AMANDIL · OPERATIONAL CONTEXT GRAPH

Foundational infrastructure for vertical AI agents.

One context graph per industry. Every agent that runs on top of it cites the regulation, standard, and benchmark it leans on.

Language models know language — not your operations. Generating the agent is the easy part now. Grounding it in an industry's real operations — every process, rule, regulation, and benchmark — is what nobody else has built. Amandil is that graph: the substrate agents query before they answer.

01
GROUNDED vs UNGROUNDED

Same prompt. Two agents.

A healthcare agent's answer is either grounded in the graph or it isn't. Here's the difference, side by side — same question, one model bolted to a workflow, the other querying the Operational Context Graph before it speaks. Healthcare shown; the same pattern instantiates against the standards and regulations of every vertical — see §04 Pattern.

WITHOUT A CONTEXT GRAPH
> agent: "Why is our denial rate elevated?"
"Elevated denial rates are typically caused by documentation issues, authorization gaps, and coding errors. Consider reviewing your front-end processes, training staff, and implementing better workflows to reduce denials."
GENERIC ADVICE · NO PROCESS NAMED · NO BENCHMARK · NO DEPENDENCY MAP
VERDICT · Fluent. Generic. Ungrounded.
WITH A CONTEXT GRAPH
> agent: "Why is our denial rate elevated?"
"Elevated denials map to L3.01.06.04 Denial Research & Analysis. The benchmarks to hold it to (HFMA MAP): write-off rate <30%, first-level appeal success >50%, throughput 15–25 denials/FTE/day. Root cause back-traces upstream to L2.04.01 Patient Registration and L2.08.01 Prior Authorization — where most preventable denials originate. The prevention loop spans 10 L3 steps under L2.01.06, each with its standard, owner, and governing regulation. Benchmark off public CMS data, or drop in your own — no integration required."
GROUNDED · HFMA MAP · L3.01.06.04 · L2.04.01 · L2.08.01 · L2.01.06
VERDICT · Grounded. Benchmarked. Actionable.
02
FIND · BUILD · SHIP

Find. Build. Ship. The path is the same regardless of who uses Amandil.

A health system uses it in-house. A GSI uses it on behalf of a customer. A platform vendor uses it inside their own product. All three ship agentic AI fast.

01 · Find
Find the gap.

Pick by benchmark, not vendor pitch. Start at the CMS gap, at the automation score across every process, or by describing the problem in natural language. The graph brings the right components to bear. Defensible on day one. Scoreable on day ninety.

'Prior authorization: 5x projected ROI vs the other six we ranked — closes a CMS interoperability gap the new mandate already requires us to address.'

02 · Build
Build the agent.

Massively compress the discovery phase — start at the gap, inject context, build the blueprint, validate against evals, generate the code. Prior authorization, already encoded with every standard, role, regulation, and benchmark. Done once. Yours in days. The spec is production-grade.

'Six weeks of discovery in days. The blueprint validated against our evals — runbooks, success criteria, HITL spec, integration map — and tied to a closeable CMS benchmark gap.'

03 · Ship
Ship it. You own it.

The agent is yours — and it runs where you run. Build once, deploy to your own cloud or any of six runtimes. It stays wired to the graph via MCP — the connection that keeps it grounded as the industry itself evolves: operations, standards, regulation, benchmarks, all of it. You own the agent; the graph keeps it current. An agent frozen at build time goes stale; one on the graph moves with the industry.

Most agents ship as islands. Built on Amandil, an agent has the graph behind it — failure modes, state machine, every standard and regulation. The graph stays in the loop, tuned as the reference evolves.

Six runtimes, one substrate. You own the agent and run it where you run — your cloud, your runtime. It stays connected to the graph via MCP wherever it lives: portable across runtimes, grounded by the graph.
Anthropic AWS Azure Databricks GCP LangChain

All product names and trademarks are the property of their respective owners; their listing reflects runtime compatibility and does not imply affiliation or endorsement.

03
DEVELOPER SURFACE

One call. The full operational context for an agent.

The graph and its agents are accessible via REST, MCP, or direct embed. Same shape, every vertical — switch the tab to see the call land in healthcare, financial, energy, or retail. No fine-tuning. No vector store. No prompt engineering against PDFs.

Designed for the agent loop.

The agent calls the graph the way it would call any other tool. The graph returns structured, token-budgeted operational context — typed nodes, typed edges, citations, sources, peer benchmarks. The agent reasons on the structure, not on the corpus.

