AI Solutions
AI, built into everything we ship.
Automation, agents and retrieval, wired into how you operate and market. Seven of our own products run on the architecture sketched below, which means every claim on this page is something we already pay for and get woken up by.
OpenAI
Anthropic
Google
DeepSeek
Z.ai
Mistral
Meta
OpenRouterOne routing layer in front of every provider. Model choice lives in config, so the cheapest model that clears the bar gets the job.
What's included
Agentic AI
Agents that plan a sequence, call your tools, and act on what comes back. They run to completion without somebody watching. Lead qualification, ticket triage, a research question chased through forty tabs before anyone on your team opens one.
Automation
The parts of your week that are the same every week. Someone on your team is copy-pasting between two tabs for an hour a day right now, and that hour is recoverable.
Integrations & RAG
Wired into your CRM, the support desk, whatever internal docs you have. Retrieval runs first and the model writes second, which is what separates an answer about your business from an answer about businesses in general.
Workflow AI
Internal tools for the work that eats afternoons. Reporting, research, first drafts. A person still signs off before anything leaves the building.
Extraction & enrichment
Pages, PDFs and free-text fields turned into clean structured records. This is the least exciting thing on the list and the one that pays for itself fastest, because the outputs are checkable and the failure mode is a blank field.
Monitoring & anomaly detection
A daily pass over your own numbers. It flags the spend spike, the traffic drop, the campaign that changed while nobody was looking, and tells you in plain English what it thinks happened.
How we build it
Most AI projects die in the same spot. A demo meets real data and nobody budgeted for the gap. Here is the sequence we run on client work and on our own products, in that order.
01
Sort the tasks before writing any code
Before the model question comes a sorting question: how obvious would a wrong answer be? Draft subject lines are safe. A fabricated compliance requirement reads identically to a real one, complete with a plausible section number, and nobody catches it until somebody acts on it. Where a use case lands on that scale decides how much checking it needs, and checking is where the budget goes.
02
Build the smallest honest version
One workflow, wired end to end, on your real data, because curated sample data hides everything that will hurt you later. An afternoon gets you a demo. Getting to something you can put in front of a customer takes weeks, and the teams that budget for the afternoon are the ones that abandon the project in week six.
03
Ground it in your data, constrain the output
Retrieval over your own documents, tickets or product data. Every answer names the source it used, and that source gets checked for existence before anything ships. A schema on the response, so downstream code fails loudly. Low-confidence cases go to a person by default, a threshold you can move once the thing has earned some trust.
04
Log everything, then measure it
Every input and output stored, so when somebody complains six weeks later you can go and look. Evals written alongside the agent. Swap a model without them and you have no way of knowing whether the output got better or just moved.
05
Hand it over as something you can run
Model choice in config. Keys server-side, spend caps in place, and a written record of why each of those calls went the way it did. If we vanished tomorrow your team could keep it running, and change it.
The providers we've used, and what for
No single model wins every job. What follows is what is running behind our own products this week and why each one earned its slot. You can go and poke at the projects yourself.
Model providers
OpenRouter
The layer everything else sits behind
Every product we run calls models through OpenRouter. One key, one invoice, and the model name lives in a config file. When something cheaper or better ships, and something cheaper or better ships most months, we change a string.
Anthropic (Claude)
Careful reasoning over messy source material
Claude Haiku 4.5 runs the price check on every plan page on CloudHosting.io, holding one tier against comparable tiers from 30+ other providers and pulling out the renewal traps that live in the fine print. Bigger Claude models do the code-heavy work.
OpenAI (GPT)
Relationship extraction and embeddings
GPT works out who integrates with whom across the hundreds of companies on PlexWyn's map. Embeddings on the news crawler go straight to OpenAI, because routers do not serve embeddings and that one dependency has stayed direct since day one.
Google (Gemini)
Second-opinion review pass
Gemini reviews CloudHosting.io changes before they ship. Two models disagreeing catches the thing one model was confident and wrong about, which happens often enough that we keep the pass in.
