Laravel 13 shipped first-party AI primitives. The useful question for Philippine business apps is not whether your developer uses Copilot. It is whether AI belongs inside the product — drafts, triage, extraction — on top of data you already trust.
This is about the Laravel AI SDK (laravel/ai) in a real application. Coding assistants are a separate article.
What the AI SDK actually is
The official SDK is a Laravel-shaped API over providers (OpenAI, Anthropic, Gemini, and others): agents with instructions and tools, structured output you can validate, embeddings, and the usual generate/stream path.
That matters because most SME “AI features” fail at the glue, not the model:
- Prompts living in a random controller
- No structured response — staff paste JSON from ChatGPT
- No audit of what the model saw or suggested
- No way to fake the provider in tests
A first-party SDK does not make the business case. It makes the engineering less sloppy once the case exists. For the business case, start with practical AI for Philippine SMEs.
Where it belongs in a business app (not a demo)
We only consider in-app AI when a human already does a repetitive step on structured records:
- Inquiry drafts — first reply using real pricing rules and availability, staff send
- Triage — classify contact-form and Messenger text into queues your Laravel app already has
- Document extraction — invoice or ID fields into the same tables finance uses
- Ops summaries — weekly plain-language digest from live bookings or orders, not a CSV emailed on Friday
Notice what is missing: autonomous refunds, silent price changes, customer-facing answers with no review. Those are still how you lose money and trust.
What must exist before you composer-require it
The SDK will happily generate confident nonsense if your domain is a mess.
You need:
- One source of truth — bookings, stock, or invoices in the database, not three spreadsheets
- Permissions — the agent only sees what that role may see
- A human approval step for anything that commits the business
- Logging — prompt version, model, input IDs, output, who accepted or edited
- Budget caps — tokens are an operating cost, not a magic line item
If that list feels like “we need a system first,” that is the point. Same sequence we use in domain modeling before controllers.
AI SDK vs MCP vs “the developer’s ChatGPT”
Laravel now has three easy-to-confuse pieces:
| Piece | Who uses it | Job |
|---|---|---|
| AI SDK | Your application | Features your staff or customers touch |
| Laravel MCP | External agents (ChatGPT, Cursor, Claude) | Expose your tools and data on a protocol |
| Boost / IDE agents | Your developers | Write Laravel code faster |
This post is the first row. MCP as a product surface is Laravel MCP for business systems. Do not expose MCP to the public internet because a keynote made it sound inevitable.
A build sequence that does not embarrass you
- Pick one bottleneck you can time (first-response minutes, re-typed invoices).
- Ship the workflow without AI — states, roles, audit.
- Add the SDK as a draft or extract layer with structured output.
- Measure: time saved, edit rate, error rate. If staff ignore the draft, turn it off.
- Only then talk about more tools or an MCP server.
What we will not sell
- “Laravel 13 means your app is AI-native” — upgrading PHP is not a product strategy. See should you upgrade to Laravel 13.
- A chatbot on the marketing site as proof of transformation.
- Training on customer personal data without a retention and access story — pair this with cybersecurity basics that actually matter.
Straight next step
If a vendor pitched “AI inside Laravel” and you cannot tell whether they mean Copilot, a chat widget, or a feature on your real records, send us the pitch and a screenshot of the workflow. We will say what the AI SDK can honestly do in your app — and what to build before you pay a token bill.
