Agentic Readiness Index

The 4 Levers CIOs Must Control Before Agents Reach Production

Free self-scoring diagnostic · 20 minutes · 2026 benchmarks

AI readiness measures whether you can adopt AI. Agentic readiness measures whether you can operate systems that act on their own. Four operational levers decide the answer, and they are not the ones most AI governance frameworks emphasize.

Run the Diagnostic Or the Full 30-Day Audit

Editorial illustration: staircase ascending from shadow toward an amber horizon representing agentic maturity

WHY A SEPARATE FRAMEWORK

Agent incidents look nothing like AI incidents

By the end of 2025, every major enterprise AI incident we reviewed fell into one of four patterns. A policy that was too coarse to stop the agent from taking a technically-allowed action that no human would have approved. A toolchain that fractured because two agents raced the same endpoint or a vendor shipped a breaking schema change. A handoff that surfaced on a dashboard nobody was watching. A cost trigger that fired after the run ended, not during. None of these were governance failures in the traditional sense. The governance documents were in place. The maturity model scored the organization at Level 3. The agents failed anyway.

The four levers below are what differentiate organizations that can run agents in production from organizations that have copilots and call them agents. A score under 60 on any single lever means agentic deployments belong in supervised mode until the lever is remediated. A score over 80 on all four is the bar for autonomous production agents, and as of Q2 2026, very few enterprises clear it.

THE FRAMEWORK

Four levers, scored independently

Each lever is scored 0–100 from the operational signals below. Levers are independent; a strong policy with a weak toolchain still yields weak agentic readiness. Remediate the weakest lever first, it defines the practical ceiling of what you can ship.

Can your policy distinguish between actions the agent should take and actions the agent could technically take?

Most AI policies are written for humans using AI tools. They say "do not share customer data with external systems" and "review outputs before publishing." Agents do not review their own outputs, and the question of what counts as "sharing" collapses when the agent is reading a CRM, writing a draft email, calling a search API, and filing a ticket in the same run. Policy granularity is whether your rules are specific enough to give the agent something to act on.

Strong signals (scores 80+)

Policy is written at the action level: specific tools, specific operations (read, write, delete), specific data classes
Every production agent has a documented allow-list of tools and an explicit deny-list of destructive or externally-visible operations
Policy distinguishes supervised (copilot) from autonomous (agent) modes and applies different rules to each
Exceptions require a documented request, a risk rationale, and a named approver, not a silent flag flip

Weak signals (scores below 40)

AI policy is one document that covers ChatGPT use, copilot tooling, and production agents interchangeably
Agents inherit service account permissions rather than agent-specific permissions
Policy exceptions are managed in Slack DMs or spreadsheets
The answer to "what can this agent do?" is "whatever the API allows"

Can your tools survive concurrent agent access, protocol drift, and vendor churn?

MCP became the dominant agent-tool protocol in 2025. Nearly every major vendor implemented it. That standardization hides a harder problem: tools break in ways that copilot-era infrastructure never had to handle. Two agents hitting the same endpoint race each other. A tool schema changes mid-run because a vendor shipped a breaking update. An agent written for Anthropic's MCP fragments when pointed at a slightly different implementation. Toolchain interoperability measures whether your tool infrastructure is load-bearing or incidental.

Strong signals (scores 80+)

Tool definitions are versioned, and agents declare the version they were tested against
Rate limiting is scoped to the agent identity, not the service account
Tools expose idempotency semantics and agents know how to use them
Breaking changes to tool schemas go through the same deprecation cycle as external API changes, 90-day notice, compatibility shim, telemetry on old-version usage
You can swap the underlying model (Claude to GPT to Gemini) without rewriting tool glue

Weak signals (scores below 40)

Tools are added to agents by "whoever needed it that week" with no registry
Two agents share a single service account; audit logs cannot tell them apart
Tool definitions live in prompt strings rather than versioned schemas
A vendor update broke an agent and the first sign was a customer complaint

When the agent escalates, does a human actually catch it, in time?

Every production agent will, eventually, hit a decision it should not make. Handoff protocols determine what happens next. Weak handoffs surface on dashboards nobody watches, page oncall engineers who do not have context, or time out silently and let the agent proceed. Strong handoffs route to the right person with the full trace, block the action until resolution, and include a rehearsed fallback for when the human is unreachable. This is the lever where post-mortems most often reveal that the organization thought it had a handoff and did not.

Strong signals (scores 80+)

Every agent has documented escalation triggers (confidence below threshold, ambiguous tool call, novel action not in training distribution)
Escalations route to a named on-duty human with full context (recent actions, the decision at hand, and a one-click deny)
Handoff SLAs are defined and tracked: median time-to-human, median time-to-decision, rate of timeouts
A monthly drill tests the handoff path end-to-end, including the case where the primary approver is unreachable
Agents pause on escalation; they do not proceed with a fallback after a timeout

Weak signals (scores below 40)

Escalations go to a Slack channel with 200 members and no owner
Timeouts default to "agent proceeds with best guess"
The on-call rotation for agentic systems is the same as the general platform rotation
Nobody has tested what happens when the primary approver is on PTO

Will you know the agent is burning budget before the budget is gone?

