ORYX AI / ENTERPRISE AI AGENTS
AI agents inside your business processes
Not “let’s roll out a chatbot”, but an engineering layer on top of what the company already runs: CRM, ERP, document management and the contact center.
WHY NOT JUST A MODEL SUBSCRIPTION
The model is a commodity. The value is in the integration layer
Employees have already tried public AI tools. Our role is not to give access to a model, but to wire it into the company’s permissions, data and policies — so every result can be verified, traced and rolled back.
- 01 Public chatbot No company context
- 02 Employees writing their own prompts Not integrated with company systems
- 03 Agent + connectors + guardrails The Oryx AI approach
What we do
- 01
API connectors
To the customer’s CRM, ERP, document management, billing, 1C and internal services.
- 02
Access control
The agent acts only within the permissions of the employee whose work it performs.
- 03
Action log
Who did what and when — kept separately from application logs.
- 04
Infrastructure
From cloud APIs to a fully on-premise setup for sensitive data.
CONTROLLED AUTONOMY
Three levels of agent autonomy
Nobody hands real operations to an agent on day one — and nobody should. We start with suggestions a person reviews and expand the agent’s authority once it has proven reliable on the customer’s data.
-
Level 1
Advisor
Suggests — a person decides
The agent drafts a reply, a document or a CRM entry. Nothing is sent or saved without an employee’s approval. Minimal risk, fastest launch.
-
Level 2
With guardrails
Acts within set boundaries
The agent performs operations on its own, but only within the rules — up to a set amount, for specific document types. Everything is logged, and an immediate stop mechanism is available.
-
Level 3
Autonomous
Triggers itself on events
The agent reacts to system events without a manual request. People step in only for exceptions. Requires solid monitoring and a process that has already been tested.
USE CASES
From the back office to the front line
Each scenario is a standalone delivery project with its own integrations, its own level of agent autonomy and its own measurable effect.
01 Insurance / health insurance Document anonymisation before external review
- 01Document scan
- 02Text recognition (OCR)
- 03On-premise model finds personal data
- 04Redaction and log of what was hidden
As is
A specialist manually strikes out names, policy numbers and diagnoses before a document goes to a third party. Pattern matching fails here: OCR errors, varying forms, inflected names.
To be
The agent finds personal data by meaning, not by rigid patterns, and returns an anonymised document with a log of what was hidden. Data never leaves the perimeter: the model runs on the customer’s servers.
Under the hood
- On-premise model
- OCR
- Human spot checks
02 Distribution / retail Supplier documents straight into 1C Level 2
- 01Incoming PDF, scan or Excel
- 02Reading line items without templates
- 03Matching against the 1C catalogue
- 04Goods receipt note created
As is
Accountants key in line items from every supplier’s documents by hand. Legacy recognition needs a template per supplier and breaks when a supplier changes its form.
To be
The agent reads any format by meaning, matches line items against the 1C catalogue and creates the goods receipt note. A person reviews only what the agent could not match with confidence.
Under the hood
- Vision-capable model
- 1C integration
03 B2B sales CRM assistant: the call ends, the deal is already updated Level 1
- 01Call or meeting transcribed
- 02Summary: agreements, next step, stage
- 03CRM updated via API
- 04Sales rep confirms or edits
As is
After every call the sales rep fills in the deal card by hand. With a steady call flow this becomes a bottleneck, and some agreements are never recorded.
To be
The agent transcribes the call or joins the online meeting, extracts the agreements and updates the deal fields. The rep sells instead of filling in forms.
Under the hood
- Speech-to-text
- CRM API
04 Telecom / internet providers Request triage: routine requests closed, complex ones routed Level 2
- 01Chat, email or ticket
- 02Topic and priority detection
- 03Answer from knowledge base and billing
- 04Complex cases to a specialist with a summary
As is
First-line operators sort incoming requests by hand. Routine questions — balance, plan, ticket status — take as much attention as the genuinely complex cases.
To be
With read-only access to the knowledge base and billing, the agent answers routine questions and hands complex ones to a specialist with a short summary, so no one starts from zero.
