcare.
Financial services systems are judged by their worst hour
The constraint: the system is graded on its worst hour, in retrospect, with the ledger in front of a regulator. The ledger is the financial record the firm has to stand behind later.
AI earns its cost where it helps people find and explain the right work faster:
- Fraud scoring inline on the transaction stream, with the queue ranked by expected loss rather than arrival.
- Anti-money-laundering triage, cutting a false-positive rate that consumes most analyst time.
- Onboarding document handling.
- Reconciliation exceptions.
Where we stop: anything that moves money without a human, and anything whose decision cannot be reconstructed. More on financial services.
Retail has ordinary months and a few hours that matter most
The constraint: eleven ordinary months and a handful of hours that pay for them, and those hours are the least tested.
AI earns its cost in eCommerce and retail where small improvements compound across stock, search and support:
- Demand forecasting and replenishment, where small accuracy gains move working capital.
- Search and merchandising, which is a retrieval problem before it is a model problem.
- Support triage on the large share of contacts that are order-status questions.
- Catalogue enrichment from inconsistent supplier feeds.
Where we stop: automated pricing without a floor and human review. More on retail.
Education systems have to work at the busiest moment
The constraint: everyone logs in at the same moment, on the oldest device in the building, over bandwidth the school cannot upgrade.
AI earns its cost in education when it helps teachers and students at the pace of the lesson:
- Formative feedback on drafts, where fast imperfect beats slow perfect that arrives after the topic moved on.
- Lesson preparation.
- Early identification of students falling behind, surfaced to a teacher.
- Accessibility support.
Where we stop: summative grading affecting a record without a human deciding. More on education.
Construction software has to survive the site
The constraint: the plan is always out of date and the software has to survive a gloved hand in the rain with no signal.
AI earns its cost in construction and real estate when it helps people work with changing documents, evidence and costs:
- Reading drawings and specifications to find clashes sooner.
- Progress assessment from site photography.
- Variation and claim drafting from a site note.
- Cost forecasting from historic project data almost nobody uses.
Where we stop: anything with a safety consequence, and anything that signs. More on construction.
Telecommunications turns tiny error rates into large customer problems
The constraint: volume changes what correctness means. One in ten thousand is thousands of customers a day.
AI earns its cost in telecommunications when it reduces noise and finds the few causes that matter:
- Fault correlation, turning ten thousand alerts into three causes.
- Fraud and abuse detection on traffic patterns.
- Support deflection at enormous volume.
- Capacity forecasting against real historic traffic.
Where we stop: automated changes to live network configuration. More on telecom.
Logistics work is mostly about exceptions
The constraint: the happy path is already automated. Exceptions are the job.
AI earns its cost in logistics and supply chain when it helps people see which exceptions matter first:
- Exception triage ranked by expected cost rather than arrival order.
- Customs and proof-of-delivery document handling.
- ETA prediction from your own performance rather than the carrier’s estimate.
- Demand and capacity forecasting.
Where we stop: autonomous commitments to customers. More on logistics.
Professional services firms are selling the unit AI compresses
The constraint: the billable hour is eroding and the firm is selling the unit being compressed.
AI earns its cost in professional services when it helps qualified people use the firm’s own knowledge faster:
- Retrieval over the firm’s own precedent, which is the highest-value and least glamorous application in the sector.
- First drafts a qualified person owns.
- Review at volume.
- Intake and scoping.
Where we stop: advice reaching a client without a named professional responsible. More on professional services.
Manufacturing software is a guest in a physical process
The constraint: the line does not stop, and software is a guest in a building governed by physical process.
AI earns its cost in manufacturing when it helps with signals, inspection and knowledge already inside the operation:
- Predictive maintenance on telemetry the machines already produce.
- Visual quality inspection.
- Yield optimisation across parameter combinations a manual search would never reach.
- Capturing retiring expertise.
Where we stop: closed-loop control without hard safety interlocks. More on manufacturing.
Travel inventory expires and availability has to agree everywhere
The constraint: inventory expires at midnight and availability must be true in several places at once.
AI earns its cost in travel and hospitality when it works inside the limits of expiring inventory and live availability:
- Demand forecasting and dynamic pricing inside visible limits.
- Guest messaging across languages and time zones.
- Itinerary assembly from components never designed to combine.
- Review analysis at a volume that makes patterns visible.
Where we stop: unbounded automated pricing. More on travel.
Media has a fast publishing clock and a strict rights clock
The constraint: two clocks. The publishing clock wants speed, the rights clock wants rigour, and they are not natural allies.
