CE Practice · Nov 2026

Accountable AI
by Design

The practitioner's guide to agent governance — what the industry calls "guardrails", what it actually requires, and where the SI's differentiated play sits.

81%
of BFSI firms run AI agents — yet only 14% have security approval. The fastest-adopting sector is the most under-governed. (AWS GEAR, 2026)
21%
of enterprises have mature governance for deployed agents (AWS GEAR, 2026)
$15.8B
AI governance software market by 2030 · 30% CAGR (Forrester)
Aug 26
EU AI Act Art. 9+17 enforcement — version-controlled test metrics legally required
Taxonomy

AgentOS, Harness, Framework — what's what

The industry uses these terms interchangeably. They are not the same thing. Your governance investment lands in the wrong layer if you confuse them.

AgentOS scope — L3 + L4
L4
Agent Management Platform
Operate
Fleet visibility, cost attribution, and governance across agents built on any framework. Gartner published its inaugural MQ for this category Nov 2026 — three Leaders named. By 2029, Gartner expects 60% of enterprise agents managed through an agent management platform.
Microsoft Agent 365 IBM watsonx Orchestrate ServiceNow AI Control Tower AWS (Challenger)
L3
Agent Harness / Runtime Governance
← the SI configures & extends
Runtime enforcement — RBAC, circuit breakers, audit trails, PII redaction, cost gates, HITL. This is what "we need guardrails" actually means in an architecture review. Built into cloud platforms; the SI configures these, not replaces them.
AWS AgentCore Azure AGT Databricks MLflow 3.0 Snowflake Cortex Gateway Google Vertex AI Agent Builder
L2
Agent Orchestration
Coordinate
Supervisor patterns, handoffs, routing between agents. Coordinates — does not govern.
LangGraph AG2 / AutoGen Google ADK
L1
Agent Framework / Runtime
Build
The perceive → reason → act → observe loop. Builds individual agents — does not govern them.
Pydantic AI smolagents Google ADK
the SI's play — above L3, not inside it
the SI configures and extends your existing cloud harness. The differentiated value is domain utilities and opinionated configurations — each with a measurable before/after output. Details in the next section.
Gartner · MQ 2026
AI Agent Management Platforms
Leaders: Microsoft, IBM, ServiceNow (Gartner MQ Nov 2026)
Forrester · Q3 2026
Agentic Development Platforms
25 vendors; Wave report Q4 2026
Forrester · Blog 2026
Agent Control Plane
"Still lacks a robust standards stack"
Gartner · Market Guide Feb 2026
Guardian Agents
Reviewer · Monitor · Protector >$3B by 2030
SI Differentiated Play

Not the harness — what makes the harness useful

Platform companies build L3 infrastructure. the SI builds the opinionated utilities and configurations that sit on top — specific to enterprise domains, measurable before/after. The same model Tredence used for semantic-layer token efficiency: don't build the database, build the utility that tests the database's quality for AI use.

Semantic Layer AI Readiness Score
Build
Audits LookML, dbt metrics, or Tableau extensions for AI-readiness. Runs test agent queries before and after, measures token reduction % and answer accuracy delta. Produces a ranked fix list.
Anchors in BI practice — the SI already implements semantic layers for clients.
Golden Dataset Factory
Build
Generates domain eval suites from existing BI reports. Extracts known Q&A pairs → structures as RAGAS-compatible eval cases → wires to CI. Compresses 6-week eval construction to days.
High reuse across verticals — every client has BI reports; almost none have structured eval suites.
Context Compression Benchmarker
Build
Tests chunking strategies, embedding models, and retrieval parameters against a client's document corpus. Measures token efficiency vs. answer faithfulness. Outputs an optimal "context recipe."
40–70% token reduction routinely available. Most clients never measure it.
Agent Output Consistency Checker
Build
Tests whether an agent returns semantically equivalent answers to rephrased questions. Flags divergence beyond a threshold. Produces a "consistency score" — a key audit metric for financial agents under EU AI Act Art. 9.
BFSI regulators require version-controlled test metrics. This is the simplest metric that maps to that requirement.
Data-to-Agent Contract Validator
Build
YAML metadata spec overlaid on Unity Catalog or Snowflake Horizon — annotates data assets with AI-consumption rules: PII sensitivity, allowed agents, business rules. Validates agent data access before promotion.
Extends existing data platform engagement into AI governance with no custom harness required.
OWASP AIT10 Reference Test Kit
Champion
PromptFoo test suite covering all 10 Agentic AI failure modes (ASI01–ASI10): prompt injection, tool poisoning, excessive agency, and more. Sector variants for BFSI and CX. Standard pre-go-live gate.
Referenced in BFSI RFPs by name since Dec 2025.
JPMorgan Chase FSB submission — gateway pattern reference architecture Barclays BCBS 239 Wave 2 — LeapLogic + LLM lineage harness NSA advisory — MCP has no protocol-level enforcement; gateway required
Structural Gap

Training-time safety ≠ runtime governance

Constitutional AI and Model Spec are behavioral priors baked into model weights. They reduce the probability of harmful outputs. They provide no RBAC, no audit trail, no circuit breakers. This is a structural gap — not a product lag.

