AI integration services for the product you already run
Adding an LLM to an existing product is mostly plumbing: reaching your data, permissioning it, and wiring the answer back into the system people already use. The one we shipped reconciled 14 data sources in 5 languages into an ERP that already existed — 87% less manual processing at 99.2% extraction accuracy.
$15,000–$30,000, 2–4 weeks, fixed price · a written scope inside 48 hours of the discovery call · or hire AI engineers onto your team from $4,000 a month
What does AI integration actually involve?
Six pieces of work, and the model is the smallest of them. On the platform behind every figure on this page — a multilingual logistics operation, client anonymised — the model call was a fraction of the build.
Reaching the data
Connectors to wherever the answer lives: databases, document stores, inboxes, internal exports. There it was 14 sources, from invoices and customs paperwork to email attachments.
Retrieval over your data
Ingestion, chunking, embeddings and ranking — RAG, built for how often the data moves. That corpus is reindexed daily, which is why it is retrieval and not a fine-tuned model.
Permission-aware answers
Enforcing at query time who may see which document. It is the requirement that rules out fine-tuning for most enterprise cases, because training data reaches everyone who reaches the model.
Writing back into your systems
The answer lands where people already work — the ERP, the CRM, the ticket — behind a review step wherever the model is unsure. That platform syncs validated values into the client’s ERP.
One interface to the model
OpenAI, Anthropic, Gemini or a self-hosted model behind a single provider layer, with streaming, fallback and token budgets, so changing models is days of work rather than a rewrite.
Evaluation from week one
A golden dataset built from real cases across languages and document types, so quality is a number you can regress against instead of an opinion.
What does an AI integration cost to run?
The build price is one number. The running cost is another, and it is the one quotes usually leave out.
That platform averages about $0.011 per answer across roughly 40,000 queries a month — about $0.002 of retrieval and $0.009 of generation. Answers take 1.9 seconds on average and 4.2 seconds at p95, measured to the last token rather than the first.
The split is the part worth carrying to your own case. Retrieval infrastructure is the cheap fifth; what retrieval really costs sits inside the generation line, as the extra input tokens every retrieved passage adds to every prompt. Context length is the cost lever; a cheaper vector database is not.
Estimate your integration with the cost calculator · Work out its running cost · The case study these figures come from · How RAG fits into an existing SaaS · RAG or fine-tuning?
How long does it take to integrate AI into an existing product?
Two to four weeks for an LLM feature. The variable is not the model; it is how much of your existing codebase has to be touched to reach the data.
Workflow and data audit
Which question the feature answers, where the answer lives, and a genuine random sample of your documents — not the tidy examples.
Golden dataset
Real cases with known correct answers, built first, so every later change is measured against the same bar.
Pipeline in slices
Retrieval and generation shipped one flow at a time, highest-volume first, rather than as one big-bang release.
Integration with checkpoints
The answer wired into your product or system of record, with a review step for the cases the model is unsure about.
Launch with monitoring
Accuracy dashboards, cost alerts and a feedback loop for prompt and retrieval improvements after go-live.
Run it, or hand it over
We support it after launch, or your team takes it with the evaluation set, so a regression is a failed build rather than a support ticket.
When is AI integration the wrong move?
Three situations where we would rather say so on the first call than in week three.
The data does not contain the answer
No pipeline fixes missing information. Sampling the real corpus before scoping anything is a day of work and it settles the question.
Nobody can say what a correct answer is
Without that definition there is nothing to evaluate against and quality becomes an opinion. Settling it is a week of work, and it comes first.
It is a new product, not a feature
When there is nothing to integrate into yet, the model and the business are the same build — a full AI product, priced at $40,000–$120,000 over 8–12 weeks — see AI development services.
Common questions
What comes up on nearly every first call about an integration.
What this has shipped
The delivery behind every figure on this page.
Have a product that needs an LLM in it?
Tell us what it should answer and where the data lives. We come back with a fixed scope and price in writing within 48 hours of the call.
Message received
We’ll review your enquiry and respond within one business day.