A lot of what gets sold as custom AI is a prompt wrapped around someone else’s model. We pick the right technology for each part of the job: trained vision models for images and video, predictive models for your numbers, and large language models where reading and writing actually matter.
A few examples, far from the full list. We use whatever suits each part of your tool, often several together. Choosing them is one of the first technical decisions we make with you, because it decides how accurate the tool is and what it costs to run.
Large language models are one option among many. Brilliant at language, and an expensive way to do almost anything else.
Three example builds, split into the parts that do the work. Click any part to see which kind of AI runs it, and why we’d choose it.
Frames are sampled from fixed cameras and site phones every few seconds, faces are blurred, and each frame is tagged with the site, zone and time.
Video is expensive to move and store. We keep only the frames the models need, which keeps the hosting bill down and the privacy question simple.
Finds people, helmets, vests, plant and exclusion zones in every frame, and raises an alert when someone without a helmet walks into zone 4.
An object detection model trained on labelled frames from your own sites is more accurate on this narrow job than a general model, runs in a fraction of a second, and has no fee per image.
Reads weekly photos of each block and classifies the stage it has reached: rebar, formwork, slab poured, blockwork, MEP first fix.
Your programme already defines the stages. A classifier trained on them gives the same answer for the same photo every time, which is what a progress report needs.
Predicts which blocks will miss their programme dates, from progress rates, crew numbers, deliveries and weather.
This is a numbers problem, so it gets a numbers model. It shows which factors moved each forecast, so a project manager can argue with it.
Writes each project manager a daily report and answers questions like ‘why is Block C behind?’ in plain English.
Language is the one part of this build that needs a large language model. It reads the structured results from the parts before it, never raw footage, so each question costs very little.
Reads scanned claim forms, invoices and lab reports, in Arabic and English, and pulls out each field and line item.
OCR and layout models built for documents are faster and more consistent than a general model at pulling fields off forms, and they don’t charge by the page.
Maps each line item to the right procedure and diagnosis codes, with a confidence score.
Your adjudicated claims are years of labelled examples. A classifier trained on them is consistent, auditable and quick.
Scores each claim against normal patterns for that provider, procedure and member, and shows the reasons behind every flag.
Fraud keeps changing shape. A model that learns what normal looks like catches new patterns that fixed rules miss, and the reasons it shows can be explained to a regulator.
Approves clean claims inside policy automatically and routes everything else to the right person, with the evidence attached.
Your policy rules stay as rules. The models only feed them scores, so every decision can be traced and every rule changed by your team.
Reads the free-text doctor’s notes on claims a person reviews and drafts a short summary for them.
Only the claims that need a human get a language model, so token spend stays small while reviewers move faster.
Joins bookings from the PMS, channel data, competitor rates and the city events calendar into one clean history per property.
Every model after this one is only as good as this table. It’s usually where most of the build time goes.
Predicts demand for each room type and night out to a year ahead, and updates as bookings come in.
Demand forecasting is a classic machine learning problem with years of your own history to learn from. A language model has no business guessing it.
Chooses the rate that earns the most across the stay pattern, inside your floors, ceilings and parity rules.
Finding the best price under constraints is what optimisation models are built for. Your revenue rules are built in, never overridden.
Sends approved rates into your PMS and channel manager. The revenue manager can approve, change or lock any of them.
The tool lives in the systems your team already uses, so nobody has to log in somewhere new.
Writes the morning note: which rates moved, by how much, and why.
A short daily note is the right job for a language model. It works from the forecast and the optimiser’s output, so it costs very little to run.
Example builds to show the thinking. Every real build is scoped on your data, with your team.
Every call to a hosted model is charged, for as long as the tool runs. A build that sends everything to the biggest model looks fine in a demo and gets expensive in production. We design to reduce token burn from the first sketch.
Bars are illustrative. Your quote states an estimated monthly running cost, and after go-live we report the real figure for three months.
Dr John Hawkins decides which technology goes where, how the models are trained and how the build is tested. Our consultants, trained at global consulting firms, make sure the tool fits how your business actually runs. AI is the only work we do.
Intersect AI is Certified in the Claude Partner Network Services Track.We stay model agnostic all the same: OpenAI, Anthropic, Google, open-weight models, or one we train for you, part by part. “Intersect AI learned our complex procedures and found non-intrusive ways to capture the data, so we can finally apply AI to deliver better patient outcomes.”
Julian AndrieszCEO, Verita“Intersect AI built AI agents that stripped the manual grind out of our influencer marketing workflows. Work that used to take our team days now takes hours, freeing them to focus on strategy and creators instead of spreadsheets.”
Monique SümeghyFounder, Brand InfluenceScoping involves some back and forth with your team. It’s where the model choices on this page get made, and it’s what makes the quote accurate.
Only for the parts where a language model is the most accurate and cost-effective option. Often a model trained on your own data does the job better and costs less to run. We explain the choice for every part of the build.
Whichever suits each task: OpenAI, Anthropic, Google, open-weight models, or a model we train for you. You aren’t locked in to one vendor.
Your quote includes an estimate based on the volumes the tool will handle. After go-live we report the real figure for three months.
Yes. That’s where a lot of projects start. We design how the missing data gets captured, then clean and structure what you already have.
Yes. We agree up front who owns what, and we write code your team can read and maintain.
Wherever suits you. We aren’t tied to a cloud provider, and some models, like vision models on a site, can run on your own hardware.
A few sentences is enough. One of our AI leads reads every brief and comes back to you to start scoping it properly.
Rather talk it through? Book a call with the team
You’ll speak to the senior team from the first call. Last updated