Agent token burn is bimodal. Most runs are cheap. A small fraction (the ones that hit a recursion, a context-window spiral, or an unbounded search) consume more tokens in minutes than a normal run consumes in a month. Cost escalation triggers determine whether you detect these runs while they are running, not after. Weak triggers fire on monthly invoice review. Strong triggers fire on per-run budgets, per-agent budgets, and cross-agent spend velocity, with automated kill-switches before the alert is even read.

Strong signals (scores 80+)

Every agent has a per-run token budget and a per-hour spend ceiling; both are enforced in code, not policy
A run approaching its budget triggers a soft stop (agent summarizes state and hands off) rather than a hard kill
Spend-velocity alerts fire within 5 minutes of a step-function increase
Budget breaches have documented owners and a post-incident review cadence
Finance and engineering share a real-time agentic cost dashboard

Weak signals (scores below 40)

Agent cost is reviewed monthly from the cloud provider invoice
A single runaway run could exceed the entire monthly budget before anyone noticed
Kill-switches exist on paper but have never been exercised
Nobody can answer "what did we spend on agents yesterday?" within 30 seconds

THE DIAGNOSTIC

12 questions, one score per lever

Answer each question for your most-autonomous production AI system. Yes = 33 points for that lever. Partial = 17. No = 0. If you have no production agents, score the system you are closest to deploying. A lever with any "no" is capped at 66 regardless of the other answers, a single blocker defeats the lever.

Policy Granularity

Can you produce, in under five minutes, the exact list of tools and operations each production agent is allowed to invoke?
Does your policy distinguish rules for copilots (human-committed) from rules for agents (agent-committed)?
Is there a documented process for adding, modifying, or removing agent permissions, with named approvers and audit trail?

Toolchain Interoperability

Are your tool definitions versioned, with agents pinned to tested versions?
Can your audit logs distinguish which agent (not which service account) performed a given action?
Have you successfully swapped the underlying model of a production agent in the last 12 months without rewriting tool glue?

Human-Agent Handoff

When an agent escalates, does it route to a named on-duty human with full context, or to a shared channel?
Have you rehearsed, in the last 90 days, what happens when the primary approver is unreachable?
Do you track time-to-human-decision as a first-class SLA, with targets and alerting?

Cost Escalation Triggers

Is there a per-run token budget enforced in code for every production agent?
Can you answer "what did we spend on agents in the last hour?" from a live dashboard?
Has your kill-switch been exercised in a drill (not just an incident) in the last quarter?

Reading your score

80–100 on all four levers: Production-ready for autonomous agents in bounded domains. Expand cautiously; monitor the weakest lever as you scale.
60–79 on all four levers: Supervised autonomy only. Run agents in production with a human in the approval loop for every action in the weakest-lever domain.
Below 60 on any single lever: Do not run autonomous agents in the domain that lever governs. Copilots are fine; autonomy is not.
Below 40 on any single lever: Halt agentic rollout in that domain and remediate. The remediation is usually measured in quarters, not weeks.

Q2 2026 BENCHMARKS

Where your peers actually score

Aggregate scores from organizations we have assessed, reviewed, or benchmarked against public disclosures. The gap between "enterprise with mature AI program" and "frontier AI labs" is not vision or talent, it is operational infrastructure measured here.

Segment	Policy	Toolchain	Handoff	Cost
Frontier AI labs / tier-1 tech (2026)	85	80	75	85
Enterprise with mature AI program	60	55	45	50
Enterprise with copilot deployments	40	35	25	30
Enterprise with ChatGPT-era policy only	15	10	10	10

Handoff is consistently the weakest lever across every segment except frontier labs. It is also the lever most organizations overestimate, the gap between "we have an escalation path" and "the escalation path has been exercised" is where most 2025-2026 agent incidents lived.

ADJACENT FRAMEWORKS

Where this sits next to governance maturity and the readiness audit

The Agentic Readiness Index complements, not replaces, the broader frameworks. Use it to answer a specific question: can this organization run agents in production, right now, without creating incidents the governance model cannot catch?

	Agentic Readiness Index	Governance Maturity Model	30-Day AI Readiness Audit
What it measures	Operational capacity to run autonomous agents in production	Institutional governance scaffolding for AI generally	Six-dimension org-wide readiness for AI adoption
Primary audience	CTO, Head of Platform, CAIO	CAIO, CRO, General Counsel	CEO, board, executive team
Output	0-100 score per lever + specific remediation moves	Level 1-5 positioning + transition playbook	Board-ready report + 6-12 month roadmap
Cost	Free self-assessment	Free self-assessment	$25,000-$50,000 paid engagement
Time to complete	20 minutes	15 minutes	30 days
Depth	Deep on four operational levers	Broad on governance controls	Deep across six organizational dimensions

Two related pieces to read alongside: Agentic AI ROI covers the economic case once agents are running; Agentic AI Security covers the adversarial dimension. For the architecture patterns themselves, the authoritative reference remains Agentic AI Architecture: Patterns, Diagrams, and the Orchestration Decision.

