Ask It. Build It. Embed It.

This guide comes out of ongoing whiteboard working sessions with a trusted circle of law students and attorneys, where the working rule is shared ideas rather than instruction. It is offered here as a starting source, not a final word. Product names, plan limits, and prices change quickly; the technical distinctions are the stable part.

Three access modes, one stack of capabilities.

A hosted AI assistant is a vendor-managed application that accepts natural-language and multimodal input, assembles a working context, invokes one or more models and tools, and returns text, images, files, code, or analysis through a conversational interface.

What happens technically: the application receives your prompt, files, and history; selects or routes to a model, possibly with tools such as search or code execution; the model performs inference — computing a probable and useful output from the supplied context; the result returns in human-facing form; you review, correct, refine, and verify.

Accuracy is not one universal threshold; the acceptable error rate depends on what the output is for. Brainstorming tolerates roughness. Legal authority, deadlines, client facts, and filed representations require direct verification.

A coding agent is an AI-enabled software agent that can inspect a project workspace, edit multiple files, execute commands, run tests, read logs, and iterate toward a specified result. Its practical capability comes from the combination of a model, tools, workspace access, execution, memory or state, and a control loop.

A complex result should not be treated as one giant generation. It decomposes into stages with intermediate artifacts and checks: preserve and inventory the sources; extract text and metadata repeatably; apply the defined checklist or rule set; generate structured findings with source references; run tests and quality checks; render the outputs; inspect the final artifacts and record remaining limitations.

The agent does not merely produce a confident final memo. It leaves a chain of inspectable work products that can be reviewed, rerun, corrected, and compared.

An application programming interface is a defined contract that allows one software system to request data or operations from another. In a model API, client software sends an authenticated request containing a model selection, instructions, inputs, tools, and parameters; the service performs inference and returns a structured response the client can display, store, validate, or use in another operation.

Metered usage, simply: API cost = input usage + output usage + tool and service charges + storage or runtime charges. Cost control means choosing an appropriate model, reducing unnecessary context, caching repeated material, using structured outputs, batching work, setting spend limits, and routing simple tasks to cheaper or local components.

Local and hybrid systems: a hosted API runs on a vendor's infrastructure under the vendor's terms. Local inference runs model weights on hardware you control — no vendor token charge, but hardware, electricity, maintenance, licensing, and capability tradeoffs. A locally running model can expose a local API to other local software. Sensitive or routine work can remain local while higher-capability tasks call hosted APIs; the best design depends on data risk, required quality, latency, cost, and maintenance capacity.

A legal example — authority and rulemaking monitor: official-source APIs provide new opinions, Federal Register documents, dockets, and bills; the application stores the original source and retrieval metadata; a model classifies relevance and proposes links to existing matters; the system requires source verification and human approval before anything becomes legal analysis; approved findings update a matter dashboard or draft memorandum. FederalRegister.gov, Regulations.gov, Congress.gov, and CourtListener all expose public endpoints — coverage and terms must be checked per project.

Intelligence, connection, triggers, and orchestration are separate layers. AI supplies probabilistic analysis or generation. Automation runs defined steps because a schedule, event, rule, or user action triggered it. An API is a connection — it does not decide when to run. An agentic workflow combines a model with instructions, tools, state, iterative planning, guardrails, and stopping conditions. A human in the loop is a designed review point before consequential output is used.

The more consequential the work, the more explicit the controls.

A professional reliability stack: control the source set; separate retrieval from analysis — the model is not the source of the law or the record; require provenance for each material finding; validate structure with schemas and cross-checks rather than accepting free-form prose; current-check authority for negative treatment, amendments, and jurisdiction; review by consequence — a qualified reviewer approves legal conclusions, filings, and deadlines; preserve an audit trail of inputs, versions, outputs, corrections, and approvals.

Confidentiality and access control: minimum necessary data and narrowest practical permissions; read-only access to originals, writes to separate output locations; understand the vendor's retention, training, sharing, deletion, and breach policies before disclosing protected information; protect API keys like passwords; separate matters so an agent cannot browse unrelated information; explicit approval gates for consequential actions; log enough to investigate errors and demonstrate supervision.

Technical language paired with the simplest accurate explanation.

• The source exists and was opened.
• The source is authoritative for the proposition.
• The quoted or summarized language is accurate; the pin cite is correct.
• The authority is current with the correct precedential weight.
• The factual statement is supported by the actual record.
• No prohibited source was used; no protected information disclosed.
• The output identifies uncertainty instead of concealing it.
• A qualified human reviewed the result before consequential use.
• The inputs, output, corrections, approval, and final artifact were preserved.