izz.ONE in action

Six situations where izz.ONE makes all the difference

Outages, attacks, runaway costs, new models, sensitive data and business teams with big ideas. That’s everyday life with AI in a large company. Here’s what happens in those moments when izz.ONE sits in the middle.

Pick a building block to see where it comes into play:

Customer Service
An AI provider goes down in the middle of peak season
- Status Check
- LLM Router
- LLM Gateway
Customer Portal
An assistant is ready to go live. But is it safe?
- Vulnerability Checker
- Custom GPTs
- Audit Log
Shared Services
The AI bill is growing faster than the value
- LLM Router
- LLM Gateway
- Budgets
IT Architecture
A better model launches, and switching is no big deal
- LLM Router
- LLM Gateway
- Evaluation
Human Resources
Confidential data, with the works council watching closely
- LLM Router
- PII Protection
- Audit Log
Legal
The business builds its own assistant, and IT stays in control
- Custom GPTs
- Roles & Permissions

The challenge
It’s the week before Christmas, and a retailer’s AI assistant is answering thousands of customer questions every hour. Then its model provider slows down and starts returning errors. Without a backup plan, chat and ticket handling grind to a halt.
How izz.ONE solves it
Status Check spots the rising response times within seconds. The Gateway moves every request to an approved model of equal quality, and IT Operations gets an alert in Teams. Once the provider is stable again, traffic shifts back step by step.
The result
Not a single customer noticed the outage. The team heard about it from an alert, not from complaints.

For developers: Failover rules

Failover targets are set per application and respect region and data classification.

failover:
  app: service-assistant
  primary: anthropic/claude-sonnet@eu
  fallback:
    - azure-openai/gpt-4.1@swedencentral
    - mistral/large@eu
  trigger:
    p95_latency_ms: 3000
    error_rate: 0.05
  notify: [teams:it-betrieb]

14:02:11
Anthropic (EU) response times climb to 4.2 sStatus Check
14:02:14
Provider flagged as disruptedStatus Check
14:02:14
00% of requests rerouted to Azure OpenAI (EU)LLM Gateway
14:02:15
Alert posted to the “IT Operations” channel in TeamsLLM Gateway
14:41:03
Provider stable, traffic returning in 10% stepsLLM Router

Switchover time: 3s
Customer-facing errors: 0
Rerouted requests: 41200

211 of 214 passed

1 critical and 2 medium findings, go-live blocked

Critical
System instructions exposed through role-playPrompt injection
Medium
Internal premium calculations quoted in a response Data leak
Medium
Policies bypassed through a translation requestJailbreak
Passed
211 more attacks blockedAll categories

The challenge
An insurer wants to launch an assistant in its customer portal. IT Security asks exactly the right question: can someone trick it into revealing internal information? Manual testing only covers a fraction of the known attack patterns.
How izz.ONE solves it
The Vulnerability Checker hits the assistant with hundreds of known attacks. These include prompt injection, jailbreaks, attempts to extract its hidden instructions, and data leaks. Every finding comes with an example and a recommended fix. Once the fixes are in, the test runs again, and after that it runs automatically every night.
The result
IT Security approves the go-live based on a clear, traceable test report. New attack patterns are tested on an ongoing basis.

For developers: Testing in the CI pipeline

The scan can be added as a step in your pipeline, and it blocks any deployment with critical findings.

# .github/workflows/deploy.yml
- name: izz.ONE Sicherheitsprüfung
  run: izz-one scan --target gpt:kundenportal --fail-on critical
  env:
    IZZ_ONE_KEY: ${{ secrets.IZZ_ONE_KEY }}

The challenge
In just one year, a large corporation has built a dozen AI applications, all running on the most powerful model available. The monthly bill keeps climbing, but nobody can say what each application costs or whether it even needs that much horsepower.
How izz.ONE solves it
The Gateway breaks costs down by application and task. The Router sends simple jobs like sorting tickets to smaller models, and complex analysis stays on the heavy hitters. Department budgets send a warning before anyone hits their limit.
The result
Much lower costs at the same quality, and every department can see what its applications use.