DeepSeek
High-volume work at low cost
DeepSeek V4 Flash does the clustering on Site Structure. That means reasoning about every single page of a full site crawl, and at a few thousand pages per scan the price per token stops being a footnote in the spreadsheet and starts being the business case.
Z.ai (GLM)
Bulk summarisation and classification
GLM handles the repetitive summarise-and-tag passes on PlexWyn and the news crawler. You would not hand it your homepage copy. Hand it two thousand articles and ask for two thousand clean records and it works away all afternoon for the price of a coffee.
Perplexity
Verification against the live web
A checking pass. Perplexity searches the live web, so the URLs in its citations open when you click them, which is the whole reason it sits in the pipeline. On Defense.Direct and PlexWyn it is the third source that has to agree before a claim gets published.
Data, crawling & retrieval
Firecrawl
Hosted crawling and structured extraction
Runs the crawl behind Site Structure and the page-level extraction in SpendPane. Hosted crawling is worth paying for up to a point, and working out where that point sits for a given workload is a judgement call we have now made several times with real invoices in front of us.
Crawl4AI
Self-hosted crawling at volume
PlexWyn's crawl volume grew until hosted pricing stopped making sense, so we moved it onto self-hosted Crawl4AI in Docker on our own Oracle Cloud box. Same job. Fixed monthly cost, and no per-page meter running.
DataForSEO
Search, keyword and backlink data
Live SERP, keyword and backlink data for SEO.io. Every call runs under a monthly spend cap and a per-site cap, both set in config, so a runaway job stops itself before anyone reads about it on an invoice.
Model-agnostic, on purpose
Price and capability spread by more than an order of magnitude across models. Most jobs run fine on something well down that list. What matters is being able to change your mind in six months, when the list has reshuffled again.
Routing and cost discipline
- One layer. Every call goes through a single routing layer, so the model name lives in exactly one config file. Upgrades take ten minutes.
- Cheapest model that clears the bar. Classification and extraction run fine on small fast models. Reasoning across a forty-page document does not, and paying frontier prices for a tagging job is the most common way we see a budget disappear.
- Hard spend caps in config. Monthly and per-site limits enforced in code. A runaway job hits a wall on a Tuesday afternoon and somebody gets a message about it.
- Token spend watched from day one. A feature costing pennies in testing can cost real money at production volume, and that surprise usually arrives as an invoice.
- Self-hosted when the maths flips. Hosted providers until volume makes them uneconomic, then open-weight models on your own infrastructure. We made that move on PlexWyn, so we know what it costs.
Building the checking in
Stanford's RegLab measured hallucination rates between 69% and 88% on specific legal queries. The purpose-built legal tools, sold to law firms with retrieval and curated databases behind them, still landed between 17% and 33%.
- Answers grounded in your documents and required to cite the source, with the citation checked against something that exists.
- A schema on the output shape. Malformed responses fail loudly at the boundary.
- A second model, or a live-web search pass, disagreeing with the first before anything publishes. On Defense.Direct that job is Perplexity, and it has killed claims both earlier models were perfectly happy with.
- Every input and output logged. Teams skip this until their first incident, then rebuild it in a hurry.
We wrote more about this in where AI earns its keep, and where it quietly costs you.
Where we'll tell you not to use it
The fastest way to waste a budget on AI is to point it at the wrong job. Here are the jobs we push back on, including when you are keen on them.
Anything stated as fact to a customer, unread
Pricing, availability, specifications, policy, legal or medical claims. A chatbot that invents a refund policy has made a commitment somebody on your team has to honour or explain away.
Anything needing current information you did not supply
Models do not know your stock levels, this week's prices, or what your terms said before last month's edit. If it was not retrieved and passed in at request time, the model is guessing. Fluently.
Work that still needs checking line by line
If a person has to verify every line against a source anyway, the model has moved the work sideways. Sometimes that still pays. Often it does not, and we will say so before you sign anything.
A use case chosen because it demos well
The jobs that pay are usually classification, extraction and routing. They photograph badly and they never make it into an AI strategy deck. We would still start you there.
Curious where AI fits?
Tell us what eats the most time on your team. If AI is the wrong tool for it, we'll say so.