It stalls at pilot
Most generative-AI proofs of concept never reach broad deployment, and a large share of agentic AI projects are expected to be canceled outright.
CONVERSATIONAL CONTROL PLANE
A single query reaches every system you run.
A conversational control plane placed between people and enterprise systems. Users do not learn the system; the system understands the conversation.
- 12
- Analysis models
- 5
- Autonomous agents
- 3s
- Response target
- 5
- Logical layers
01 / THE PROBLEM
Few chatbots complete the work
Analysts converge on the same cause of failure: not model quality, but operating structure. The problem is not the conversation — it is that nothing happens after it.
It ends at the answer
Retrieval chatbots relabeled as agents are a widely noted pattern. Without an execution path and permissions, you have built one more search box.
Every integration is a new project
Rebuilding a custom integration per system makes cost scale with the number of targets, and every stack change forces rework.
Governance is missing, so it cannot be enabled
Few organizations have mature governance for autonomous agents, and most stalled projects name governance and security as the bottleneck.
Orbiton.AI Agent solves these four through product structure, not through a conversational interface.
02 / CATEGORY
From a search box to a control plane
Conversational AI has moved through four stages. Orbiton.AI Agent sits at the last one.
- 01FAQ chatbot“It finds”
Fixed scenarios and keyword responses
- 02RAG assistant“It cites”
Document retrieval with summary and sources
- 03Copilot“It suggests”
Calls tools to read state and recommend action
- 04Conversational control plane“It executes and proves”
Policy, approval, execution, verification and evidence — completed inside the conversation
Orbiton.AI Agent
The global market splits into ecosystem-embedded products (Microsoft Copilot Studio, Salesforce Agentforce, ServiceNow, Google Agentspace), knowledge-retrieval products (Glean) and customer-service specialists (Sierra, Decagon, Moveworks). Embedded products weaken sharply outside their own ecosystem, retrieval products hold no execution rights, and specialists depend on a vendor cloud. Orbiton.AI Agent targets the position that belongs to no ecosystem while still owning execution and control.
03 / OVERVIEW
Turn a query into executed work
Web, Slack, Teams, API or voice — every channel connects to the same engine.
It understands
It decomposes intent, target, scope and time from a natural-language query, then selects the knowledge and tools it needs and builds an execution plan.
- Intent decomposition
- Planning
- Agentic RAG
- Short and long-term memory
It acts
It does not stop at retrieval. It calls tools in internal and external systems over standard protocols, and executes only what policy and approval allow.
- MCP tool calls
- Policy evaluation
- Approval gate
- Verification and rollback
It learns
Conversations and action outcomes become training material. The evaluation set grows on its own, and only improvements that pass validation reach production.
- Feedback capture
- Evaluation-set growth
- Shadow validation
- Tenant isolation
Kafka is simply the one driver connected in the first customer environment. Target systems are swapped as plugins.
04 / ARCHITECTURE
Five logical layers
Layers couple only through standard interfaces, so replacing one leaves the rest running.
05Evidence layerEvidence
- Immutable conversation and rationale store
- Action audit trail
- Regulatory reporting
- Automatic ticket attachment
04Control layerControl
- Policy engine (OPA/Rego)
- Approval gate (HITL)
- Guardrails and injection defense
- Verification and auto rollback
03Connectivity layerConnectivity
- MCP tool servers
- A2A and ACP agents
- System drivers
- Business connectors
02Knowledge and memory layerKnowledge
- Agentic RAG
- Knowledge graph
- Short and long-term memory
- Permission-aware index
01Reasoning and planning layerReasoning
- Intent and scope decomposition
- Execution planning
- Tool selection and routing
- Self-verification
05 / CAPABILITIES
Six core modules
Understanding and planning
Intent decompositionIt breaks one sentence into intent, target, period and permission, then plans how many steps a response requires.
Knowledge retrieval
Agentic RAGResolve queries a single search cannot answer. Retrieval, evaluation and re-retrieval repeat, and queries about entity relationships go to the knowledge graph.
Short and long-term memory
Working · episodic · semanticConversation context, past incident history and the organization’s vocabulary and rules are held in three separate layers, so nothing needs explaining twice.
Tool execution
MCPExtend into new systems without retooling. Tools in internal and external systems are called over a standard protocol, and a new system is integrated by registering a tool, with no code change.
Automatic learning loop
Active learningConversations, feedback and action outcomes feed back into the evaluation set and index, and only validated improvements are promoted.
