We dey use analytics make di experience better. You fit update your choice any time.
Glossary
Definitions for di terms wey matter when building accountable AI work automation systems.
Action Proposal
platform
Action proposal na structured suggestion wey person fit review to change connected business system. Automation create am, but e never execute. E name target system, operation, and exact parameters, so person or policy fit approve, edit, or reject am before anything happen.
Why propose action instead of executing am direct?
Propose first separate intent from effect. E let approval policy and reviewers inspect di exact operation and parameters, so automated mistake no reach system of record.
Wetin action proposal contain?
Di target integration, operation to perform, resolved parameters, supporting evidence, and policy decision about whether approval dey required before execution.
Agent Delegation
security
Agent delegation na controlled grant of scoped, time-bound authority to AI agent so e fit act for user or another agent. Di delegation specify exact capabilities, tenants, and actions wey dey allowed, so agent operate under explicit, revocable, auditable limits.
Di capabilities agent fit use, tenant where e fit act, actions e fit propose or execute, and expiry, so authority narrow, time-bound, and revocable.
How delegation stay accountable?
Every delegated action dey attributed to both di agent and di delegating principal and recorded in audit trail, with sensitive actions still routed through approval policy.
Agent2Agent Protocol (A2A)
ai
Agent2Agent protocol na open standard for autonomous agents to discover one another, exchange tasks, and coordinate work across organization boundaries. E define how agent advertise capabilities and how another agent delegate task and track am to completion.
MCP connect model to tools and data. A2A connect agents to each other, defining how one agent hand task to another and follow status, instead of how model call one tool.
How A2A tasks dey tracked?
A2A task map onto tracked work record so im lifecycle, evidence, and outcome dey auditable, just like work wey come from person or form.
Answer Engine Optimization (AEO)
marketing
Answer engine optimization na practice of structuring content so AI answer engines and chat assistants fit find, cite, and summarize am correctly. Where SEO target ranked links, AEO target di synthesized answer itself through clear definitions, structured data, and machine-readable source files.
SEO optimize so page rank as clickable link on results page. AEO optimize so AI-generated answer select, quote, and cite content, rewarding precise definitions, structured data, and clean machine-readable feeds.
Which signals help answer engine cite page?
Definition-first writing, valid schema.org structured data, llms.txt index, FAQ markup, and stable canonical URLs all make content easier for answer engine to retrieve and attribute.
Approval Workflow
governance
Approval workflow na governed sequence of checkpoints wey proposed action must pass before e execute. Each step route di decision to right reviewer based on risk, role, or policy, and record who approve wetin so outcome fully accountable.
Requirements fit apply by workflow, channel, risk class, money threshold, or action type, so only steps wey genuinely need oversight pause for reviewer.
How approval workflow stay auditable?
Every request, approval, edit, and rejection dey recorded with actor and timestamp, producing end-to-end trail wey prove who authorize each governed action.
Chunking
ai
Chunking na process of splitting source documents into smaller retrieval units before embedding dem. Di chunk size and boundary strategy determine how precisely retriever fit locate relevant fact, balancing recall, precision, and embedding cost across knowledge base.
Good chunk dey semantically self-contained, sized so one fact no split across boundaries, and carry stable metadata so e fit be filtered, refreshed, and cited reliably.
How chunking affect answer quality?
Chunks wey too large dilute relevance and waste tokens, while chunks wey too small break context and lose meaning. Boundary choices directly shape recall and grounding of generated answers.
Embedding
ai
Embedding na numeric vector wey represent meaning of text, images, or other data inside high-dimensional space. Items with similar meaning produce vectors wey sit close together, so systems fit compare, cluster, and retrieve content by semantic similarity instead of exact matches.
Vectors from different models no comparable. Storing model version with each embedding let you detect drift and reindex safely when you upgrade embedding model.
Embeddings fit reverse back to original text?
No exactly, but embeddings fit leak sensitive information, so dem suppose inherit di same tenant isolation and access controls as source content wey dem represent.
Evidence Citation
ai
Evidence citation na practice of attaching verifiable source references to every claim AI system make. Each cited passage link back to document, record, or knowledge asset where e come from, so person fit confirm answer get grounding before trusting or acting on am.
At minimum, source identifier and exact passage used; ideally with stable link and timestamp so reviewers fit confirm evidence current when answer was produced.
Why citations essential for governed automation?
Citations make answer auditable. Without dem automated response no accountable, but with dem reviewer fit verify grounding and audit trail fit prove which evidence drove decision.
