AI deployment by industry: where the difficulty sits
Sector pages about artificial intelligence usually describe opportunity. These describe difficulty, because the opportunity is not what a project runs into in week two. What it runs into is one specific constraint that differs by industry and that determines everything about how the work should be sequenced: in financial services it is that the decision was already regulated, in healthcare that anything clinical sits behind a different order of approval, in retail that product data was never clean, in logistics that the deadline is physical, in professional services that the knowledge lives in the most expensive people in the firm, and in manufacturing that operational technology is separated from information technology by design. Each page also names two things nobody puts in sector marketing: the loud use case that is usually the wrong first project, and where the ground truth actually lives.
Financial services
- The constraint:
- Every decision that touches a customer outcome already has a regulator with expectations about explanation and record-keeping, and those expectations predate the technology.
- The distraction:
- Fraud detection, which is already well served by existing models and is rarely where the unmet need is.
- A first project that works:
- Document extraction in onboarding or claims: high volume, checkable output, and a person reviewing the result anyway.
Healthcare
- The constraint:
- Anything touching a clinical decision carries a regulatory burden of a different order, so the work that lands is almost always administrative rather than clinical.
- The distraction:
- Diagnostic support, which attracts attention and sits behind the heaviest approval path available.
- A first project that works:
- Correspondence triage, coding support or discharge summarisation: administrative, reviewed, and where staff time is genuinely scarce.
Retail
- The constraint:
- Product data is fragmented across suppliers, systems and years of inconsistent entry, and the quality problem is upstream of anything a model can fix.
- The distraction:
- Personalisation, which is the loudest use case and depends on the data quality nobody has fixed yet.
- A first project that works:
- Product data normalisation and enrichment: unglamorous, measurable, and it unblocks everything else on the list.
Logistics
- The constraint:
- Operational decisions have hard physical deadlines, so a system that is accurate and slow is useless in a way it would not be elsewhere.
- The distraction:
- Route optimisation, which is a well-studied operations research problem and usually not where the manual effort actually goes.
- A first project that works:
- Exception handling on shipping documents: high volume, structured output, and currently done by people reading scans.
Professional services
- The constraint:
- The knowledge being automated lives in senior people who are billable, which makes the input a system needs the most expensive resource in the firm.
- The distraction:
- Drafting, which works well and produces value that is hard to capture when the business model bills for time.
- A first project that works:
- Retrieval across the firm’s own prior work, which is a search problem the firm has failed at for twenty years.
Manufacturing
- The constraint:
- Operational technology is separated from information technology by design and often by regulation, so getting data out is an architectural question rather than a permission one.
- The distraction:
- Predictive maintenance, which requires failure data most plants have too little of to train on.
- A first project that works:
- Technical documentation retrieval for maintenance and quality teams, which needs no plant-floor integration at all.
The pattern across all six
Read the table and the same shape appears six times. The use case that gets discussed at conferences is high variety, high consequence and politically weighted. The project that actually works first is high volume, low variety, and recoverable when wrong.
This is not a counsel of timidity. It is about what a first project is for. Its deliverable is not the time it saves; it is evidence that this can be made to work inside this organisation, which is what makes the second and third projects authorisable. A first attempt at the hardest problem either fails or overruns, and both outcomes remove the possibility of a second attempt for about a year.
Ground truth is the thing to find first
Every industry has a record of decisions a human already made correctly, and almost none of them think of it as a dataset. Adjudicated claims. Coded clinical records. Catalogue entries a merchandiser fixed. Resolved shipping exceptions. Prior matters and their outcomes. Maintenance logs.
That record is what an evaluation set is built from, and without one there is no way to tell whether a system is good, which means no way to tell whether a change helped. Finding it is frequently the single highest-value hour of a first engagement, and it is invisible from outside the sector, which is the strongest practical argument for having somebody on the project who has worked in it.
The characteristic obstacle is not that the record does not exist. It is that it lives with a team who were not part of the purchase decision and have no obligation to help, which is the same obstacle as stage two everywhere and is why that stage consumes the calendar.
What these pages deliberately do not contain
No market sizing, no adoption percentage by sector, and no claim about how much any industry will save. Figures of that kind are available and they are averages over organisations whose variation on the same process is larger than the effect being reported.
No named clients and no case studies. We do not publish work we cannot name, and inventing an anonymised composite that reads as a real engagement is the kind of thing this site exists to avoid.
What is here instead is the constraint, the distraction and the first project, which are the three things worth knowing before a first conversation in any of these sectors.
