How to add AI to an existing website or app without rebuilding it

Can you add AI to a website or app without rebuilding it?

Usually, yes. An AI feature is one more service your product calls: the page sends a request to your server, which adds the context it may share, calls the model, checks the reply and passes it back. Pages, checkout and accounts carry on as before.

Whether it works on your stack depends on three conditions, not on the framework:

• A place to run server code: an API route, a backend, a serverless function or a small separate service. The model key lives there, never in the page.
• Readable data: the content or records the feature should use can be exported or queried, with rules on who may see what.
• An owner: someone who decides what the feature may answer and reads what it got wrong.

A rebuild is justified only when one of these cannot be met: a builder that accepts no server code or outside service, say, or records so tangled nobody knows which is correct. Even then, a small service beside the old site often costs less than replacing it.

Which AI features can be bolted on to an existing product?

Four patterns cover most requests. They differ less in the model than in what they may read and change.

Start where mistakes are cheapest to catch: a wrong draft a manager reviews costs a minute, a wrong public answer costs trust.

What does an AI integration touch besides the model?

Calling a model takes a few lines of code. The work is everything around the call, and most of it maps to OWASP's 2025 Top 10 for LLM applications.

• Keys and routing. OpenAI's production guidance says to keep API keys out of your code and load them from environment variables or a secret manager, so every call passes through your server.
• Permissions. The model sees only what the signed-in person could open anyway; your server enforces that, not the prompt (sensitive information disclosure, LLM02).
• Untrusted input. Visitor messages and files can smuggle in instructions (prompt injection, LLM01), so the feature gets no tool or permission it does not need (excessive agency, LLM06).
• Output handling. A reply is data to check, not a command to run: prices and dates come from your records, and markup is escaped before it reaches the page (LLM05).
• Limits. Per-visitor rate limits and a monthly spend cap stop a loop or a scraper from running up the bill (unbounded consumption, LLM10); OpenAI lets you set spend alerts and a hard monthly limit.
• Logs and fallbacks. Requests, answers and failures are recorded, and when the model is slow or down the page falls back to its old behaviour.

How much does it cost to add AI to an existing website or app?

Market guides price whole programmes: discovery, data preparation, custom interfaces and testing. Compare scope lines, not totals.

Why the gap? The studio price covers one feature on the stack you already pay for, using your content as it stands. It leaves out a new interface, copywriting and data clean-up, the lines that push market quotes into five figures; if your data needs that work, the estimate says so.

What will the AI feature cost each month after launch?

The build is paid once; running costs return every month and grow with use. In Antino's breakdown, pre-trained model APIs take $1,000–$15,000 of a project, against $5,000–$30,000+ for data preparation and $5,000–$50,000+ for infrastructure: the model is rarely the biggest line.

An illustration only: an exchange that sends 3,000 tokens and gets 500 back costs about $0.011 on Claude Sonnet 5 (3,000 × $2 plus 500 × $10, per million), so 10,000 a month come to roughly $110. What protects the budget is a spend cap and a log, not the per-token rate.

What should you prepare before asking for a quote?

A reliable quote is built from evidence, not a wish list. Bring five things:

• One normal case and one that went wrong: a real question, form entry or record, with the answer a good employee would give.
• The source of truth: which pages, documents or tables the feature may read, and who keeps them current.
• Access: a staging copy or read-only credentials, the repository or CMS admin, and the hosting account.
• The boundary: what the feature must never do alone, such as confirm a refund, a price or a booking.
• The owner: who accepts the result and reads the failure log after launch.

With those, VITON¹³ replies by the next working day with a scoped estimate naming the pattern, the data it reads, the review step and what stays outside.

What must be in place before the AI feature goes live?

• A frozen evaluation set: real cases, awkward and failing ones included, rerun after every prompt or model change.
• An off switch: a setting that hides the feature and restores the old page without a new release.
• Disclosure: in the EU, Article 50(1) of the AI Act expects a system that converses with people to tell them they are dealing with AI unless that is already obvious. The Commission's FAQ gives 2 August 2026 as the date these transparency duties became enforceable.
• Personal data: know what leaves your server, where the provider processes it and under which data-processing terms; send the model only the fields a task needs.
• Monitoring: daily cost, error rate, refusals and handoffs on a dashboard someone actually opens.
• A human route: any answer the feature cannot support ends with a way to reach a person.

When should you not add AI to your product at all?

Sometimes the honest scope contains no model. Skip AI, or wait, when:

• A filter, a clearer menu or ordinary site search would answer the question: deterministic tools cost nothing per request and never invent.
• The answer must be exact every time, such as a contract clause, a dosage or a final price, and nobody is there to check it.
• Nobody owns the content: an assistant over outdated pages repeats them with confidence.
• Volume is tiny: a feature used a few times a week rarely repays its set-up and monitoring.
• Your data may not leave your servers, and no suitable model can run where it is allowed to go.

Then a content clean-up or a plain integration does more than a model, and we would rather say so in the estimate than sell a feature you switch off in a month.