our AI method
Under the hood.
Here is how Kaiyan works: the method, how agents are orchestrated, the engagement levels, where the models run, how costs are governed.
Running an agent is not enough. What counts is anchoring it in your processes and your data, then keeping it alive in production. That is what this page is about.
4 pipelines
Dev · QA · Infra · Docs, orchestrated in parallel
3 levels
of confidentiality: public, managed private, sovereign
Native FinOps
every agent's cost tracked from token to euro
What we believe
AI does not remove the complexity
of your information system.
The demo looks impressive: it all seems to be front end and back end. Then you walk into a real company, and the architects show up. In truth, AI in production deals with three layers. The one the demo shows. The one underneath that never goes away, security, storage, sovereignty. And the one that ties them together: integration with the real system. That is where most AI projects fail, and that is where our work happens.
The visible layer
Front end, back end, the agent that produces a screen. This is what the demo shows.
The persistent layer
Security, storage, cloud, compute, sovereignty, monitoring. It sits under every project and never goes away.
The integration layer
Hooking into real systems, migrating the data, going live. This is where our work happens.
Which is why the model matters less than people think.
The reasoning model is powerful, but it is a commodity: you pick one, you rent it, you swap it out. Mistral, Anthropic or sovereign, it does not change what we offer. What makes AI valuable in production is your governance rules, the human in the loop, your business context and end-to-end traceability.
Method
We do not test the code.
We test the need.
From your data through to the delivered product, a chain of five links, each one led by an expert: the Product Builder. The method is encoded, not improvised, which is what makes it repeatable from one project to the next.
The Product Builder
One expert, human and AI combined. They steer, orchestrate and sign off the deliverables at every step.
They do not carry the knowledge, the method or the engine alone.
Your data and your business context, governed and sorted by relevance. The lungs of Kaiyan.
The reasoning model, agnostic and interchangeable.
Multi-agent pipelines, orchestrated and supervised, running in parallel.
Data Product · App Product · AI Product, driven by a portfolio of business initiatives.
One rule of sequence, non-negotiable: the data product and the governance come first before natural language can produce reporting you can trust. On a chain of agents, the smallest error spreads: the butterfly effect of hallucination. That is why every link is checked, expert sign-off, end-to-end traceability, continuous supervision: the cost of an error is simply too high to work any other way.
FINOPS
FinOps by design: intelligence spend, under control.
What AI costs to run is about to become the big question of the coming months. Kaiyan was built from day one to steer that cost, not absorb it.
Usage → €
from the token spent
to the euro billed
A single screen tracks what each agent consumes, converted into euros, task by task. Model routing is automatic: the right model for the right job, with cost-optimization algorithms.
The kind of analytical view of intelligence spend that no CFO or CIO has today.
Engagement
Two ways in, depending on how far along you are.
Start with one, the other, or both. Both run on the same method and the same governance.
Best-of-breed AI, orchestrated. Kaiyan + Joule, Kaiyan + Cortex: we assemble the best agent from each platform rather than forcing a single one on you. The deeper the consultant's business expertise, the stronger the assembly.
FASTER
Augmented integration
Track 1
We stay the integrator on your S/4HANA conversion, data platform, EPM or CRM. Our consultants and Kaiyan deliver far faster. The easiest track to start, and the quickest to pay for itself.
HIGHEST VALUE
Agent deployment
Track 2
Two routes, often combined: switch on the native agents of the platforms you already run (Agentforce, Snowflake CoWork, SAP Joule), or build your own agents on your data and your processes. Kaiyan orchestrates the whole thing.
Production
We go live the SDD way:
we start from the need, not the code.
A pipeline is a chain of agents that takes on one development discipline end to end. Four specialized pipelines run in parallel, supervised and signed off by one of our experts.
The code
Dev pipeline
Lead Dev
Backend Dev
Frontend Dev
The tests
QA pipeline
Lead QA
QA Engineer
The deployment
Infra pipeline
Architect
DevOps
The docs
Infra pipeline
Tech Writer
Business Analyst
Human-in-the-loop
People sit inside the build of the agent, at the critical checkpoints.
We do not park teams inside your day-to-day operations. The expert steps in where the risk sits: when the agent is designed, and when sensitive actions are approved. Three modes, depending on how much autonomy you grant.
Mode 00
Proposed resolution
The agent documents the analysis, the root cause and the path to a fix. Nothing is executed in your systems: you act.
Mode 01
Controlled automation
Execution runs through standard APIs, once a human has approved the sensitive actions. Everything is logged.
Mode 02
Expert assistance
A real-time copilot for your teams, with no direct connection to your systems.
REAL EXAMPLE · TICKET HANDLING
A user cannot receive a purchase order. The agent diagnoses it: the order is sitting in “not approved” status.
In mode 00, it explains the issue and leaves the process owner to fix the status. In mode 01, once approved, it changes the status, saves, and goods receipt becomes possible. In mode 02, it coaches the support agent handling the ticket in real time, without touching the system itself. Same diagnosis, three levels of autonomy.
Deployment
The engine runs wherever your data requires.
The same agent runs on any of three confidentiality levels. You stay in control, from shared public to sovereign cloud, depending on how sensitive your data is and how far you want to commit.
For your sensitive data
Managed private
A closed model hosted in an EU region, with no reuse of your data.
Infrastructure
Infrastructure that keeps running, even with your laptop closed.
Agents run on persistent infrastructure. You kick off a run, you shut your machine, the agents carry on. The work is waiting for you the next day.
Persistent execution
Agents run on Kubernetes. Jobs carry on outside your session, state is stored, nothing is lost when you close your machine.
System connectivity
Agents read, write and execute in SAP, Salesforce or Snowflake through a standard, secured protocol.
Clean Core on SAP
Everything goes through standard extensions, with no change to the SAP core. Maintainability and upgrades stay intact.
Value roadmap
Start small. Prove it. Then scale.
Rollout follows a curve: start with a POC, integrate, optimize. Real, measured gains at every phase, never an exponential promise.
TARGETED GAINS
POC
Phase 01
First tests, tight scope. We prove the value on one case and measure the starting point.
GROWING GAINS
Integration
Phase 02
Connection to the systems, scaling up, migrating history. The value spreads into live workflows.
OPTIMIZED GAINS
Optimization
Phase 03
Feedback loops, controlled automation, quality sharpened continuously. Cost becomes steerable, quality closes in on the optimum.
Augmented application maintenance (AMS)
Installed at the client, Kaiyan plugs its engines into your ticketing system. It picks up the knowledge base of your systems (SAP, OneStream, Pigment) and your configuration. A ticket can end three ways:
Resolved
The ticket is handled end to end, once sensitive actions have been approved.
Diagnosed
The root cause and the path to a fix are handed over; you execute.
Escalated
The agent proposes, but hands the decision back to the process owner on critical cases.
Our commitments
Your requirements,
already built into the architecture.
Nothing is bolted on afterwards: every guarantee follows from what comes before. These are what get Kaiyan through a CIO's door.
Sovereignty
Three confidentiality levels, from public to sovereign cloud, with separate dev / test / prod clusters.
No reuse
Your data stays yours. It never trains a third party's model.
Agnosticism
Automatic model routing, no lock-in: the engine can be chosen, rented, replaced.
Compliance
GDPR and EU AI Act rules encoded into the method, applied at every step.
