BloomSpec

Enterprise Data Grounding

The problem of assembling live, structured, transactional enterprise data — a CRM opportunity, an SAP invoice, a ServiceNow ticket — correctly matched, permissioned, and formatted, into a prompt for a generative AI tool, across systems no single native platform grounding tool spans. Scoped to the cross-system, vendor-neutral layer specifically: native single-platform grounding (Salesforce Agentforce within Salesforce, Microsoft Copilot Studio within the Graph) is real and adjacent, but structurally excluded by definition, since it never reaches a system the platform doesn't already own.

113 pages

Agent patterns4

Compliance Evidence Compiler Agent
Manual reconstruction of this evidence during an actual examination is exactly the burden Vantage and Colton's own scoping notes describe; producing it as a…
Entity Resolution Maintenance Agent
Resolving entities only at request time means the same ambiguous pair gets re-evaluated on every query that touches it; maintaining the match set as a standing…
Grounded Context Assembly Agent
This is the domain's own core value proposition made concrete: the QBR scenario in the source material that takes 30–90 minutes manually is exactly this…
Prompt Governance & Routing Agent
Governance evaluated after the fact cannot prevent the action it's meant to govern; the policy engine has to sit in the path of the request itself, the same…

AI opportunities16

Clarifying-question generation for underspecified requests
Producing a targeted, specific clarifying question when parsed intent is ambiguous, rather than a generic 'please clarify' prompt.
Classifier-based sensitive-field detection in unstructured text
Detecting sensitive content (an SSN, a diagnosis code) that lands inside a free-text field never explicitly tagged as sensitive by schema.
Deterministic and probabilistic record linkage with LLM-assisted disambiguation for near-misses
Matching the same real-world entity across systems using rule-based keys first, fuzzy matching second, and LLM judgement only for genuinely ambiguous remaining…
Deterministic parallel API call orchestration
Fanning out calls to multiple connected systems in parallel and joining their results against a resolved intent.
Instruction/data channel-separation and anomaly detection for retrieved content
Keeping retrieved data structurally separate from instructions in the assembled context, and flagging anomalous embedded instructions before they reach the LLM.
LLM-assisted prompt drafting suggestions for template authors
Suggesting improved phrasing or structure to a prompt creator drafting a new template, without authoring the template autonomously.
All 16 AI opportunities →

Capabilities29

Ambiguity Disambiguation & Clarification
Detecting when a parsed intent is underspecified or has more than one plausible resolution, and either asking a targeted clarifying question or offering the…
Anomaly & Misuse Alerting
Flagging usage patterns that look like misuse or excessive access before they're discovered in a periodic review — one user's prompts pulling an unusual volume…
Compliance Evidence Generation
Producing regulator-ready documentation directly from provenance logs — which model, which team, which decision, at what cost and latency — as a byproduct of…
Connector Resilience & Degradation Handling
Continuing to produce a usable, clearly-labelled partial result when one or more connected systems are unreachable, rather than failing the whole request — and…
Cross-System Entity Resolution
Determining that a record in one system and a record in another refer to the same real-world entity, in the absence of a shared universal identifier — the…
Cross-System Permission Normalisation & Enforcement
Translating each connected system's own permission model — SFDC sharing rules, SAP authorisation objects, ServiceNow ACLs — into one filterable model, then…
All 29 capabilities →

Entities8

Canonical Entity Match
The resolved, cross-system identity record for one real-world entity — the confirmed link between "Atlas Fabricators" in a group CRM and the same policyholder…
Context Bundle
The assembled block of retrieved, resolved data attached to a Grounded Prompt Request before it reaches the LLM — structured text, YAML/JSON, or a short…
Data Field Mapping
A single crosswalk entry translating one field in one connected system's native schema into the canonical model — SAP's `KUNNR` and a legacy system's…
Data Source Connection
A registered, credentialed connection to an external system — an SFDC instance, an SAP instance, a hospital's EHR — including which protocol it exposes (MCP,…
Grounded Prompt Request
One specific instance of a prompt — whether invoking a registered Prompt Template or ad-hoc — being resolved, grounded, and executed. Carries the resolved…
Policy Route Decision
The record of how a specific Grounded Prompt Request was classified and routed — executed immediately, logged and auto-approved, held for review, or blocked —…
All 8 entities →

Glossary15

Canonical Entity Match
A confirmed record linking two or more systems' representations of the same real-world entity — the same company, the same patient — when no shared identifier…
Channel separation
Keeping retrieved data structurally distinct from instructions in the assembled context, so a model cannot mistake content it retrieved for something it was…
Confused deputy
A failure mode where an AI system, unable to distinguish trusted instructions from untrusted retrieved data, acts on an instruction hidden inside content it…
Context Bundle
The assembled block of retrieved, resolved data — structured text, a table, or JSON — formatted for injection into an LLM prompt.
Data Field Mapping
A single crosswalk entry translating one field in a connected system's native schema into the domain's canonical model.
Data Source Connection
A registered, credentialed connection to an external system, including which protocol it exposes and which canonical fields it can answer for.
All 15 glossary →

