Applied AI: retrieval, generation, extraction, routing
Applied AI means putting existing models to work on one specific business process. OK Next does not train foundation models. It builds systems that read your material, produce something the business can use, and can be checked afterward.
- MODELS
- Provider-agnostic, chosen per task
- COST
- Budgeted per run, stated before you commit
- DELIVERED BY
- Automato AI
01 — WHAT THIS IS
Applied AI is the practice of applying existing language models to a defined business process, with the retrieval, evaluation and fallback engineering that makes the output reliable enough to act on. It is not a chat window bolted to a website.
OK Next builds retrieval systems over a company’s own documents, so answers come from your contracts, policies and catalog. Generation pipelines that produce proposals, quotes, reports and briefs in your exact template. Classification and extraction that route inbound work by rules you can inspect and change.
The engineering matters more than the model. Which model, at what cost per run, with what context, with what fallback when the provider is down, with what evaluation set proving a prompt change improved things. Those are normal engineering decisions with normal answers, and the running cost is stated before you commit.
Where a deterministic rule will do the job, OK Next writes the rule. Not every problem should be handed to a model.
The pipeline runs in eleven steps, in this order.
- Your documents — the material your company already owns.
- Chunk — the documents are split into passages.
- Embed — each passage is turned into a vector the index can search.
- Index — those vectors are written to an index.
- Query — a question arrives. It comes from outside the ingestion chain, not from the index.
- Retrieve passages — the query and the index together return the passages that bear on the question.
- Assemble context — those passages are assembled into the model’s context.
- Generate — a primary model writes the answer, with a fallback lane behind it for when the provider is down. Cost and latency are budgeted per run.
- Validate against schema — the first gate. The output either passes or is rejected.
- Human review — the second gate. A person reads the output.
- Out of the company — only what has passed both gates goes out.
The reject exit is drawn because it is the honest half of the mechanism: a malformed answer is never shown. It returns to the generate step instead.
An evaluation set sits off the pipeline and feeds the generate step. It is run before a prompt or model change ships, so a regression is caught before a customer finds it.
Two gates stand between the model's answer and the outside: a schema it has to pass, and a person who reads it before it leaves the company.
02 — WHAT WE BUILD
ASK
What does our policy say about refunds after delivery?
- Source A.SOURCE · RETURNS POLICY
- Source B.SOURCE · SUPPLY CONTRACT
- Source C.SOURCE · PRODUCT CATALOG
Refunds are accepted after delivery where the item is unused.Cited from source A.
The supply contract sets a shorter window for wholesale orders.Cited from source B.
SCHEMA · VALIDATEDNOT CITED · C
Simulated interface. Every value in it is a placeholder. Two of three retrieved sources are cited by the answer, sentence by sentence; the third is retrieved and not used.
03 — HOW IT IS BUILT
01
Scope
The process as it runs today, written down, with the manual steps and handoffs named.
02
Build
Short cycles against a staging URL you can open at any time.
03
Instrument
Logs, alerting and an admin view, so a failure surfaces to a human.
04
Operate
Monthly operation, or a clean handover to your team. Both are quoted at scoping.
04 — HOW WE SCOPE THIS
05 — QUESTIONS
Questions
What is applied AI?
Applied AI is the practice of applying existing language models to one defined business process, with the retrieval, evaluation and fallback engineering that makes the output reliable enough to act on. OK Next builds applied AI systems; it does not train foundation models. In practice that means retrieval over your documents, document generation in your templates, and classification that routes work.
What is RAG, and do we need it?
Retrieval-augmented generation means the system looks up passages from your own documents and gives them to the model before it answers, so the answer comes from your own material. You need it when the correct answer lives in documents only your company has. You do not need it when a short, stable set of rules can be written directly.
How do you stop it from making things up?
Four mechanisms, used together: retrieval so the model answers from your documents; structured outputs validated against a schema so malformed answers are rejected; an evaluation set that catches a regression when a prompt or model changes; and a human review step wherever the output leaves the company. Where none of those is enough, OK Next writes a deterministic rule instead.
Will our data be used to train a model?
OK Next tells you which provider processes what, per project, and configures retention accordingly. Enterprise API tiers at the major providers do not train on submitted data by default, and that setting is confirmed in writing at scoping. Where the data cannot leave your systems at all, the system is built to run against a model you host.
Tell us what needs building.
Describe the process, who does it now, and the systems it touches. OK Next reads it and replies within one business day.