A8gent

Free tool

AI Agent Risk Scorer

Score an AI agent workflow for customer, financial, legal, data, brand, and operational risk before launch, and see which steps still need a human.

Contents
  1. Fast answer
  2. Quick facts
  3. What this covers
  4. Why it matters
  5. Where to go next
  6. Outcomes
  7. Step by step
  8. The methodology: exactly how the risk score is calculated
  9. What the score can and cannot tell you
  10. A worked example: scoring two clinic workflows
  11. Mapping your risk score to the NIST AI Risk Management Framework
  12. What OWASP says can actually go wrong with agents
  13. The regulatory clock and the incident curve
  14. What changed in 2025 and 2026
  15. The limits: five questions is not a risk assessment
  16. When not to use this tool, and what to use instead
  17. At a glance
  18. Mistakes to avoid
  19. FAQ
TL;DRAnswer-first

Fast answer

An AI agent risk scorer helps decide whether a workflow can run automatically, needs human approval, should only draft recommendations, or should not be automated yet. Agent risk is not only model accuracy; it includes bad data, unclear accountability, sensitive information, customer impact, financial exposure, and whether mistakes are easy to detect. Start by identifying sensitive actions, flagging anything that sends customer messages, changes CRM records, updates billing, makes promises, or touches private data. Rate impact and reversibility, since a mistake that can be reviewed and undone is lower risk than one that reaches a customer, regulator, or payment system. Set approval levels using draft-only, human approval, sampled review, or automatic action depending on risk and confidence. Define failure handling for missing data, conflicting sources, and low confidence, then monitor edits, overrides, complaints, and failed tool calls after launch. An agent can act without approval only when the action is low-risk, reversible, well-tested, logged, and monitored, because review reduces risk but never removes it.

Facts last verified 2026-07

Quick facts

NIST AI RMF functions
Govern, Map, Measure, Manage (AI RMF 1.0, Jan 2023)
Top agent-specific OWASP risk
LLM06:2025 Excessive Agency; LLM01 is Prompt Injection
EU AI Act fully applicable
August 2, 2026, including high-risk obligations and penalties
Reported AI incidents
362 in 2025, up from 233 in 2024 (Stanford AI Index)
Projects at risk
Over 40% of agentic AI projects predicted canceled by end of 2027 (Gartner)
Gen AI risk areas named by NIST
12, including confabulation and human-AI configuration (NIST-AI-600-1)

On this page

What this page covers

A tool visitor should leave with a decision, not just a number: build now, prepare first, choose another workflow, or follow a course path.

  1. 01Risk inputs
  2. 02Approval matrix
  3. 03Data exposure
  4. 04Failure modes
  5. 05Monitoring plan
  6. 06FAQ

Why does this matter now?

Agent risk is not only model accuracy. It includes bad data, unclear accountability, sensitive information, customer impact, financial exposure, and whether mistakes are easy to detect.

Internal path

Where to go next from this page

These links are part of the A8gent learning and conversion path. Use them to move from concept, to diagnosis, to workflow build, to course.

Start with readiness

What you should be able to do after this

  • Identify risky actions
  • Choose approval levels
  • Plan monitoring
  • Decide what not to automate

How do you do it, step by step?

1. Identify sensitive actions

Flag anything that sends customer messages, changes CRM records, updates billing, makes promises, or touches private data.

2. Rate impact and reversibility

A mistake that can be reviewed and undone is lower risk than one that reaches a customer, regulator, or payment system.

3. Set approval levels

Use draft-only, human approval, sampled review, or automatic action depending on risk and confidence.

4. Define failure handling

Write rules for missing data, conflicting sources, uncertain answers, low confidence, and escalations.

5. Monitor after launch

Track edits, overrides, customer complaints, failed tool calls, and examples that should update the workflow.

