AI in practice
Not the demo version. Six things that reliably work in production, described by the problem they solve rather than the technique behind them — plus the cases where the honest answer is that AI is the wrong tool.
Answers are buried in documents nobody has time to read
Policies, manuals, contracts, past tickets — turned into something people can just ask. Answers come back with the source attached, so anyone can check them in seconds.
What that looks like
Customer support that resolves common questions without a human. An internal helpdesk that stops interrupting your senior staff.
Someone is retyping information from PDFs into a system
Invoices, forms, applications, delivery notes, email attachments. Fields get pulled out into your systems, with anything uncertain flagged for a person instead of guessed at.
What that looks like
Invoice processing that takes minutes instead of days. Onboarding that no longer waits on manual data entry.
Usually the least exciting option on this page and the one that pays back fastest.
Triage is slow, inconsistent, and depends who is on shift
Tickets, leads, applications and complaints classified and sent to the right place the moment they arrive, using the same criteria every time.
What that looks like
Urgent issues surfacing immediately rather than sitting in a queue. Leads reaching the right person while they are still warm.
Skilled people spend their day writing the same things repeatedly
First drafts of replies, reports, summaries and content, in your house style, for a person to approve. The model does the typing; your team keeps the judgement.
What that looks like
Support replies that need editing rather than writing. Long threads and documents summarised before a meeting instead of during it.
You find out about problems after they have cost you money
Demand, cashflow, stock and usage projected from your own history, with alerts when something departs from the pattern.
What that looks like
Knowing what to reorder before you run out. Catching an unusual spend pattern the week it starts.
This is conventional machine learning rather than an LLM, and it needs a decent history of clean data.
Your search only works if people already know the right word
Search that matches meaning rather than keywords, so a customer describing a problem in their own words still finds the right product, article or part.
What that looks like
Catalogue search that survives people not knowing your terminology. Internal search staff actually use.
Before you commit
Four questions we ask on every first call. You can answer them yourself before speaking to anyone — including us.
The best candidates already happen — often enough to be annoying, and consistently enough to describe. If nobody does it now, you are inventing demand as well as building software.
If yes, retrieval will probably work, because the knowledge exists in a form a system can reach. If it needs ten years of instinct, be sceptical.
If you can judge outputs, we can build an evaluation set and improve against it. If nobody can say what correct looks like, no amount of engineering will get there.
If the answer is "someone notices and fixes it", proceed. If it is "we lose a customer or breach a regulation", the design changes — a person stays in the loop.
A page claiming AI improves everything is a page nobody believes. These are the situations where we would tell you to spend the money elsewhere.
Models are probabilistic. If a wrong answer is unacceptable rather than inconvenient, the model can assist a person but must not decide alone.
If the logic fits on a page and rarely changes, write it as rules. It will be cheaper, faster, and you will be able to explain exactly why it did what it did.
Retrieval needs documents; forecasting needs history. If the information only exists in someone's head, that is the first problem to solve, and it is not an AI problem.
Automating something that happens four times a month rarely repays the build and upkeep. Volume is what turns a small saving into a real one.
Describe what is slow, expensive or breaking. We will tell you which of these fits, what it would take, and whether it is worth doing at all.
If you want the engineering detail behind any of it, that is inhow we work andinsights.
Describe your problem