AI in industry: vertical playbooks
What changes when a language model meets a specific industry: the sign-off that has to happen, the format the data arrives in, the metric that decides whether it works, and the failure that costs money.
Almost nothing about a language model changes between industries. The same context window, the same token prices, the same tendency to answer confidently when it should abstain. What changes completely is everything around it: who is legally on the hook for the output, what the source documents look like, how wrong an answer is allowed to be, and what happens on the day it is wrong.
That is why a generic use-case list is worthless. “Summarise documents” is the same sentence in insurance and in agriculture, and the two projects share no engineering at all — one is a reason-code architecture wrapped around policy wording, the other is a store-and-forward pipeline for a device with no signal. These pages start from the constraint rather than the capability.
Nothing here is legal, medical, financial or professional advice. Obligations are described in general terms because they differ by jurisdiction and they move; where a page says a decision needs a qualified human, that is an engineering statement about where the accountability sits, not a summary of any particular rule.
AI in Customer Support: What Deflects and What Annoys
Why the usual deflection metric counts abandonment as success, what to measure instead, and how to design the handoff so the escalated customer does not start again from zero.
4 min read
AI in Legal: Contract Review and Its Limits
What contract review actually asks of a model — recall over a document set, not accuracy over a document — and what the lawyer who signs the memo needs in order to sign it.
5 min read
AI in Healthcare: Documentation, Triage, Diagnosis
The same technology sorted by how close its output sits to a clinical decision, which is what determines the oversight it attracts and the evaluation it needs.
5 min read
AI in Finance: Analysis, Reporting and Compliance
Why finance work has to tie to a control total, and the extract-then-reconcile architecture that follows from a model that cannot be trusted with arithmetic.
5 min read
AI in Recruitment: Screening and Its Legal Exposure
Why a screening model trained on past hiring decisions learns the past hiring decisions, what an adverse-impact analysis looks like as an engineering artefact, and the recruitment work that carries none of this risk.
5 min read
AI in E-commerce: Search, Descriptions, Recommendations
Why on-site search usually fails on missing catalogue attributes rather than on ranking, what a two-word misspelt query does to embeddings, and the experiment design that tells you whether any of it earned money.
5 min read
AI in Marketing: What Scales and What Backfires
The structural reason mass-generated pages hurt a site regardless of any search engine's policy, how to specify brand voice so it can be checked, and the claims that need a human before they go out.
4 min read
AI in Manufacturing and Supply Chain
The boundary between what classical models own on a factory floor and what only a language model can read, and why the network topology decides the architecture before anything else does.
5 min read
AI in Education Technology
A tutor's job is to withhold the answer, which is the opposite of what a model is trained to do — the architecture that fixes it, and the trial that would show whether it taught anyone anything.
4 min read
AI in Real Estate
Lease abstraction as the worked example: the fields, the clause that is prose encoding a formula, per-field evaluation, and a cost model you can run on your own portfolio.
5 min read
AI in Insurance: Claims and Underwriting
Having to give someone a reason for a decision rules out an architecture, and the one it leaves has the model extracting facts while the policy wording decides.
5 min read
AI in Government and Public Services
Two constraints that are genuinely specific to public bodies — a fixed-scope contract for a system whose behaviour changes underneath it, and logs that are records — plus why eligibility answers must only ever run one way.
5 min read
AI in Media and Publishing
A disclosure taxonomy that says something, the provenance metadata your own CDN strips, and why a fabricated quote is a different category of failure from a bad sentence.
4 min read
AI in Logistics and Field Operations
A voice latency budget worked out line by line for a driver with both hands busy, why the endpointing silence is the biggest term in it, and what to do when there is no signal at all.
5 min read
AI for Small Businesses: A Realistic Starting Point
Three workflows chosen for a business with no engineer, the shadow-run that verifies them without a labelled dataset, and the arithmetic that decides whether any of it is worth it.
5 min read
AI in Agriculture
Ground speed sets the vision budget, connectivity rules out the API, and one growing season per year sets how fast anything can improve — plus the label-reading task that is a genuinely good fit.
5 min read
AI in Energy and Utilities
Why forecasting stays with statistical models that imbalance settlement already prices, where the language work is, and what the control-room boundary does to the design.
4 min read
AI in Non-Profits and Research Organisations
What is buildable on a budget measured in tens of pounds a month, why beneficiary data is the most sensitive category in this cluster, and the line between assembling your evidence and inventing it.
5 min read
AI in Translation and Localisation
Why the LLM-versus-dedicated-MT answer is necessarily different per language pair and per content type, and the evaluation — hard checks, a neural metric, human error annotation — that answers it for your own strings.
6 min read
Building an Internal AI Assistant for Your Company
The permission model is the whole design: an index built without access control is a leak, and the leak is usually a document somebody shared broadly three years ago.
5 min read
Other topics
- LLM fundamentals & architecture
- Tokens, tokenization & context windows
- Prompt engineering
- Reasoning models & test-time compute
- Multimodal AI: vision, audio, video
- RAG & retrieval
- Embeddings & vector search
- AI agents & tool use
- Structured output & function calling
- Fine-tuning & post-training
- Inference, serving & latency
- Evaluation, benchmarks & LLM-as-judge
- Observability & LLMOps
- Hallucination & failure modes
- LLM cost engineering
- AI security & prompt injection
- Privacy, compliance & data residency
- AI governance, policy & society
- Building reliable AI applications
- AI hardware, GPUs & compute
- Open-weight models & local inference
- AI for developers & coding agents
- AGI, superintelligence, alignment & the long future
- Machine learning foundations
- NLP fundamentals & classical tasks
- Data engineering for AI
- Synthetic data & dataset curation
- AI product design & UX
- Search, ranking & recommendation
- Enterprise adoption & change management
- AI careers, skills & teams
- Reading AI research
- AI in science & discovery
- Robotics & embodied AI
- AI economics, markets & business models
- AI myths, hype & media literacy
- Context engineering
- Shipping AI features: patterns & anti-patterns