The conversation is the evidence
Evidence by conversationWhat was used as grounds, what was decided, who approved it and what was executed remain as one conversational record, submitted to audit as-is.
The six modules form the single path that processes one query.
06 / HOW IT WORKS
Seven steps from query to response
Every query passes the same seven steps. No path skips a step.
- 1Interpret intent
Decompose the query into intent, target, period and required tools
- 2Check permission
Determine the reachable scope from the user role and tenant boundary
- 3Gather evidence
Query knowledge, memory and live systems as many times as needed
- 4Reason and self-verify
Re-query when the response and the evidence disagree
- 5Evaluate policy
When action is needed, policy decides allow, deny or approval-required
- 6Approve and execute
After approval, set an idempotency key and call the tool
- 7Respond and record
Present the response, evidence and outcome together, then store them
No response without evidence
If retrieval does not support the query, the agent reports that it cannot confirm rather than guessing.
Nothing outside the permission scope
Permissions are stored at index time, so documents the user cannot see never appear in results.
Execution always passes the approval gate
Proposal and execution are separated; an action that fails policy or approval is never called.
The response target is three seconds. Queries that include an action are measured excluding approval wait time.
07 / KNOWLEDGE & MEMORY
Retrieve, evaluate and retrieve again
Conventional RAG retrieves once and summarizes, which does not resolve complex business queries. Keyword and semantic search run together, reranking filters the result, and the agent searches again when that is not enough.
Retrieval
- Keyword search BM25
- Handles queries where the term must match exactly — error codes, device names, defined vocabulary.
- Semantic search pgvector
- Embedding search retrieves documents that carry the same meaning in different wording.
- Reranking BGE
- Merges both result sets and reorders them by relevance to the query.
- Repeated retrieval Agentic loop
- When the result does not satisfy the query, the agent reformulates it and searches again.
Memory
- Working memory
- Holds the current conversation; when the session ends, only a summary is retained.
- Episodic memory
- Holds past incidents, the actions taken and their outcomes, and is consulted first in similar situations.
- Semantic memory
- Holds the organization’s vocabulary, rules and runbooks, and is updated only after validation.
Keyword search is precise but misses paraphrase; semantic search is flexible but returns loosely related documents. Combining both and filtering by reranking outperforms either alone — a result confirmed repeatedly in information retrieval research. Agent-controlled iterative retrieval and conversation-level memory on top of that resolve the business queries a single search cannot.
08 / AI TOOLBOX
Twelve analysis models the agent calls directly
The agent does not only reason on its own. It selects a purpose-built model for the query, calls it, and responds from that result.
Forecasting and judgment
| Model | Role | Implementation |
|---|---|---|
| LSTM | Forecasting | PyTorch · PyTorch-Forecasting |
| Prophet | Forecasting | Prophet · holidays-KR |
| Temporal Fusion Transformer | Forecasting | PyTorch-Forecasting (Lim 2021) |
| N-BEATS | Forecasting | PyTorch (Oreshkin 2020) |
| Anomaly ensemble | Anomaly | Isolation Forest · USAD |
| Causal inference | Cause | EconML · DoWhy (Pearl 4-step) |
Learning and quality
| Model | Role | Implementation |
|---|---|---|
| RL policy | Action selection | Stable-Baselines3 PPO |
| Federated learning | Distributed training | Flower FedAvg |
| DP federated learning | Privacy | Opacus DP-SGD · Flower |
| Continual learning | Knowledge retention | EWC (Kirkpatrick 2017) |
| Drift detection | Quality monitoring | Evidently PSI·KS · River |
| Hyperparameter tuning | Search | Optuna TPE sampler |
“Why has processing latency risen since last week?” — that one sentence calls the forecasting, anomaly and causal models in sequence. The user never needs the model names.
09 / AGENTS
Agents that carry out defined roles on their own
Above a chatbot that waits for queries sit agents that carry out defined roles on their own — all on the same knowledge, tools and policy.
Classifies incoming incidents first
On a new incident it judges severity and candidate causes, attaches similar past cases and likely owners, and places it in the queue.
Closes out finished incidents
It gathers the timeline, rationale and action history of a closed incident, drafts the root-cause document and proposes prevention items.
Assists the on-call engineer continuously
It organizes the context of alerts raised during on-call hours before delivering them, and guides the procedures and approval paths needed at night.