  • One signature, four verticals. Switch the vertical: key — the response shape doesn't change.
  • MCP runtime, native. Claude, Cursor, and any MCP-compatible agent framework — wired in without glue code.
  • REST API for legacy stacks. Direct integration into proprietary systems where MCP isn't an option.
  • Token-budgeted responses. The graph trims to your context ceiling — never overrun a window.
  • Provenance on every response. Citation, version, and source path travel with the data — your audit trail is built-in.
claude.ai · amandil-mcp · session 0x4F2A
// agent: prior-auth resubmission await amandil.query({ vertical: "healthcare", subgraph: "prior-auth", scope: { payer: "BCBS-CA", cpt: "72148", denial: "medical-necessity", }, return: ["policy", "prior-decisions", "benefits", "peer-recovery"], }) // returns typed nodes + edges + citations // agent drafts resubmission grounded in graph // audit trail emitted alongside the response
// agent: KYC packet rebuild await amandil.query({ vertical: "financial", subgraph: "kyc-review", scope: { account: "US-CORP-04821", jurisdiction: "US-NY", risk_flag: "ofac-near-match", }, return: ["policy", "prior-decisions", "watchlist", "peer-resolution"], }) // returns typed nodes + edges + citations // agent drafts CDD packet grounded in graph // audit trail emitted alongside the response
// agent: outage incident pack await amandil.query({ vertical: "energy", subgraph: "outage-response", scope: { iso: "PJM", asset: "GEN-04211", severity: "category-3", }, return: ["policy", "prior-incidents", "tariff", "peer-restoration"], }) // returns typed nodes + edges + citations // agent drafts NERC EOP report grounded in graph // audit trail emitted alongside the response
// agent: markdown decision await amandil.query({ vertical: "retail", subgraph: "markdown-decision", scope: { sku: "SKU-9241-RD", channel: "store", category: "apparel", }, return: ["policy", "prior-markdowns", "elasticity", "peer-velocity"], }) // returns typed nodes + edges + citations // agent drafts markdown plan grounded in graph // audit trail emitted alongside the response
04
ONE PATTERN · FOUR INDUSTRIES

Same substrate. Instantiated per industry.

Public-data spine, reference standards, regulatory frameworks, canonical operational example — the dimensions are identical. The substance is per-vertical. Healthcare is live; financial, energy, and retail follow.

Public-data spine
CMS Medicare PUF
NCQA · Joint Commission
FDIC Call Reports
FFIEC · SEC · NAIC
FERC eLibrary
NERC · EIA
Census · BLS
NRF · GS1
Reference standards
X12 · FHIR · HL7v2
CPT · HCPCS · ICD-10
Basel III · IFRS
ISO 20022 · FIX · FpML
NERC CIP · IEEE 1547
NEMA · IEC 61850
GS1 · EDI 850/856
ISO 8000 · UNSPSC
Regulatory frameworks
HIPAA · CMS CoP
NCQA · Joint Commission
Dodd-Frank · BSA/AML
SOX · GLBA · Reg CC
FERC · NERC · NEPA
PHMSA · OSHA
CPSC · FDA · FTC
state UDAP
Canonical example
Prior authorization
KYC / AML review
Outage response
Markdown decision
Status
22 domains · 1,926 processes
9,694 edges · stitched to CMS
Design-partner cohort
Available
Design-partner cohort
Available
Design-partner cohort
Available
05
WHO IT'S FOR

Three customer segments. Every vertical.

Three customer segments — the same across every vertical. The org that runs the operations. The firms that consult into them. The vendors that embed Amandil in their own products.

A · OPERATORS
Payers · Providers · Banks · Insurers · Utilities · Retailers

“Where should we invest in AI?” You run the operations. Most teams pick what to automate based on a vendor pitch or what feels problematic to an executive. You need a benchmark-anchored answer — and agents that cite the regulation, name the process, and return the audit trail.

You get: diagnose against the benchmark library → prioritize by automation score reweighted by your gaps → simulate the top candidates → ship agents grounded in the graph via MCP.

B · ADVISORS
Consulting partners · Healthcare, financial, energy, retail advisory

“How do we scale our AI engagements?” Every project opens with a discovery phase. Weeks of work to rebuild what's broadly the same in every operation. The sameness goes unclaimed — and the artifact ages on contact with reality.

You get: the discovery phase compressed to a 30-minute branded analysis. White-label per client. Branded blueprints, business cases, diagnostics. Multi-client workspaces with scoped grants. The Agent Pipeline as the consulting workflow re-imagined.

C · BUILD PARTNERS
Health IT · Fintech · Energy software · Retail tech vendors

“What grounds our AI features?” You're shipping AI inside your product. Every LLM-integrated feature hits the same wall — the model has no grounded knowledge of the operation. Building the graph yourself is a multi-year, multi-million-dollar undertaking.

You get: a native MCP server (13 tools, 3 read-only resources). OAuth 2.0 plus headless bearer tokens, instantly revocable. Grant-scoped per partner. Versioned, validated graph surface. Multi-year build, solved on day one.

06
WHAT'S NEXT

The intelligence is here.
The operational context is what's missing.

Healthcare is live. Financial, energy, and retail are in design-partner cohort. Tell us your industry and your operational question — we'll send back a roadmap.

If you're in healthcare: we'll send back a roadmap with your public CMS footprint, your peer cohort, and the agent surfaces we'd deploy first.

If you're in finance, energy, or retail: design-partner cohorts are open. You get the first live graph in your industry; we get the operational depth to make it real.

If you're an applied-AI team: we'll get you a sandbox MCP key and the integration patterns within a week.

Press, academia, builders, the curious: just say hi. hello@amandil.ai