REMEDIATION ORDER

Fix the weakest lever first, always

Multi-lever remediation programs consistently underperform single-lever remediation programs followed by the next-weakest lever. The reason is brittle by nature of the work: policy, toolchain, handoff, and cost work as a system, and attempting three at once produces three half-finished projects. Sequence as follows.

Identify the weakest lever. If two levers score within 10 points of each other, pick the one your agents exercise most often in their current workload.
Set a ceiling, not a floor. Cap agent autonomy in the domain that lever governs until the lever crosses 70. This is non-negotiable and should be visible to every team shipping agents.
Run a 90-day remediation sprint. Policy granularity: rewrite the agent section of the AI policy with named tools and named operations. Toolchain: version every tool definition, instrument agent-identity audit logs, add schema deprecation cycles. Handoff: name on-duty humans, set SLAs, run a monthly drill. Cost: implement per-run and per-hour budgets in code, build the live dashboard.
Re-score and re-plan. Re-run the diagnostic at day 90. The lever should cross 70. If it does not, the plan was wrong; extend by 60 days before moving to the next lever.
Move to the next weakest lever. Repeat. Full four-lever remediation typically takes 9–15 months in a mid-sized enterprise, longer in regulated industries.

Frequently Asked Questions

What is agentic readiness?

Agentic readiness is an organization's capacity to deploy and operate autonomous AI agents that take multi-step actions on their own, not just RAG-augmented chatbots. Where traditional AI readiness asks whether you can adopt AI, agentic readiness asks a harder question: can you operate systems that decide, act, spend tokens, call tools, and occasionally fail in unexpected ways without human review of every step? Four operational levers determine the answer: policy granularity, toolchain interoperability, human-agent handoff protocols, and cost escalation triggers.

How is this different from the AI Governance Maturity Model?

The governance maturity model measures the institutional scaffolding around AI: policies, risk registers, compliance mapping, board reporting. The agentic readiness index measures the operational infrastructure required for a specific class of AI system: one that acts autonomously. An organization can reach Level 3 governance maturity and still be agentic-unready because its tool-call logs are sampled, its cost triggers fire only after the fact, and no one has tested what happens when an agent loops.

How is this different from the 30-day AI Readiness Audit?

The AI Readiness Audit is a paid 30-day engagement scoring six organizational dimensions (delivery, workforce, architecture, data, governance, leadership) against Gartner benchmarks. The Agentic Readiness Index is a free self-scoring diagnostic focused specifically on the four operational levers required to run agents in production. Most organizations who complete the audit score well on general AI readiness and poorly on agentic readiness. The capabilities are adjacent, not overlapping. Teams typically start with this index and commission the audit when they need an enterprise-wide roadmap.

Why only four levers instead of a bigger framework?

Every agentic failure we have observed across 2024-2026 fell into one of four categories: a policy that was too coarse (the agent did something technically allowed that no one would have approved), a toolchain that fractured under load (two agents fighting over the same tool, or a tool changing shape mid-call), a handoff that failed silently (the agent escalated to a human who was not watching), or a cost trigger that fired too late (the run was over before the budget alert arrived). Data quality, model selection, and prompt design matter too, but none of them separate agentic readiness from AI readiness generally. Four levers is tight enough to remember and specific enough to act on.

What score indicates we are ready to scale agents in production?

A score of 80+ across all four levers, with no individual lever below 70. At that threshold an organization has granular enough policy to prevent over-action, tool infrastructure that survives agent concurrency, handoff protocols that catch failures before they escalate, and cost triggers that fire before budgets blow. Below 60 on any lever, agentic deployments should stay in supervised pilot mode. Below 40 on any lever, do not run autonomous agents in production at all. Run copilots with every step human-approved until the lever is remediated.

We already have copilots in production. Do we need this?

Copilots and agents are different risk categories. A copilot suggests; a human commits. An agent commits; a human audits. The jump from copilot-in-production to agent-in-production is where most 2025-2026 incidents happened: the same infrastructure that was safe for suggestions became unsafe when the same system started acting. The index is most useful precisely at this transition, when leadership believes the organization is agent-ready because copilots work, but the operational infrastructure has not caught up.

Self-scored, and not sure what the score means?

The 30-day AI Readiness Audit takes the same four levers and seven other organizational dimensions, validates them with stakeholder interviews and architecture review, and produces a board-ready roadmap. Most teams run the index first; the audit comes in when the score surfaces a gap too big to close internally.

See the 30-Day Audit Or a Fractional Engagement