The methodology: exactly how the risk score is calculated

The scorer asks five questions and each answer is a position on a five-point scale, stored as 0 through 4. Autonomy runs from draft only, through suggests actions, acts with approval and acts with human spot checks, to fully autonomous. Data sensitivity runs from public, through internal, confidential and personal data, to regulated or financial. Who sees the output runs from internal only, through internal plus logged, partners and customers after review, to customers directly. Reversibility runs from trivial to undo, through undo with effort, hard to undo and costly to undo, to irreversible. Volume runs from a few actions a day, through tens, hundreds and thousands, to continuous.

The five factors are weighted, and the weights are small: reversibility 1.3, data sensitivity 1.2, autonomy 1.1, who sees the output 1.0, and volume 0.9. Your answer index is multiplied by its weight, the five products are summed, and that total is divided by the maximum possible weighted total, which is 4 times the sum of the weights, or 22. The result is multiplied by 100 and rounded. Answering the middle option on everything gives 50.

Three bands convert that number into a level. Sixty or above is High, 33 to 59 is Medium, and below 33 is Low. Each level carries a fixed control list: three items at Low, five at Medium and seven at High, moving from logging and a kill switch, through mandatory approval before any external or written action and PII redaction, to two-person sign-off on irreversible steps, automated fact and policy checks, tested rollback per action type and a staged rollout with capped volume.

Be clear about what those weights are: they are an editorial judgement, not a measurement. Nobody has established that reversibility is 1.44 times as important as volume, and if we set every weight to 1.0 most workflows would land in the same band. The weights exist to make the two factors that generate real-world incidents pull hardest, and reversibility pulls hardest of all because an irreversible mistake is the only kind you cannot fix on the day you notice it. Read the score as an ordering device across your own workflows rather than an absolute measurement of danger.

What the score can and cannot tell you

It can tell you whether one workflow is riskier than another workflow you scored the same way, and it can tell you which control set to start from. That is genuinely useful, because most teams either apply no controls or apply the same heavy controls everywhere until people route around them. Used as a relative ranking across five or six candidate workflows, the score does honest work.

It cannot tell you that a workflow is safe. There is no threshold in this model below which nothing bad happens, and a Low band means only that the four or five things this tool asks about all pointed the same way. A workflow can score 20 and still cause harm through a failure mode outside the five questions, and it will not be the tool's fault, because it never claimed to look.

It does not model your regulatory exposure at all. Sector rules on consent, record retention, professional conduct, marketing communications and automated decision-making are absent from the questionnaire. A recruiting screen and a marketing draft can produce identical scores while sitting in completely different legal categories, and in Europe a workflow touching hiring, credit, insurance or education can attract obligations that no score in this tool would hint at.

It does not model data residency, vendor insolvency or model deprecation either, and each of those can break a workflow quietly. Where your data physically sits and which subprocessors touch it is a contractual question this tool never asks. Whether your platform will still exist, or still be owned by the same company, in eighteen months is a diligence question about the vendor rather than the workflow. And a model deprecation notice can change your agent's behaviour on a date somebody else chose, which is why a workflow with no evaluation set is fragile no matter what it scores here.

A worked example: scoring two clinic workflows

First, an appointment reminder cascade at a medical clinic: a confirmed appointment approaches, each stage fires only if the previous one did not resolve, and a reply gets routed to the front desk. Score it. Autonomy is acts with approval at index 2, sensitivity is personal data at index 3 because the message names a patient and a time, who sees the output is customers after review at index 3, reversibility is undo with effort at index 1 because a wrong reminder can be corrected with a phone call, and volume is hundreds at index 2.

The arithmetic: 2 times 1.1 is 2.2, 3 times 1.2 is 3.6, 3 times 1.0 is 3.0, 1 times 1.3 is 1.3, and 2 times 0.9 is 1.8. That sums to 11.9, divided by 22 is 0.541, so the score is 54 and the band is Medium. The control list you get is the five-item one: approval before external or written action, citations on claims, PII redaction before anything reaches the model, logging of inputs, outputs and approver, and rate limits with a kill switch. For a reminder cascade that is a fair description of a competent build.