AI earns its cost in media and entertainment when it makes owned material easier to find, adapt and understand:
- Archive enrichment, making decades of material findable and therefore sellable.
- Format adaptation.
- Captioning, translation and accessibility at newly economic volume.
- Audience insight.
Where we stop: publishing without an editor. More on media.
Energy systems are bought for a long regulated life
The constraint: software that has to be right for twenty years, procured against a regulator.
AI earns its cost in energy and utilities where forecasts, inspection data and contact handling have direct operational value:
- Load and generation forecasting with direct settlement value.
- Asset condition assessment from inspection data.
- Outage prediction and restoration sequencing.
- Contact handling at the volume a storm produces.
Where we stop: any direct control action on physical infrastructure. More on energy.
The same AI system changes when the constraint changes
The sector constraint is central to the design. Four examples show the same nominal system built completely differently because of where it sits.
Healthcare puts the audit record first
A document extraction system in healthcare spends most of its engineering on the audit record and the data boundary. It must record what model was used, what version, what exactly was sent, who accepted the output, and how that record is retained for the life of the clinical record. The extraction itself is the easy part.
Logistics puts the exception path first
The same system in logistics spends its engineering on the exception path. Ninety percent of documents are routine; the value is entirely in how the awkward ten percent reach a person with the right context, ranked by what they cost.
Education puts the performance budget first
The same system in education spends its engineering on the performance budget. It has to work on a four-gigabyte Chromebook over a shared connection at the start of a lesson, which rules out several architectures that would be obvious anywhere else.
Energy puts documentation and handover first
The same system in energy spends its engineering on documentation and handover, because the thing will outlive everyone who built it and the buyer is procuring against a twenty-year horizon.
Four systems, one nominal problem, four different builds. A supplier who proposes the same architecture across all four has not understood any of them.
Three questions expose the real constraint
If your sector is not on the list, or you want to check your own reading of it, these usually surface it quickly.
- What does the failure cost, and who pays? A wrong answer in retail costs a return. In healthcare it costs a patient. In utilities it can cost a region’s power. The answer sets how much engineering the business case can carry and where the stopping line falls.
- What shape is the load? Smooth, spiky, seasonal, or a wall at a fixed moment. A school bell, a campaign, a storm and a trading day are four different problems and only one of them is solved by an average.
- Who has to be able to explain it afterwards, and to whom? A regulator, a court, a customer, a board, or nobody. This single question determines your audit requirements, and getting it wrong is the most expensive mistake available because it cannot be fixed retroactively.
Three questions, twenty minutes, and they will tell you more about what to build than any amount of technology evaluation.
The useful AI work across all twelve sectors looks familiar
Read the list and three things repeat.
- The best applications are boring. Document handling, triage ranking, forecasting, retrieval over material the organisation already owns. Almost nothing on the list demonstrates impressively, and almost everything on it pays for itself.
- The stopping line is always accountability, never capability. In every sector, the boundary falls where a human must carry responsibility for a specific decision. That line is set by regulation, liability and professional practice, and it does not move because a model improved.
- The constraint is never the technology. It is a bell, a settlement cycle, an audit, a physical process, a rights window. Understanding that constraint is the work. The model is a component you choose afterwards.
Some AI engineering problems are common to every sector
The differences dominate, but some things do not change. Those are the things worth standardising rather than rediscovering.
- Retrieval quality sets the ceiling. In every sector, the most common cause of a bad answer is that the system never saw the right material. Fixing chunking, where documents are split into usable pieces, adding hybrid search and filtering on metadata pays more than any amount of prompt work, regardless of industry.
- The refusal path is what makes it usable. A system that declines when unsure is trusted; one that always answers is verified constantly and therefore saves nobody any time. This is universal and it is the single most reliable predictor of whether staff keep using something after the first fortnight.
- The human queue is the dominant cost. Across every costing we have done, the escalation volume times a loaded hourly rate exceeds the inference bill, which is the cost of running the model. It is the line most often left out of business cases and the one that decides whether the project is worth doing.
- Somebody has to own the number. A named person, outside engineering, who is accountable for whether the thing is working. Projects without one drift into being declared successful because nobody is measuring, and that pattern is identical in a hospital, a bank and a warehouse.
Use the constraint before you choose the technology
Find your sector, read the constraint, and ask whether the project you are considering respects it or fights it. Projects that fight the constraint fail regardless of how good the technology is, and they fail late.
Then ask what your version of the boring applications is. In our experience the highest-return project in any organisation is almost never the one that was proposed enthusiastically, and almost always the one somebody described as too obvious to be interesting.
The twelve pages are at industries, each with the systems we build, the regulatory reality and the case studies where they exist.
