Operating surfaces
Where people and approved agents ask, review, and act.
Website
CMS
Built-in Chatbot
Relationship Intelligence
Analytics
Internal tools
Why a connected layer
AI earns its keep when the whole system shares context.
A bolt-on chatbot starts from a blank prompt and copied context. This is the opposite: a controlled layer between where work happens and the systems that hold the source of truth.
Connected AI layer
Where context, routing, generation, tools, and approvals come together.
Intent
Context assembly
Brand Soul
Knowledge
Routing
Review
Open interfaces
How models, agents, applications, and events reach the system.
MCP
A2A
OpenAPI
Direct APIs
Webhooks
Business systems
The owned records, content, documents, and services that ground the work.
CMS content
Relationship Intelligence records
Analytics
Documents
Calendars
Services
Permissions
Privacy
PII rules
Budgets
Rate limits
Audit logs
Governance spans every layer
From request to reviewed result
Every request takes the same governed path.
Context in, reviewed work out. Each stage narrows the task: what is asked, which context is relevant, what may be accessed, and where a person decides.
StageWhat happensIn this example
Understand
Interpret the request, the module it came from, and the outcome expected.
"Which stalled opportunities need a follow-up this week?"
Assemble
Pull only the brand, business, and relationship context the task needs.
Qualified opportunities, recent activity, service knowledge, Brand Soul.
Route
Pick the model, specialist agent, retrieval path, or deterministic tool that fits.
Relationship Intelligence specialist, internal knowledge, draft-only access.
Apply boundaries
Enforce source scope, role, consent, read-versus-write access, and budget.
No records changed. Nothing sent.
Prepare
Return an answer, structured output, or a proposed action for the workflow.
Account-specific follow-ups, each with why the record was selected.
Record
Log the outcome, cost, latency, feedback, and any knowledge gap.
6 drafts queued. The owner edits, approves, or discards each one.
Command Center
Ask once. The routing stays visible.
An SEO suggestion does not need the same model, context, or permissions as a relationship analysis. Command Center sends each task down the path that fits, so no request runs through one oversized prompt.
01
SEO metadata suggestion
Writing profile
Reviewed title and description
02
Relationship analysis
Relationship Intelligence specialist
Summary and proposed follow-up
03
Availability request
Deterministic tool
Structured available times
Shared context
Every module starts from the same understanding.
The system assembles the context a task needs instead of sending every request through one oversized prompt. Brand Soul and approved knowledge keep the output closer to your language and your facts.
01Primary context
Brand Soul
Voice, audience, positioning, preferred vocabulary, prohibited phrases, claims guidance, content pillars, and approved examples.
02Controlled grounding
Approved business knowledge
Published content, curated Q&A, uploaded documents, structured business information, and selected connected sources.
03When permitted
Relationship context
Current page, Relationship Intelligence contact context, prior conversations, recent activity, and visitor signals where consent and project rules allow.
04Task state
Operational context
The initiating module, requested action, expected output, current workflow state, available tools, and applicable review rules.
Content Intelligence
Missing context becomes visible work.
Weak retrieval, negative feedback, and repeated unanswered questions can become knowledge gaps. A dedicated Content Intelligence workflow can then prioritise a source update, curated answer, or draft brief for editorial review.
Governance and observability
Boundaries the AI cannot cross. A record you can inspect.
AI capabilities inherit the boundaries of the business workflow. Operators can see what was used, what happened, what it cost, and what still needs review.
01Control
Access
Module scopes, tool permissions, read-versus-write access, and public-versus-internal knowledge.
02Control
Privacy
Consent-aware context, configurable PII masking, retention rules, and controlled provider data flows.
03Control
Approval
Confirmation and human review before consequential publishing, sending, booking, or record changes.
04Control
Spend
Task-level model selection, rate limits, budget alerts, optional hard stops, and visible estimated cost.
05Control
Audit
Initiating module, provider, model, retrieved context, tool activity, status, latency, tokens, and errors.
06Control
Resilience
Failure classification, provider health, fallback behaviour, and visible items that need attention.
Request record
See what the AI used, did, and cost.
The exact fields depend on the build. The principle stays the same: consequential AI work should leave an inspectable operational record.
Module
command-center
Task
ris.follow-up.prepare
Provider
configured by task
Context
internal · 4 sources
Tools
ris.query · knowledge.search
Approval
required
Outcome
6 drafts queued
Status
completed · nothing sent
Open by design
Connected through open interfaces, not locked to one vendor.
Your architecture, knowledge, and workflows should not be trapped inside one model vendor's application.
01MCP
Models using approved tools and context.
A consistent interface for permitted assistants and internal copilots to search knowledge, query records, or prepare reviewed work.
02A2A
Agents exchanging structured tasks.
Agent2Agent workflows can support discovery, delegation, progress updates, and structured results between approved agents.
03APIs
Predictable actions staying predictable.
Typed APIs and webhooks handle forms, Relationship Intelligence updates, bookings, publishing events, and synchronisation without forcing every action through a model.
MCP for tools. A2A for agents. APIs for predictable actions.
Connected modules
One intelligence layer. Multiple working surfaces.
FAQ
Before you put AI inside the workflow.
The right architecture depends on the systems involved, the data they contain, and which decisions must remain with people.
01Is AI Engine a standalone SaaS product?
No. AI Engine is a set of reusable foundations Alapan configures inside a custom website, platform, Relationship Intelligence workflow, content system, or internal tool. The project is scoped around the client's data, team, infrastructure, and review process.
02Which AI providers can be used?
A build can support Anthropic, OpenAI, Gemini, or a compatible custom provider. Models can be routed by task so retrieval, structured output, writing, chat, and tool-assisted workflows do not all need the same configuration.
03Does all business data stay inside the application?
Provider keys stay server-side and each workflow controls which context is sent. When an external model provider is used, relevant request data may be processed by that provider. Data flows, retention requirements, and provider terms are reviewed during project design.
04Can AI publish content or change records automatically?
Only when a project is deliberately designed and permissioned for that behaviour. The default is reviewed assistance: AI can retrieve, explain, summarise, draft, and propose; people approve consequential write actions according to the client workflow.
05Do we need every AI Engine capability?
No. Shared context, retrieval, specialist agents, Command Center, open interfaces, provider routing, observability, and governance controls can be configured or omitted around the operational problem being solved.
06Does approved knowledge eliminate hallucinations?
No. Controlled grounding improves the context available to a model, but results still depend on source quality, retrieval quality, model behaviour, and review. Weak retrievals and negative feedback can be logged so the system can improve over time.
Build with Alapan
Build AI into the system, not beside it.
Connect your knowledge, models, tools, and workflows through a governed AI layer designed around how your business actually operates.
Start a project