The Agent Readiness Checklist:The 26-Point Pre-Deployment Matrix (ARM-26)

Stop Deploying AI to Undefined Workflows

The single most expensive mistake enterprise technology leaders make is assuming that artificial intelligence can compensate for a broken, undocumented, or unmeasured business process.

When a team says, “Our ticket triage is chaotic and our people are overwhelmed, let’s put an AI agent on it,” what they usually get is accelerated chaos. The model encounters ambiguous priorities, conflicting instructions, and missing fields. It makes erratic decisions, frustrates the human team, and is quietly disabled within eight weeks.

Before writing code, procuring software licenses, or drafting an engineering roadmap, you must determine whether a workflow has the structural maturity to support autonomous software.

At Kodense, our Forward Deployed Engineers use the Agent Readiness Matrix (ARM-26): a rigorous, 26-point operational diagnostic that evaluates target workflows across four non-negotiable dimensions.

Scoring Methodology

Score each of the 13 criteria below using a strict ternary scale:

• 0 Points: No / Not currently true / Cannot be verified.
• 1 Point: Partially true / True in some instances / Requires minor preparation.
• 2 Points: Fully verified / Documented in writing / Implemented in production systems.

Maximum achievable score: 26 Points.

Category 1: Value & Economic Impact (6 Points Max)

An autonomous agent must move an audited business metric. If the financial or operational return cannot be stated in numbers, the project will be defunded during the next budget review.

1.1 Volume Threshold

“The target workflow consumes at least 20 human hours per week, or directly affects revenue, operational cost, or critical compliance risk.”

• 2 Points: The workflow consumes 20+ hours weekly across the team, or directly throttles revenue velocity.
• 1 Point: Volume fluctuates seasonally or totals 10 to 19 hours weekly.
• 0 Points: The task occurs intermittently (e.g., once a month) and takes fewer than 5 hours.

1.2 The Moving Metric

“We can state the exact metric an agent would move (resolution time, throughput volume, error rate, cost per transaction, or reclaimed capacity).”

• 2 Points: The target metric is clearly identified and tracked in an operational dashboard (e.g., Jira, Zendesk, Salesforce).
• 1 Point: The team has a conceptual idea of what will improve, but no formal metric is logged today.
• 0 Points: Success is defined purely by user sentiment or vague “productivity gains.”

1.3 Pre-Deployment Baseline

“We know the current baseline for that metric, or could establish a verified 14-day baseline measurement before deployment.”

• 2 Points: We have historical baseline data for the past 90 days.
• 1 Point: We do not have historical data, but we can measure and log the baseline over the next two weeks.
• 0 Points: The process is untracked, unmeasured, and baseline data cannot be extracted.

Category 2: Workflow Determinism & Scope (6 Points Max)

Frontier models excel at natural language parsing and multi-step reasoning, but they require bounded state transitions to execute actions reliably.

2.1 Process Describability

“The workflow follows a describable, step-by-step process, even if it contains numerous branch conditions and edge-case exceptions.”

• 2 Points: A standard operating procedure (SOP) or flowchart exists and accurately reflects how the team actually works.
• 1 Point: The process lives primarily in tribal knowledge, but team leads can describe the rules consistently when interviewed.
• 0 Points: Every case is treated ad-hoc with zero consistency between individual operators.

2.2 Judgement vs. Execution Ratio

“Fewer than 50% of incoming cases require subjective creative or strategic human judgement; the remainder consist of rules, lookups, and deterministic actions.”

• 2 Points: 70%+ of incoming volume consists of standard lookups, cross-referencing records, and predictable state updates.
• 1 Point: Roughly half the volume requires nuanced interpretation of customer intent or policy exceptions.
• 0 Points: Nearly every transaction requires senior executive discretion, political negotiation, or creative writing.

2.3 Observable “Done” State

“The definition of ‘Done’ is clearly defined, unambiguous, and visibly recorded in a software system.”

• 2 Points: A completed transaction is marked by a specific state change in a database, CRM, or ticketing system (e.g., ticket marked Resolved, invoice marked Approved).
• 1 Point: The task is finished when an email is sent, but no central ledger or database records the completion.
• 0 Points: “Done” is subjective and depends on informal verbal agreement between colleagues.

Category 3: Data & Systems Accessibility (6 Points Max)

An AI agent cannot execute without programmatic interfaces. If your internal systems cannot be accessed via modern APIs or structured data exports, the agent cannot function.

3.1 Interface Programmability

“The systems involved have documented REST/GraphQL APIs, database connectors, structured exports, or at least standardized email/document schemas.”

• 2 Points: Modern APIs or database read/write access exist for all primary software tools involved.
• 1 Point: Some tools have APIs, but one legacy tool requires flat CSV exports or automated web scraping.
• 0 Points: The workflow relies entirely on proprietary desktop software with zero API, zero export capabilities, and strict anti-automation controls.

