AI systems · Laravel & WordPress
I build AI into the software your business already runs on: Laravel applications and WordPress sites that read the email, extract the invoice, draft the product page and answer from your own documents — then hand the one decision that matters to a person. Built on the official Laravel AI SDK and the WordPress AI Client, with your data, your keys and your choice of model.
{ intent: quote, qty: 40, deadline: Fri, urgency: high }
01Where it saves timeFive workflows, before and after
Most businesses already use AI somewhere — 88% in McKinsey’s 2025 survey — yet more than 80% say it has not moved their profit. The difference is almost always where it sits. A chat window beside the work saves minutes. The same model inside the workflow — reading what arrives, filling in your systems, asking a person only when it should — takes whole steps away.
A customer emails
Today, by hand
With the system
Left for you
A PDF arrives
Today, by hand
With the system
Left for you
A new product comes in
Today, by hand
With the system
Left for you
Someone fills in your form
Today, by hand
With the system
Left for you
A colleague has a question
Today, by hand
With the system
Left for you
02ScopeWhat I build
A search box or a chat that answers from your documents, policies, product data and past tickets — not from the internet. Your content is split, embedded and stored so it can be searched by meaning, and every answer links to the passage it came from. When the answer is not in your material, it says so instead of inventing one.
Invoices, purchase orders, CVs, contracts, forms, inbound email: read by a model that is required to return a fixed structure, validated like any other input, and written into your database or your accounts system. The ones it is unsure about go to a queue for a person, with the reason.
Product pages, descriptions, FAQs, alt text and translations generated as drafts inside WordPress and WooCommerce, in your voice, for an editor to check. I have built this in production: product pages written from a single photograph and carried into seven languages in one action.
An agent is a model that can use tools: look up an order, update the CRM, book a slot, issue a refund. Each tool is ordinary, tested code with its own permission check, and anything with consequences waits for an approval before it runs. The Laravel AI SDK builds this approval step in; I use it.
WordPress now has an Abilities API and an official MCP adapter, which let AI assistants discover and run the actions your site allows — publish a draft, look up an order — and nothing else. I register the abilities that make sense for your team, each with its own permission check.
Most of this work is not a new product. It is a Laravel app or a WordPress site that works, gaining one feature that removes the most tedious step in someone's day. I read your code first and add to it in the same style, rather than bolting on a separate service.
03MethodThe engineering around the model
Calling a model takes ten lines. Trusting what it does inside your business takes everything around those ten lines — and that is most of what I build.
Answers are built from passages retrieved from your own content, and cite them. That is the single most effective defence against a confident wrong answer, and it lets anyone check the work in one click.
Where the result feeds another system, the model must return a defined structure — fields, types, allowed values — and it is validated before anything is saved. A malformed answer is a failed job that gets retried or flagged, never a corrupt record.
Sending, paying, deleting and publishing wait for a human. Drafting, sorting, extracting and looking up do not. Where that line sits is a decision we make together, per workflow, and it can move as trust grows.
Before launch there is a test set drawn from your real cases, and the system is scored against it. Changing a prompt or a model is then a measured change, not a hopeful one.
Code talks to a provider-agnostic layer, so moving between OpenAI, Anthropic and Gemini is configuration, not a rewrite. Every call is logged with its tokens, cost and time, so you see what each workflow costs to run.
If a spreadsheet formula or a plain rule would do the job, I will say so. AI is the right tool for reading unstructured things — text, documents, images — not for arithmetic.
04ProcessSmall first, then proven
05QuestionsAsked often, answered plainly
No — it removes the retyping, sorting and searching from their day, and leaves them the decisions. Every system I build has a person approving anything that matters. In practice the same team handles more work with fewer mistakes, and stops doing the part of the job nobody wanted.
Not on the business API tiers I use: OpenAI, Anthropic and Google do not train on API data by default. Your documents stay in your own database; only the passages needed for a given answer are sent with the request. Where data must not leave a region or a server, that constrains which model we choose, and I will say so up front.
It is designed to be caught. Answers cite the source they came from; results that feed a system must pass validation; anything with consequences waits for approval; and a test set of real cases is re-run whenever a prompt or model changes. The goal is not a model that is never wrong — it is a system where a wrong answer cannot quietly do damage.
Usually far less than people expect: most business workflows cost cents per item in model usage. Every call is logged with its cost, so you see the real figure per workflow from the first week of the pilot rather than an estimate.
Whichever does your task best at the right cost — typically a model from Anthropic, OpenAI or Google. The code talks to a provider-agnostic layer (the Laravel AI SDK, or the WordPress AI Client), so changing model later is a configuration change, not a rebuild.
Yes, and that is most of this work. I read the existing code first and add the feature in the same style — a new job, a new admin screen, a new block — rather than a separate service you then have to keep in sync.
A prototype on your real data in about a week, and a first workflow in production in roughly three to six weeks, depending on the systems it has to connect to. Starting with one workflow is deliberate: it proves the value before you invest in the next.
Next step
A paragraph is enough to start. If AI is the wrong answer for it, I will say so — sometimes the honest fix is a plain rule or a small Laravel application.