Under the hood
- Knowledge-base answers (RAG)
- Billing API, read-only
Effect
Lower first-line load at peak hours
05 Manufacturing Searching regulations inside a closed network Level 1
- 01Engineer’s question in plain language
- 02Semantic search across the archive
- 03Answer with the source and its version
As is
An engineer searches for the right standard, regulation or manual through folders and keyword search — and sometimes works from an outdated version without noticing.
To be
The agent searches by the meaning of the question and answers with a link to the source document and its version. The documents are restricted, so everything runs inside the company network.
Under the hood
- On-premise model
- Semantic search (RAG)
Effect
Fewer errors caused by outdated regulation versions
06 Retail / logistics Onboarding assistant instead of a queue to the mentor Level 1
- 01New employee’s question in chat
- 02Company policies and HR records
- 03Answer; unusual cases go to HR
As is
Frontline staff with high turnover ask mentors and HR the same questions: how to log a shift, where a policy is, how to request leave.
To be
The agent answers from internal policies around the clock and pulls personal details, such as remaining leave or the schedule, from the HR system. Anything non-standard goes to a person.
Under the hood
- Policy answers (RAG)
- HR system API
Effect
HR relieved of routine questions; faster ramp-up for new staff
07 Banking / contact center Quality control across the whole call stream Level 1
- 01All calls, not a sample
- 02Transcription tuned to banking terms
- 03Script deviations and customer sentiment
- 04Digest of recurring issues for the manager
As is
Quality control listens to a small random sample of calls. Systemic problems — a weak script, a confusing product term — stay invisible until complaints pile up.
To be
The agent transcribes and analyses the entire call stream and surfaces recurring script deviations and causes of dissatisfaction. The manager sees the pattern, not an isolated case.
Under the hood
- Speech-to-text
- Model-based text analysis
Effect
Every call reviewed, not a sample
08 Management Company health on one screen Level 1
- 011C, CRM, HR and billing via read-only APIs
- 02Metrics gathered and cross-checked
- 03Live dashboard
- 04Questions in plain language
As is
Sales, costs and profit arrive as a month-end summary prepared by hand from several systems.
To be
The agent collects metrics from company systems, flags discrepancies between them and answers questions such as “why did the margin dip in January?” with figures from company data.
Under the hood
- Read-only APIs
- Questions in plain language
- Real-time updates
THE INFRASTRUCTURE ARGUMENT
Not everything belongs in the cloud
Medical documents, restricted regulations and bank call recordings share one constraint: the data cannot leave the customer’s infrastructure. In these cases the model is deployed inside the customer’s network, and company knowledge is connected through document search — without lengthy model retraining.
- 01Data stays inside the perimeterSuitable for personal data and banking secrecy requirements.
- 02Runs in an isolated networkIncluding high-security facilities without internet access.
- 03Predictable costsNo per-request pricing for an external model.
- 04Company knowledge does not leakDocuments never end up in public models.
SECURITY REVIEW
We name the risks first
If we do not raise these four points ourselves, the customer’s security team will — so the answers are ready in advance.
- 01
The agent can make things up
Answers are based strictly on company documents with a source reference. Wherever a decision touches money or obligations, the final word stays with a person.
- 02
Action log
Every agent action — who, what, when — is recorded separately from application logs. It is the first thing the security team asks for at acceptance.
- 03
Access rights
An agent never holds more permissions than the employee whose work it performs. Boundaries are enforced at the integration level, not by prompt wording.
- 04
Stop and rollback
For levels 2 and 3, the agent can be disabled immediately and its latest automatic actions rolled back without stopping the rest of the system.
HOW WE START
Three steps to the first working process
- 01
Discovery
We review the customer’s systems — CRM, ERP, document management: which APIs exist and what shape the data is in.
- 02
Pilot on a single process
One use case at level 1: a person reviews every result.
- 03
Scale-up
Roll out to adjacent processes and expand the agent’s authority step by step.
NEXT STEP
Tell us which process to start with
Describe the process, the systems involved and the data restrictions. We will suggest a use case, an autonomy level and a pilot format.
Thank you. Your request has been sent.
We will review it and reply to the email address you provided.