Training-time safety
Constitutional AI · Model Spec · RLHF
Governs model behavior at inference. The model "tries" to refuse harmful requests. Can be overridden by a sufficiently adversarial system prompt.
✓ Reduces probability of harmful outputs
✓ Refusal classifiers in model weights
✗ No RBAC — anyone with an API key gets the same model
✗ No audit trail — model doesn't log tool calls
✗ No circuit breakers — model doesn't halt after N iterations
✗ No delegation-chain enforcement
Runtime governance
What enterprises must build or configure
Enforcement happens outside model weights — at the gateway, in the harness, in the orchestration layer. This is what regulated enterprises require.
✗ RBAC (who can call what agent / tool)
✗ Audit trail of every tool call and hop
✗ Rate / cost circuit breakers per agent
✗ PII redaction layer (GDPR, HIPAA)
✗ Prompt injection defense (OWASP ASI01)
✗ Human approval gates (EU AI Act Art. 14)

What you must configure on model providers

Control required Anthropic APIOpenAI APIRequired by
Per-agent identity❌❌OWASP ASI01 · EU AI Act Art. 9
Audit trail of tool calls❌⚡ 30-day limitNIST AI RMF · ISO 42001 Cl. 8
RBAC per agent / tool⚡ MCP auth (partial)❌AIUC-1 · internal compliance
Rate / cost circuit breakers❌❌FinOps · risk management
PII redaction❌❌GDPR · CCPA · HIPAA
Prompt injection defense⚠ Protocol warning only❌OWASP ASI01
Human approval gates⚡ SDK interrupt only⚡ SDK interrupt onlyEU AI Act Art. 14 · HITL

Business model tension: model providers monetize on token consumption; governance controls that limit consumption are economically adverse. Infrastructure providers (AWS, Azure, Databricks) monetize on platform lock-in — governance drives stickiness. That's why the enforcement layer belongs to infrastructure, not model providers.

Platform Comparison · 7 platforms · 10 governance dimensions

Who is building the Agentic Control Plane — and how fast

Assessed across governance specification depth, enforcement completeness, identity, observability, and strategic direction. Sources: Gartner MQ 2026, Forrester Q3 2026, platform documentation, Nov 2026.

Dimension Microsoft
AGT + Fabric + Agent 365		 AWS
Bedrock + AgentCore		 Databricks
Mosaic AI + Agent Bricks		 Snowflake
Cortex AI + Horizon		 Google
GEAP + A2A		 Anthropic
Claude + MCP		 OpenAI
Agents SDK

✅ Full / native  ⚡ Partial / emerging  ❌ None / must build  🔥 Best-in-class · Scroll horizontally on narrow screens.

Design Patterns · 18 spotlight patterns across 6 phases

Governance applied per lifecycle phase — not bolted on at the end

Each phase has distinct governance obligations. Spotlight patterns are the ones that matter most in enterprise architecture reviews and regulatory audits.

Regulatory Compass

Four frameworks shaping enterprise agent governance

These are the frameworks cited in procurement RFPs and architecture reviews. Know the obligations before the auditor does.

European Union
EU AI Act Art. 9 + 17
Risk management system + quality management. Version-controlled test metrics legally required for high-risk AI. Credit scoring, insurance risk pricing = Annex III.
Enforcement Aug 2026
NIST
AI Risk Management Framework 1.0
Govern / Map / Measure / Manage functions. The US enterprise governance baseline. Maps directly to harness component requirements.
GA 2023 · Agentic update 2025
ISO / IEC
ISO 42001:2023
AI management system standard. Clause 8 requires documented AI system lifecycle processes including testing and monitoring. Audit-ready evidence trail.
Published Dec 2023
OWASP
Agentic AI Top 10
ASI01–ASI10. Published Dec 2025. Fastest-adopted failure-mode taxonomy in the industry. CDOs reference it by name in BFSI RFPs. The red-team test matrix.
Published Dec 2025
AIUC-1 Certification (UiPath first certified) Singapore MAIGF v1.5 · May 2026 ABS GenAI Guardrails Handbook NIST AI 600-1 (GenAI supplement)
SI Positioning

"Accountable AI by Design" — an unclaimed position

Every major consultancy has staked a governance phrase. None own the utility-first, platform-native, eval-driven framing that enterprise architects actually need.

FirmGovernance PhrasePrimary LensWhat's Missing
IBMTrustworthy AIEthics + model biasRuntime enforcement depth
AccentureResponsible AIPrinciples + policyPlatform-native delivery
DeloitteAI AssuranceAudit + complianceAgentic-specific patterns
KPMGAI Trust (May 2025)Risk managementMeasurable utility output
the SIAccountable AI by DesignUtility-first · Platform-native · Eval-driven—
①
Utility-first, not infrastructure-first
the SI builds measurable accelerators on top of your existing harness — not a competing harness. Semantic Layer AI Readiness Score, Golden Dataset Factory, Context Compression Benchmarker. Each has a specific before/after output, like the Tredence semantic-layer token utility model.
②
Platform-native configuration
the SI configures AWS AgentCore, Azure AGT, Databricks MLflow, and Snowflake Cortex — not a proprietary harness. One less vendor to trust, one less thing to certify. Clients already paying for these platforms get governance from their own stack.
③
Eval-driven from day one
Write evals before prompts. Golden datasets built from the client's own BI reports. The 89%/52% observability-vs-eval gap is the direct entry point: most clients have monitoring, almost none have structured evals wired to CI.
Next Step

Ready to govern your agents before the incident?

the SI helps enterprise teams close the governance gap — from platform configuration and harness selection to eval pipelines and regulatory readiness. Platform-native, utility-first.