Writes the weekly operations review
It aggregates the week’s incidents, actions and SLA metrics into review material, leading with the items whose trend has worsened.
Feeds operator judgment back into the model
Approvals, rejections and corrections made by operators are collected as ground truth to raise the accuracy of the next decision.
Agents write and organize documents on their own, but any action that changes a system still passes policy evaluation and human approval.
10 / INTEGRATION
Connect internal and external systems without retooling
Orbiton.AI Agent presumes no particular vendor ecosystem. The boundary where it meets external systems is defined by three standards; the conversation, reasoning and control engines are reused unchanged.
Tool invocation
Registering a tool completes the integration. The plugin marketplace supports one-click installation and version management.
- Query tools (metrics, logs, SQL)
- Action tools (execution APIs)
- Plugin marketplace
- MCP standard support
Agent-to-agent collaboration
Agents with different roles exchange work on a single graph, and external agents connect over the same standards.
- LangGraph multi-agent
- Role-based task division
- HITL approval linkage
- A2A and ACP support
Business system integration
A standard event schema and webhooks keep your existing processes in sync both ways.
- ITSM, ticketing and approvals
- Groupware and collaboration channels
- Databases and warehouses
- SIEM and internal APIs
Example targets — messaging (Kafka, RabbitMQ, Pulsar), databases, data warehouses, Kubernetes, API gateways, ITSM, groupware, in-house systems, external SaaS and document stores. A target is integrated through a driver that implements only four methods: resource discovery, metric collection, action execution and schema sampling.
11 / LEARNING
Operational data becomes the next version’s training set
Learning runs continuously rather than as a manual batch process. Conversations and action outcomes become the material for the next version.
- 1
Collect
Structure and store queries, grounds, responses and action outcomes
- 2
HITL active learning
Capture operator feedback and approval decisions as ground truth
- 3
Update the evaluation set
Failed cases are registered automatically into the regression set
- 4
Retrain
Optuna re-searches; EWC updates without discarding prior knowledge
- 5
Watch for drift
PSI, KS and Page-Hinkley track input distribution and performance change
- 6
Promote
Applied only after shadow validation passes; otherwise the previous version stays
Customer data never leaves
Tenant data never crosses its boundary. When cross-organization learning is needed, Flower federated learning collects results rather than data, and differential privacy masks individual contributions.
Learning never degrades quality
Every update must pass the regression evaluation set. If accuracy, refusal rate, latency or cost falls below its bar, the previous version is kept automatically.
What changed and why is recorded
Training rounds, participation scope, model versions and evaluation results are logged, so model changes are auditable too.
12 / CONTROL
Execute with governance and control
Before conversational AI can change a real system, the control structure matters more than the answer.
- Proposal separated from execution
- The AI only proposes; execution rights belong to the policy engine and the approver. That boundary cannot be bypassed by configuration.
- Policy managed as code
- Automatic-execution permission, approval requirement, attempt budget, cooldown and the per-risk action list are externalized into versioned policy files.
- Duplicate and repeat execution blocked
- An idempotency key prevents the same action from being issued twice, while budgets and cooldowns prevent repeating an action on the same incident.
- Automatic rollback on failure
- Success criteria are verified after an action; falling short reverts the change and escalates to an operator.
- Separation of roles
- Viewing, requesting, approving and configuring are separate permissions. Approval requires a comment.
- Data boundary enforced
- Data outside the tenant and system scope is used nowhere — not in retrieval, not in references, not in responses.
13 / SECURITY
A countermeasure for each published threat class
The risks of conversational AI are already catalogd. Countermeasures are mapped to the LLM and agent threat taxonomies published by the international security community.
| Threat class | What goes wrong | Orbiton.AI Agent response |
|---|---|---|
| Prompt injection | Instructions hidden in documents, tickets or web pages manipulate the AI | Instruction/data separation, tool allowlists, HITL approval for high-risk actions |
| Sensitive-information disclosure | Private information leaks through question context or the answer | RBAC-aware indexing, enforced tenant boundaries, masking before response |
| Excessive agency | The AI holds more execution authority than it needs | OPA Rego policy evaluation, least-privilege scopes, execution budget limits |
| System prompt leakage | Internal rules, model names or endpoints surface in an answer | Errors and internal identifiers blocked from responses, user-facing messages separated |
| Vector and index weakness | A poisoned document enters the index and distorts answers | Source trust scoring, ingestion validation, mandatory citation links |
| Agent tool misuse | An unintended tool is called during multi-step execution | UEBA behavioral analysis, audit hash chain, anomalous-call blocking |
Threat names and numbering follow the public standard taxonomy; the countermeasures reflect the product implementation. Regulatory compliance is operated alongside a separate management system covering risk, audit and documentation.