Grounding
ai
Grounding na practice of constraining AI model output to verifiable source evidence instead of im parametric memory. Grounded answer dey supported by retrieved passages wey fit be cited and checked, and na primary defense against fabricated or confidently wrong responses.
Retrieval supply model only with relevant source passages, prompt instruct am to answer from dat evidence, and verification step reject claims wey no get supporting citation.
Wetin happen when grounding evidence no dey?
Well-designed grounded system decline to answer or escalate to person instead of inventing response, surfacing explicit gap instead of confident guess.
Hallucination
ai
Hallucination na confident but unsupported or fabricated output from language model: claim wey sound plausible but no get basis in provided evidence or reality. Hallucinations na central risk in automating knowledge work, and grounding with cited evidence na primary mitigation.
Synonyms: AI hallucination, fabrication, confabulation, ungrounded output wey dey
Why language models hallucinate?
Models predict likely text, no be verified facts. Without retrieved evidence to constrain dem, dem fill gaps with statistically plausible but unverified statements.
How you reduce hallucination?
Ground answers in retrieved sources, require citations, verify claims against evidence, and route low-confidence or unsupported cases to person instead of returning guess.
Human-in-the-Loop
governance
Human-in-the-loop na design pattern where people review, approve, or correct AI system proposals before dem take effect. E keep human judgement on critical path for high-risk or low-confidence decisions while automation handle routine volume.
Synonyms: HITL, human in the loop, human oversight, human review wey dey
When step suppose be human-in-the-loop?
Whenever decision high-risk, irreversible, low-confidence, or governed by policy. Routine, well-grounded, low-risk steps fit run automatically with human reviewing exceptions.
How this different from full automation?
Full automation act without review. Human-in-the-loop insert explicit checkpoint where person fit approve, edit, or reject proposal, preserving accountability for sensitive outcomes.
Hybrid Retrieval
ai
Hybrid retrieval combine semantic vector search with lexical keyword search to retrieve relevant passages. Vector search capture meaning and paraphrase, keyword search capture exact terms and identifiers, and fusion step merge both result sets so precise tokens and conceptual matches no dey missed.
Vector search fit miss rare exact terms like SKUs or error codes, while keyword search miss paraphrases. Fusing both recover di strengths of each and raise recall on real-world queries.
How di two result sets dey combined?
Fusion method like reciprocal rank fusion or weighted score blend rerank merged candidates, often followed by cross-encoder reranker for final precision.
Intake Automation
platform
Intake automation na process of turning unstructured inbound requests into structured, machine-readable records without manual data entry. E classify request, extract fields wey matter, and route result into workflow so work fit be answered or actioned consistently.
Email, chat messages, web forms, uploaded documents, and synced records from connected systems fit all normalize into di same structured shape for downstream handling.
Intake automation replace people?
No. E remove manual data-entry and triage burden so people focus on judgement-heavy exceptions, approvals, and high-risk decisions wey policy route to dem.
Intent Classification
ai
Intent classification na step wey determine wetin inbound request actually dey ask for, mapping unstructured text to defined category of work. Accurate classification route each WorkItem to right workflow, evidence sources, and policy, making am foundation of reliable automation.
E decide di entire downstream path. Misclassified request retrieve wrong evidence and apply wrong policy, so classification accuracy gate di quality of everything wey follow.
How classification accuracy dey measured?
Through evaluation gates over labeled set, tracking precision and recall per intent and watching for confusion between similar categories before workflow go live.
Model Context Protocol (MCP)
ai
Model Context Protocol na open standard wey let AI assistants connect to external tools and data sources through uniform interface. MCP server expose typed tools and resources wey model client fit discover and call, so capabilities fit be added without custom code per integration.
Typed tools wey model fit invoke and resources wey e fit read, each described with schema and annotations so client fit discover capabilities and call dem safely.
Why MCP matter for governed automation?
E give external assistants standard, schema-described way to act on platform, so tool calls fit be validated, scoped to tenant, and routed through same approval policy as any other action.
MongoDB Vector Search
infrastructure
Di process of turning content into vector embeddings and storing dem in MongoDB Atlas Search vector indexes with metadata for efficient similarity search.
Why use MongoDB Atlas Search for knowledge retrieval?
MongoDB Atlas Search provide fast vector similarity search with ability to combine vector and traditional queries, integrated metadata filtering, and seamless scaling inside your existing MongoDB infrastructure.
Which metadata matter?
Store tenant ID, language, URL, content hash, updated timestamp, and model version to enable filtering, freshness checks, and controlled reindexing.