Now change one answer. Make the same clinic workflow fully autonomous, index 4 rather than 2. The autonomy contribution goes from 2.2 to 4.4, the total becomes 14.1, and the score is 64, which crosses into High and swaps in the seven-item control list including two-person sign-off and staged rollout. One answer moved the workflow across a band, which is the correct behaviour and also a warning: with only five questions, the bands are sensitive, and being honest about the autonomy question matters more than being precise about any of the others.

For contrast, score an internal weekly report draft at the same clinic: autonomy suggests actions at 1, sensitivity internal at 1, internal only at 0, trivial to undo at 0, tens of actions at 1. That is 1.1 plus 1.2 plus 0 plus 0 plus 0.9, which is 3.2 over 22, or 15, comfortably Low. The gap between 15 and 54 for two workflows in the same building is the entire point of the exercise, and it is why blanket policies about AI usually fail: they are calibrated for one of these two and wrong for the other.

Mapping your risk score to the NIST AI Risk Management Framework

Your risk score is a lightweight version of what the formal frameworks do at enterprise scale, and knowing the mapping makes the result more defensible. The NIST AI Risk Management Framework, released in January 2023, organizes AI risk work into four functions: Govern, which sets policies and accountability; Map, which establishes context and identifies risks; Measure, which analyzes and tracks them; and Manage, which prioritizes and treats them. The scorer's questions about sensitive actions and reversibility are Map work, and the approval tier it recommends is a Manage decision.

NIST followed with a Generative AI Profile in July 2024, published as NIST-AI-600-1, which names twelve risk areas specific to generative systems, including confabulation, information security, and human-AI configuration. Confabulation is worth singling out for agent workflows: it is the formal name for the failure where a model states something false with full confidence, and it is why the scorer treats hard-to-detect mistakes as higher risk than obvious ones. An error a reviewer can spot in seconds is a cost; an error that looks plausible is a liability.

For a small team, the practical translation is simple. You do not need a governance committee to use the same logic: name an accountable owner for each agent workflow, write down what it touches and what could go wrong, track edits and failures from the logs, and set the autonomy tier accordingly. That is Govern, Map, Measure, and Manage at operator scale, and it means your risk decisions can be explained to a client, an auditor, or an insurer in language the frameworks already use.

What OWASP says can actually go wrong with agents

The OWASP Top 10 for LLM Applications, in its 2025 edition, is the most cited catalog of the failure modes your score is protecting against. Prompt injection sits at number one, LLM01: any agent that reads external content, whether emails, web pages, support tickets, or documents, can encounter instructions embedded in that content and act on them as if they came from you. The standard defense is exactly what a low-autonomy tier enforces: untrusted input never triggers consequential actions directly, and a human gate sits between the reading and the doing.

The risk most specific to agents is LLM06, Excessive Agency: an agent granted more functionality, permissions, or autonomy than its job requires. It earns its place on the list because it converts every other weakness into real-world impact; a prompt injection against a read-only summarizer produces a bad summary, while the same attack against an agent with write access to your CRM and email produces actions you have to undo publicly. Scoping tool access to the minimum, the scorer's core recommendation, is the named mitigation.

OWASP has since gone deeper on agent-specific threats, publishing an Agentic AI Threats and Mitigations guide in February 2025 covering memory poisoning, tool misuse, and runaway autonomy, and a full Top 10 for Agentic Applications in December 2025 built with input from more than 100 researchers and practitioners. The direction of all this work matches the scorer's tiers: agents are treated as systems whose permissions, not whose intelligence, determine the damage a failure can cause.

The regulatory clock and the incident curve

Risk scoring stopped being purely voluntary in Europe. The EU AI Act entered into force on August 1, 2024, its prohibitions and AI literacy obligations applied from February 2, 2025, obligations for general-purpose AI models followed on August 2, 2025, and the regulation becomes fully applicable, including the high-risk system requirements and penalties, on August 2, 2026. If your agents touch hiring, credit, insurance, education, or other high-risk categories and you have EU exposure, the risk tiering exercise this tool performs informally is becoming a legal expectation with documentation requirements attached.