3.2 Historical Ground Truth

“Historical examples of completed work exist (closed tickets, processed orders, approved PRs, paid invoices) from which we can construct evaluation suites.”

• 2 Points: We have at least 500 historical completed transactions with associated input data and final resolutions.
• 1 Point: We have 50 to 200 historical examples, but some context was handled over phone calls or unlogged chats.
• 0 Points: Zero historical data exists; this is a brand-new workflow with no prior track record.

3.3 Access Grant Velocity

“An authorized system administrator can grant scoped API credentials or private VPC staging access within two weeks.”

• 2 Points: System owners have the authority to issue service accounts and API keys within 5 business days.
• 1 Point: Access requests require formal approval from an IT security board, taking 10 to 14 business days.
• 0 Points: Access requires external compliance audits, multi-month steering committees, or vendor permissions that take 90+ days.

Category 4: People, Governance & Safety (8 Points Max)

The technical architecture of an agent is only half the battle. If internal stakeholders feel threatened, or if safety rails are not designed upfront, the deployment will fail on human grounds.

4.1 Accountable Workflow Owner

“A designated workflow owner exists who has operational responsibility for this queue, wants it automated, and has the authority to make weekly decisions.”

• 2 Points: A dedicated engineering manager, director of operations, or team lead is actively championing the deployment.
• 1 Point: A manager supports the project conceptually, but lacks the executive authority to approve workflow changes.
• 0 Points: No single individual owns the workflow; responsibility is diffuse across a committee.

4.2 Frontline Operator Buy-In

“The frontline staff currently doing the manual work understand the project, view the agent as leverage rather than a threat, and are willing to train and supervise it.”

• 2 Points: The frontline team is actively burned out by manual toil and enthusiastically volunteers to supervise the agent’s initial runs.
• 1 Point: The team is cautious but willing to cooperate if they see that it eliminates tedious repetitive tasks.
• 0 Points: The team is actively hostile, fears immediate job cuts, and has not been briefed on the implementation plan.

4.3 Designated Security & Compliance Contact

“Security, legal, and compliance have an assigned point of contact who can review data flow diagrams and sign off on security requirements.”

• 2 Points: A security engineer or compliance officer is already identified and has reviewed the private VPC deployment architecture.
• 1 Point: A security team exists, but they have not yet been assigned to review the agent’s threat model.
• 0 Points: Any external or autonomous tool requires an adversarial 6-month corporate review with zero designated sponsor.

4.4 Explicit Non-Negotiable Approval Gates

“The organization can explicitly name the specific actions that must NEVER be executed autonomously without human sign-off (Money, Customers, Production, Legal).”

• 2 Points: The team has an explicit written list of actions requiring mandatory human sign-off (e.g., payments > $500, outbound emails to VIP clients, database deletions).
• 1 Point: The team agrees that safety gates are needed, but has not yet defined the exact numerical thresholds.
• 0 Points: The team assumes the AI should be 100% autonomous from day one without any human supervision.

Score Interpretation & Action Plan

20 to 26 Points: Immediate Production Deployment Candidate

Your target workflow possesses the necessary volume, process determinism, technical programmability, and governance buy-in to succeed. You are ready for a Forward Deployed Engineering Sprint. An embedded FDE can enter your environment, establish the evaluation baseline, wire into your systems, and have a containerized agent running live in production within 8 weeks.

12 to 19 Points: Candidate with Targeted Preparation

The workflow has significant potential, but one or two structural deficiencies will stall an immediate deployment. Typically, the issue is missing APIs, an unmeasured baseline, or undefined approval thresholds.

• Action: Run a 2-Week Technical Scoping Sprint. Focus exclusively on logging 14 days of baseline data, securing service account credentials, and establishing the four human approval gates. Once resolved, the workflow transitions to a Tier-1 candidate.

Under 12 Points: Foundation Remediation Required

Do not deploy an autonomous AI agent into this workflow today. Attempting to automate an unmeasured, undocumented, or politically contentious process with an LLM will burn capital and destroy organizational credibility.

• Action: Step back and fix operational foundations. Standardize the human workflow, implement structured tracking in your CRM or ticketing system, and build internal trust before introducing autonomous software.

Take the Next Step: 30-Minute Workflow Mapping

Whether your workflow scored 24 or 14, the fastest way to validate your assessment is to review it with an engineer who builds and deploys autonomous agents into production every day.

In a 30-Minute Kodense Technical Discovery Call, we will:

1. Review your target workflow against the ARM-26 matrix.
2. Identify any hidden integration walls or data schema traps.
3. Provide an honest, unvarnished recommendation: whether to build custom, buy commercial SaaS, or fix the workflow first.

No sales presentations. No slide decks. Just systems engineers talking through code and architecture.

Schedule a 30-Minute Technical Discovery Call →