Use Cases › API Integration

Wire your applications straight into the intelligence.

A chat window is one way to use ZenithAI. The other is to call it from code — from inside the ERP, the trading platform, the inbox or the onboarding flow your teams already run — so reading, drafting and reasoning happen at the exact moment the work does. Below are four applications already built this way, the API routes behind them, and the pattern for building your own.

The integration pattern

Your application keeps the decision. ZenithAI does the reading.

The same shape repeats across every integration below: your application sends a specific, authorised task; ZenithAI reads, drafts or extracts on infrastructure you control; your application validates the result and carries the workflow to completion. AI never has the last word.

Your application An authenticated user, data they’re authorised to see, and one specific task — a broker note, an inbox message, a scanned form.
ZenithAI, on your infrastructure Reads, extracts, drafts or retrieves cited evidence, and returns a result — not a decision.
Your workflow Validates fields, matches master records, applies business rules, and routes exceptions to a person before anything is recorded.

Four applications, one API

Built by four different teams. Same pattern each time.

These are documented integrations, not hypotheticals — four products already calling ZenithAI in production to move paperwork, email and onboarding forward, with a person reviewing the result before it becomes a record.

01 / Compliance

e-Filix

Corporate compliance platform

A compliance team handles deadline reminders, consent notices, signed agreements and bank statements every day. ZenithAI is embedded at seven points in e-Filix — drafting communications and reading documents inside the workflow an administrator already uses, instead of a separate tool.

  • Alert templates — drafted from allowed variable names; invented placeholders are stripped automatically.
  • Consent notices — machine-labelled translation, reviewed and saved by an administrator.
  • Signed agreements & bank statements — scanned documents read into searchable text.
  • Document cover notes — an editable email or chat message drafted from the file and the user’s instruction.
Measure: minutes per approved notice, extraction corrections per document Control stays with: the administrator who approves each notice and translation
02 / Trade & finance

Commodity Trading Platform

Trade and finance operations

A broker note, a payment advice and a Bill of Lading describe the same transaction three different ways. Data Ingenious’s commodity trading platform uses ZenithAI to read free-form trade documents and return structured fields to the purchase or sale screen, where the trader checks the draft before saving.

  • Broker note → purchase order — parties, quantity, price, dates and terms extracted; matching against master records stays in the application.
  • Payment proof → reconciliation — amount, date and reference compared against the invoice, with the difference surfaced, not assumed cleared.
  • Bill of Lading — pattern matching first, AI to fill gaps; proposed values checked against the source text.
  • Vendor bills & inspection reports — varying layouts read into a consistent, confirmable set of fields.
Measure: reviewed order-entry time, payment mismatches found Control stays with: the trader who saves the record, and bank confirmation for cleared funds
03 / Email intelligence

IntelliMail

Inbox triage & response

IntelliMail turns an ambiguous inbox into a decision-ready view: category, urgency with a reason, extracted facts, a one-line summary and an editable reply. Clear-cut messages take a fast rules path; only unresolved messages go to ZenithAI — so the model is used where judgment is actually needed.

  • Eight categories — Meeting Request, Financial, Legal, Travel, Pitch Deck, Rewards, FYI, Other.
  • Urgency with a reason — not just a label; legal correspondence is urgent by policy, not by guess.
  • Reply drafting — a contextual draft with Formal, Professional, Friendly or Brief rewrites, always sent by a person.
  • Resilient by design — configured model failover, rate limits and a safe fallback keep the inbox moving.
Measure: triage time, urgency recall, draft-acceptance after editing Control stays with: the user who reviews and explicitly sends every reply
04 / Onboarding

Godil

Identity & address verification

Godil applies ZenithAI to identity and address document reading during customer onboarding. The model reads printed fields from an uploaded image or PDF — without being told the values it’s being checked against — and application rules compare the reading to the customer’s submission before routing to auto-verification, rejection or a reviewer’s queue.

  • Independent extraction — the model reads the document; it never sees the claimed values, so it can’t just confirm what it’s told.
  • Rule-based decisions — number format, name matching and address legibility are checked by application logic, not the model.
  • Reviewable outcomes — match dimensions, uncertainty and reason codes are recorded; unreadable documents go to manual review.
  • Controlled rollout — a feature switch that defaults off, so availability depends on the deployment’s own configuration.
Measure: false-acceptance and false-rejection rate on labelled samples Control stays with: this checks document-to-claim consistency — not authenticity, liveness or full regulatory KYC

Match the task to the route

Four routes cover most document and drafting work.

The same API that powers the ZenithAI chat interface is what these four applications call directly. Full reference and a live sandbox are on the API playground.

CapabilityRouteTypical useIntegration note
Drafting & structured image readingPOST /generateTemplates, extracted fields from a photo or screenshot, vision-assisted reading.Choose a compatible deployed model; validate returned fields against your schema.
Attachment-based conversationPOST /chatDocument summaries, questions over a PDF, audio-to-text.Confirm supported file types and size limits on your deployment.
Text-layer PDF extractionPOST /pdfRead text and tables from native (non-scanned) PDFs.Detect scans up front and route them to OCR instead.
Scanned PDF & image OCRPOST /ocr
GET /ocr/jobs/<id>
Background reading for scans and image-only documents.Poll for completion; handle failure, timeout and manual fallback.

Every request authenticates with an X-API-Key header, issued and scoped per workspace. Model availability and exact fields depend on your deployed version — this table is a capability map, not a full specification.

Building your own integration

Five decisions that hold up in production.

The four applications above didn’t get resilience by accident. The same five choices show up in each of them.

  • 01Define a small output contract. Ask for the specific fields the screen needs. Allow missing values. Reject a malformed response before it can touch a record.
  • 02Keep credentials and permissions in the backend. Store the X-API-Key server-side, apply your own user permissions, and send only the data a task is authorised to use.
  • 03Separate reading from the business decision. Calculations, master-record matching and approvals stay in your application logic. A cited answer supports review; it doesn’t replace it.
  • 04Design for slow and unavailable. Show progress on document jobs, bound your retries, and give every path a manual fallback. Test malformed files, empty results and timeouts before go-live.
  • 05Measure before you widen access. Run representative material with known-correct answers first, and track review effort alongside speed — on your own infrastructure, not a demo tenant.

Beyond these four

Workflows we see across other applications.

These follow the same integration pattern and published API surface, applied to a use case we haven’t yet built and measured together — each one needs workflow design and validation on your material before go-live.

Knowledge Adaptable

Answer policy questions from a service desk

Load approved leave, travel and expense policy; return an answer with the source page cited.

Sales Adaptable

Build a proposal from approved material

Specifications, scope and approved commercial inputs in; a proposal draft out, ready for review before release.

Finance Adaptable

Explain spend from a procurement register

Computed totals and comparisons with a narrative explanation — reconciled back to the source register.

Compliance Adaptable

Compare two versions of a policy or contract

A change brief referencing the relevant passages, for a specialist to interpret and approve.

Customer service Adaptable

Prepare a support response from a knowledge base

A cited answer and draft reply from your own product and support documents, integrated with ticket permissions.

Operations Adaptable

Draft minutes from a meeting recording

A transcript-based draft of decisions and action items, with names, owners and dates confirmed afterward.

See the complete set, filterable by function, on the use-case library.

Bring your own application

Tell us what it does. We’ll show you where AI fits.

Bring the workflow, the document types and a handful of real examples. We’ll walk through which routes apply, what stays in your application logic, and what a first pilot would measure.