14 / EVALUATION
Measure quality change continuously
The most common failure in conversational AI is that nobody notices the degradation. Measurement is built into the product.
- Regression evaluation set
- Failed queries are registered into the evaluation set automatically, and every later update must pass it.
- Dual scoring
- Items with a checkable answer are scored automatically; items requiring judgment use model scoring together with human review.
- Groundedness measurement
- Whether each sentence is supported by the cited evidence is measured separately to track hallucination rate.
- Conversation-level tracing
- OpenTelemetry distributed tracing binds query, tool calls and response into one unit, queried in Loki and Grafana.
- Cost observability
- Tokens and tool-call cost per query are aggregated by tenant, so cost cannot grow unnoticed.
- Load regression testing
- A load-test harness measures the latency distribution across all twelve inference endpoints and compares each release with the previous one.
Telemetry is exported in the standard format, so it is queried directly from the observability stack you already run.
15 / PERFORMANCE
Performance targets defined as design baselines
These are the product’s design baselines. Conversational inference latency and concurrency depend directly on customer infrastructure, so actual figures vary by environment.
Under 3 sEnd-to-end response
From query receipt to final response, excluding approval wait
p95 under 2 sModel inference latency
Measured on the completed response, not the first streamed token
20 concurrentConcurrent users
Concurrency held while the same response standard is maintained
10+ QPSQuery throughput
Achieved through async processing paths and horizontal scaling
- NVIDIA A100 · Standard inference tier
- The base unit handling conversational responses together with the twelve analysis models. Memory capacity determines concurrent session count.
- NVIDIA H100 · Large-scale tier
- Deployed where large models, long contexts or high concurrency are involved, with inference separated onto dedicated nodes.
- NPU alongside · Resource separation tier
- Continuously running lightweight inference — anomaly detection, classification, embedding — moves to NPU so GPU capacity concentrates on conversation and large models.
Design elements — fully asynchronous, model routing, prompt cache, horizontal scaling
16 / POSITIONING
How Orbiton.AI Agent relates to existing tools
Orbiton.AI Agent does not replace your business systems. It owns the layer above them that understands, executes and proves.
| Axis | Ecosystem-embedded | Knowledge retrieval | Orbiton.AI Agent conversational control plane |
|---|---|---|---|
| Copilot Studio · Agentforce · ServiceNow | Glean · NotebookLM | ||
| Strength | Deep integration inside its own ecosystem | Permission-aware internal knowledge search | A single path through understanding, execution and evidence |
| Execution scope | Workflows within the vendor product family | Retrieval and summarization | Every system registered as a standard tool |
| Connection model | Vendor connectors | Document source connectors | MCP and A2A standards plus in-house drivers |
| Approval control | Depends on platform workflow | No execution concept | In-product policy evaluation and approval console |
| Audit evidence | Platform logs | Search history | Immutable storage of conversation, grounds and action |
| Deployment | Vendor cloud first | SaaS first | On-premises and air-gapped included |
Customer-service specialists (Sierra, Decagon, Moveworks) compete on a different axis and form a separate category. The comparison is a positioning view based on public material, not a feature verdict.
17 / USE CASES
From query to execution, by function
“Why did payment processing slow down?”
It narrows the cause across metrics, logs and change history, executes scaling or restart under approval, then verifies and reports the outcome.
“Does this contract clause meet our standard?”
It cross-searches policies, contracts and history within permission, answers with citations, and raises the approval or ticket inside the conversation.
“Show me last month’s churn rate by region”
It converts natural language into a query, runs it, and presents the tables and conditions it used. Results appear as tables and charts.
“Check the order status and process the refund”
It consolidates state scattered across systems into one answer, and any task requiring authority goes through policy and approval.
The engine is identical. Differences between functions are absorbed by tool registration, knowledge sources and policy bundles.
18 / ADOPTION
Adoption path
Start with reading and widen execution as trust builds. Execution rights are never open on day one.
- PHASE 01
Connect and query
Connect existing knowledge sources and read-only tools. No execution tools are registered; only question and answer quality is assessed.