KPIs13

Anomaly alert precision
Share of anomaly/misuse alerts confirmed as genuine issues upon review, rather than false alarms.
Compliance evidence completeness
Share of grounded prompt requests for which a complete, field-level provenance record can be produced on demand.
Cross-system entity match accuracy
Share of automated cross-system entity matches confirmed correct in a sampled audit.
Data residency violation count
Count of any grounding event where data tagged with a residency or sovereignty constraint was routed to a processing endpoint outside its permitted zone.
Grounding-enabled task completion rate
Share of the domain's target recurring tasks (QBR-style summaries, headcount justifications, escalation responses) completed using the grounding platform…
Injection/exfiltration catch rate
Share of known confused-deputy / data-exfiltration attempt patterns, run against the pipeline in a red-team evaluation, that are actually caught before…
All 13 KPIs →

Personas4

Care Coordinator
Coordinating a patient's care across departments and, since the merger, across a hospital system that isn't yet on Praxis's own EHR.
Claims & Underwriting Business User
Producing accurate, timely policyholder and broker updates, using whatever data she can actually get to across group and acquired-subsidiary systems.
Platform Engineer
Integrating grounded data into Halworth's internal tools and customer-facing AI features, calling the grounding API directly rather than through a…
Systems Engineer
Specifying and validating engineering requirements against part numbers, requirement IDs, and configuration baselines, entirely within an air-gapped…

Products5

Agentforce (Data 360)
Salesforce's agent platform, grounded via the Data 360 unified data layer.
Glean
The unified Glean platform — search, assistant, and connector layer.
Kosmoy
Self-hosted AI governance and audit control plane.
Microsoft 365 Copilot / Copilot Studio
Microsoft's AI assistant and agent-building platform, grounded via Microsoft Graph.
Worqlo
Self-hosted conversational AI workspace for CRM/ERP data.

Scenarios3

A broker asks about a policy that predates the acquisition
A Vantage claims adjuster gets a plain-language request from a broker about a policyholder who originally held their policy with an acquired regional insurer.…
A confused-deputy attempt is caught before it exfiltrates patient data
A Praxis care coordinator asks for a routine patient visit summary. A record retrieved to answer it contains a hidden instruction, planted in a free-text field…
A platform engineer wires grounded data into an internal tool, and a source goes down mid-request
A Halworth platform engineer builds a feature against the grounding API using exact schema fields rather than natural language, and one of the two systems his…

Solution areas7

Access Control & Data Protection
Ensuring grounded data only ever includes what the requesting person or system is authorised to see, in a form that respects field-level sensitivity, and never…
Context Assembly & Prompt Composition
Formatting retrieved, resolved data into a form an LLM can actually use — structured blocks, reusable snippets, task-specific bundles — and letting the…
Cross-System Retrieval & Entity Resolution
Calling the right systems, and correctly matching the same real-world entity — an account, a case, an employee — across systems that disagree on identifiers,…
Intent & Query Understanding
Turning a business user's natural-language ask, or a program's structured request, into a resolved statement of what enterprise data is actually needed — which…
Observability & Audit
Recording what data, from which system, grounded which prompt, for whom, and making that record usable — for usage analytics, for anomaly detection, and for…
Policy Enforcement & Risk Routing
Classifying a grounded prompt by sensitivity and provenance, and routing it down the execution path that classification earns — immediate execution, logged…
All 7 solution areas →

Vendors5

Glean
Enterprise AI search and agent platform connecting to 100+ SaaS tools, with a Knowledge Graph treating every connected application as a first-class data source.
Kosmoy
AI governance control plane sitting above gateways like Portkey and Kong — AI inventory, regulatory evidence generation, and agent containment — self-hosted.
Microsoft
Microsoft 365 Copilot and Copilot Studio, grounding responses in work data and organisational context pulled through Microsoft Graph, alongside general web…
Salesforce
Agentforce, grounded via Salesforce Data 360 (formerly Data Cloud) — a unified layer combining structured CRM records, unstructured knowledge content, and…
Worqlo
Self-hosted conversational AI workspace for revenue teams, connecting natively to CRM (Salesforce, HubSpot, Zoho) and ERP (Odoo) systems, running entirely…

Verticals4

Defense / Aerospace Engineering
Engineering contractors under ITAR and FedRAMP obligations, where an AI governance board typically precedes any grounding capability by design.
Healthcare / Hospital Networks
Hospital systems formed by mergers, where PHI forces air-gapped grounding and per-hospital EHR fragmentation compounds the domain's own entity-resolution…
Insurance
Regional and national insurers, frequently grown by acquisition, where policy and claims history is scattered across systems inherited from acquired carriers.
Software / B2B SaaS
Technology companies with deliberately best-of-breed tooling and, typically, an AI gateway and prompt registry already operating before structured-data…