For developers: Budgets via API

Budgets and thresholds can also be set via API, for example when a new application is created.

await one.budgets.upsert({
  scope: "department:kundenservice",
  monthly_eur: 5000,
  alerts: [0.75, 0.9],
  on_exceed: "route_cheaper", // statt hart zu blockieren
});

Ticket classificationBefore: 4,800 €, With Router: 520 €
Customer service chatBefore: 3,100 €, With Router: 1,900 €
SummariesBefore: 2,100 €, With Router: 1,350 €
Contract analysisBefore: 1,600 €, With Router: 1,600 €

Customer service budget3,900 € of 5,000 €

Team lead alerted at 75%.

routing:  app: rechnungseingangremoved: -  model: azure-openai/gpt-4.1@euadded: +  model: mistral/medium-3@eu  data_class: intern

Evaluation, 500 invoices	Current	New
Fields correctly extracted	94.1%	95.3%
Response time (p95)	3.8 s	2.9 s
Cost per 1,000 invoices	€11.40	€7.10

Sample figures

The challenge
Every few weeks a new model arrives that is faster, smarter or cheaper. Until now, switching meant a new integration, new tests and a new deployment for every single application. So everyone stuck with what they had.
How izz.ONE solves it
The new model is added to izz.ONE and scored against the application’s real-world test cases. If it wins, a simple configuration change switches it on, first for a small share of requests and then for all of them. Not a line of the application’s code changes.
The result
New models go live in days, not quarters, and the numbers prove they’re better.

For developers: Gradual rollout

The rollout is controlled through routing weights. The application keeps calling model: "auto".

routing:
  app: rechnungseingang
  targets:
    - model: azure-openai/gpt-4.1-mini@eu
      weight: 80
    - model: mistral/medium-3@eu
      weight: 20   # Canary, Auswertung im Dashboard

The challenge
HR wants to use AI to draft employment references and prepare performance reviews. Data Protection and the works council insist on two things: personal data never leaves the company, and every use can be traced.
How izz.ONE solves it
The application is classified as “confidential”. izz.ONE masks names and IDs before anything is processed. The Router blocks all external providers and uses a self-hosted model instead. Every request is recorded in the audit log, without any readable personal data.
The result
Data Protection and the works council sign off, because the rules are enforced by the system and not just written down on paper.

For developers: Set the data class

The data class can be fixed per application or made stricter per request, but never looser.

const res = await one.chat.completions.create({
  model: "auto",
  messages,
  extra_body: {
    routing: { data_class: "vertraulich" },
    pii: { mode: "mask_and_restore" },
  },
});

Request from the HR app
Draft of an employment reference

PII Protection

3 personal details detected and masked.

Reference for [PERSON_1], employee no. [ID_1], department [ORG_1]

3
LLM Router
External providers blocked for this data class.
- ~~Azure OpenAI~~
- ~~Anthropic~~
- ~~Mistral~~
- Self-hosted
Response and restore
Placeholders in the response are swapped back for the real values
Audit Log
Time, application, model and region logged, with no readable personal data

Owner: Legal
Knowledge sources: Standard clauses, review guidelines, approval list
Access: Legal, Procurement, Sales (read-only)
Security scan: 214 of 214 passed

SharePoint1,240 calls/month
CRM860 calls/month
Procurement portal410 calls/month

1.4New clause on limitation of liability added
1.32026 review guidelines add

The challenge
Legal wants to pre-screen contracts faster. Waiting for IT to build an assistant takes months. But if the business builds its own, you end up with standalone tools that skip security checks and cost control.
How izz.ONE solves it
The lawyers set up a Custom GPT with their own review rules and standard clauses, with no coding at all. IT decides who gets to use it, and the Vulnerability Checker tests every new version. The same GPT is then available in SharePoint, the CRM and the procurement portal.
The result
One assistant, maintained by Legal, used in three systems and run under IT’s rules.

For developers: Pin a version

Production systems can lock in an approved version while Legal works on the next one.

const res = await one.chat.completions.create({
  model: "gpt:vertragspruefung@1.4",
  messages: [{ role: "user", content: vertragstext }],
});

Six situations where izz.ONE makes all the difference

An AI provider goes down in the middle of peak season

An assistant is ready to go live. But is it safe?

The AI bill is growing faster than the value

A better model launches, and switching is no big deal

Confidential data, with the works council watching closely

The business builds its own assistant, and IT stays in control

An AI provider goes down in the middle of peak season

Event log

Test report

An assistant is ready to go live. But is it safe?

The AI bill is growing faster than the value

Monthly cost by task

Configuration

A better model launches, and switching is no big deal

Confidential data, with the works council watching closely

The flow of a request

Request from the HR app

PII Protection

LLM Router

Response and restore

Audit Log

Contract review

The business builds its own assistant, and IT stays in control

Apply to izz.ai

Get in touch