Output · An answer-accuracy baseline - PHASE 02
Approval-gated execution
Register execution tools and define policy. Every action passes human approval while verification and rollback paths are exercised for real.
Output · Action success rate and rollback behavior verified - PHASE 03
Selective autonomy
Relax policy so low-risk actions with a proven success rate run without approval. The rest keep the approval gate.
Output · Shorter handling time, less operator intervention - PHASE 04
Expand and learn
Widen tools and knowledge sources into adjacent functions, and refresh retrieval, routing and models on accumulated conversation data.
Output · Adopted as the company-wide work interface
Each phase is switched by a feature flag; lowering it returns you to the previous phase immediately.
19 / STANDARDS
Standards and research behind the design
The product structure follows open standards and recent research findings.
Standards and frameworks
- MCP
- The agent-to-tool connection standard, transferred to an open foundation and adopted across major AI platforms
- A2A · ACP
- Standards for delegating work and exchanging messages between agents, with broad enterprise vendor participation
- OpenTelemetry
- The standard for exporting conversation and tool-call traces into the customer’s observability stack
- Policy-as-Code
- The open policy engine approach that separates decision from enforcement and versions both
- OWASP threat taxonomy
- LLM and agent threat items used as the basis for countermeasure design
- AI risk management
- Risk identification, measurement, treatment and monitoring fixed as operating procedure (NIST AI RMF, ISO/IEC 42001)
- MAPE-K
- The reference model for the monitor, analyze, plan, execute and knowledge control loop
Research
- Graph-based retrieval
- Reported advantage over plain retrieval on queries that span documents (GraphRAG line of work)
- Agentic retrieval
- Accuracy improves when the model controls retrieval, evaluation and re-retrieval (Agentic RAG and RL-combined studies)
- Non-parametric continual learning
- Extending knowledge by updating memory structure rather than retraining
- Agent memory structure
- Separation of working, episodic and semantic memory, with long-conversation benchmarks
- Agent interoperability protocols
- Surveys comparing MCP, A2A and ACP with a staged adoption roadmap
- Prompt-injection defense
- Combining least-privilege tooling with human approval sharply reduces attack success rate
Standards and findings are cited from public material; the product description reflects the Orbiton.AI implementation.
20 / TECH STACK
Technology
Every layer runs on open source and open standards, and the model provider itself is isolated behind a replaceable interface.
Conversation · LLM
- LangChain 1.x · LangGraph
- SSE streaming responses
- Multi-provider routing
- Ollama · OpenAI · Anthropic · Gemini
Retrieval · memory
- pgvector embedding search
- BM25 (rank_bm25)
- BGE reranking
- Single PostgreSQL store
AI / ML
- PyTorch · PyTorch-Forecasting
- Prophet · Optuna · Evidently
- Stable-Baselines3 · Flower(FL)
- EconML · DoWhy · River
Backend
- FastAPI 0.115 · Python 3.x
- SQLAlchemy · Alembic
- Redis
- Seven microservices
Frontend · observability
- Vue 3 · Vite 6.0 · Pinia
- Bootstrap 5.3 · ApexCharts
- Prometheus · OpenTelemetry
- Loki · Grafana
Security · deployment
- JWT + OAuth2 · MFA(TOTP)
- RBAC · OPA Rego
- Audit Hash Chain · SIEM Hook
- Kubernetes(Helm) · GitHub Actions
21 / ROADMAP
Roadmap
Each item extends the existing structure rather than changing it. Because the layer interfaces are fixed, capabilities are added without replacing the engine.
GPU-accelerated inference
Move the inference service to a CUDA-based image to cut latency for conversations that include model calls.
Shadow-mode OOD
Automatically classify out-of-domain questions and detect topic drift, so the agent does not sound plausible while being wrong.
Streaming resilience
Strengthen connection retention and resumption over long responses to stabilize the conversation.
Mobile app
Handle conversation and approval from iOS and Android. The skeleton implementation is ready.
Advanced UEBA
Combine behavioral clustering with anomaly detection to detect misuse of conversation and tool calls more accurately.
Federated learning at scale
Improve model quality across multiple customers without sharing data, preserving privacy.
CONVERSATIONAL CONTROL PLANE
From conversation to action, from action to evidence.
The value of conversational AI is not fluency. Orbiton.AI Agent binds understanding, execution and evidence into one path and takes responsibility for that span.