Retrieval-Augmented Generation (RAG)
ai
Retrieval-augmented generation na technique wey ground language model output in retrieved source documents instead of relying only on im parametric memory. Di system fetch relevant passages from knowledge base, supply dem as context, and ask model to answer using only dat evidence.
RAG keep knowledge inside external store wey you fit update instantly, so answers stay current and every claim fit trace to source. Fine-tuning bake knowledge into weights, which slower to refresh and harder to attribute.
Wetin RAG pipeline include?
Usually ingestion and chunking, embedding, index for vector or hybrid search, retriever, and generation step wey condition model on retrieved passages and return cited evidence.
Single Sign-On (SSO)
security
Authentication method wey let users access many applications with one set of login credentials through identity federation protocols like SAML or OpenID Connect.
E centralize identity, enforce enterprise security policies like MFA and conditional access, and speed up user provisioning across shells, packs, and governed workspaces.
SAML vs OIDC?
SAML dey XML-based and common for older enterprise stacks; OIDC, built on OAuth2, lighter and modern. Supporting both maximize compatibility with customer IdPs.
SLA Breach
operations
SLA breach happen when work miss commitment defined in service-level agreement, like response or resolution deadline. Detecting and escalating breaches automatically keep accountability visible and ensure at-risk work reach right people before commitments miss.
Each WorkItem carry im commitment timers, and system watch elapsed time against thresholds, raising escalations as deadline near and recording breach if e miss.
Wetin happen when breach dey imminent?
Policy fit escalate WorkItem, notify owners, or reprioritize queue so attention shift to at-risk work before commitment actually miss.
Tenant Isolation
security
Tenant isolation na guarantee say each customer data and configuration in multi-tenant system remain logically separated and inaccessible to other tenants. E dey enforced at every layer: storage, retrieval, and access control, so one organization never see or influence another work.
How tenant isolation dey enforced during retrieval?
Every query dey scoped to requesting tenant, and stored content carry tenant identifier so vector and keyword search fit return only that tenant own evidence.
Isolation na only about data?
No. E cover configuration, policy, embeddings, and audit logs too, so no aspect of one tenant work leak into another, even on shared infrastructure.
Vector Search
ai
Vector search find content by meaning instead of exact words. Text dey converted into high-dimensional embeddings, and similarity metric like cosine distance rank stored vectors by how close dem dey to query vector, returning related passages even when no keywords match.
Embedding na numeric vector wey represent meaning of piece of text, produced by embedding model. Texts with similar meaning land close together in vector space.
Wetin be approximate nearest neighbor (ANN) search?
ANN search trade small amount of accuracy for big speed gains, using index structures so similarity lookups stay fast as stored vectors grow into millions.
Vertical Pack
platform
Vertical pack na packaged configuration wey tailor platform to specific domain of work: im intents, extraction fields, evidence sources, policies, and actions. Packs let team launch focused workflow like IT access or vendor security without rebuilding underlying engine.
Di intents e recognize, fields e extract, evidence e ground answers in, approval policies e enforce, and governed actions e fit propose for dat domain of work.
Packs fit be customized?
Yes. Pack na starting configuration wey teams adapt in Studio: adjusting intents, prompts, evidence sources, and policies so e fit their real processes.
Work Packet
platform
Work packet na bundle of context assembled around WorkItem so people or automation fit reason about am and act on am: original request, extracted fields, retrieved evidence, applicable policy, and proposed actions. Na complete, self-contained briefing for one piece of work.
Synonyms: work bundle, context packet, task packet, work context wey dey
How work packet different from WorkItem?
WorkItem na tracked record of di request itself. Work packet na assembled context: evidence, policy, and proposals gathered around dat record to drive answer or action.
Why bundle context into packet?
Self-contained packet let model or reviewer make decision without hunting across systems, and preserve exactly which evidence dey available at decision time for audit trail.
WorkItem
platform
WorkItem na unit of work in Threada: one inbound request from email, chat, document, or form, normalized into structured, trackable record. Each WorkItem carry im intent, extracted fields, evidence, and complete history of every decision and action taken on am.
Synonyms: work item, task record, tracked request, unit of work wey dey
How WorkItem different from support ticket?
Ticket usually track conversation. WorkItem track di work itself: classified intent, extracted fields, evidence wey ground any answer, and governed actions taken, all auditable end to end.
Which lifecycle WorkItem dey move through?
Intake normalize request, intent classification route am, evidence retrieval ground proposed response, and any action pass through approval policy before WorkItem resolve and record.