The incident data explains the urgency. Stanford's AI Index counted 233 reported AI incidents in 2024, a record and a 56 percent increase over 2023, and its 2026 edition counted 362 incidents in 2025. Those are only the publicly reported cases; the operational near-misses that never make a database are the ones your logging and review tiers exist to catch. A rising incident curve alongside rising deployment is not surprising, but it is exactly the environment in which unlogged autonomous agents age badly.

Gartner connected risk practice to project survival directly: its June 2025 prediction that over 40 percent of agentic AI projects will be canceled by the end of 2027 lists inadequate risk controls as one of the three causes, alongside cost and unclear value. The inverse reading is the useful one. Workflows with defined approval tiers, escalation rules, and monitoring are cheaper to keep, easier to defend, and far less likely to end as a quiet cancellation after their first public mistake.

What changed in 2025 and 2026

Agent risk went from blog-post topic to formal guidance. Within roughly one year, OWASP shipped its agentic threats taxonomy in February 2025 and an agent-specific Top 10 in December 2025, NIST's generative AI profile became the reference for enterprise AI policies, and the EU AI Act's compliance dates started landing. A team scoring workflow risk in 2026 has named frameworks to align with, which also means a team that skips the exercise has named frameworks it is visibly ignoring.

Platform tooling caught up with the tiered-autonomy model. LangGraph's 1.0 release made human-in-the-loop pause and approve a first-class API, OpenAI's AgentKit shipped with guardrails built into its Agent Builder, and n8n, Make.com, and Zapier all support approval steps as ordinary workflow nodes. The draft-only, approve, sampled-review, and autonomous tiers your score recommends are no longer custom engineering; they are configuration choices on every mainstream platform.

The market also punished the opposite approach. The high-profile pattern in 2025 and 2026 incident reporting was not exotic model failures; it was ordinary excessive agency, agents with write access acting on bad or manipulated context. The operational lesson embedded in your score has stayed stable through all the model releases: capability upgrades change what an agent can do, while your permissions, gates, and logs decide what it can damage, and only one of those is under your control.

The limits: five questions is not a risk assessment

This is a heuristic, and treating a heuristic as an authority is the specific failure we want to avoid. Five questions cannot cover prompt injection through untrusted content, poisoned memory, tool misuse, cascading failures between chained agents, or the plain case of an agent that is confidently wrong in a way the reviewer cannot detect. Those are documented agent failure modes with published mitigations, and none of them is an input to this arithmetic.

It also assumes your answers describe reality. Autonomy is the question people get wrong most often, usually by describing the approval step they intend to add rather than the one that exists. A build that ships with approval and quietly loses it eight weeks later because the reviewer got busy is the most common way a Medium workflow becomes a High one, and no score recomputes itself when that happens. Rescore after launch, not just before.

The score says nothing about likelihood. It is a blast-radius estimate: how bad, applied to how much, seen by whom, how hard to undo. It does not ask how often the agent gets it wrong, because that number does not exist until you have run the workflow and counted. Two workflows with identical scores can have wildly different real risk if one has an error rate of one in fifty and the other one in five thousand, and the only way to know which you have is logged output and a reviewer counting edits.

Finally, it is not legal, security or compliance advice, and it is not a document you can hand to an auditor or an insurer as evidence of a risk assessment. It is a structured prompt for a conversation you should be having with whoever owns that responsibility in your business. If you have no such person and the workflow touches money, health, hiring or legal matters, that absence is the finding, not the score.

When not to use this tool, and what to use instead

Do not use it as the only gate on a regulated workflow. If your agent touches patient records, client funds, credit decisions, insurance underwriting or employment screening, the questions that decide whether you may proceed are not on this page. Score the workflow here for your own sequencing, then take the answer to whoever holds the professional or legal obligation, and expect their constraints to change the design.

Do not use it to compare against another company's number. The bands are ours, the weights are ours, and the same workflow scored by a different framework will land somewhere else. It is a ruler for your own workflows and nothing more. Comparing your 54 to somebody else's 54 is meaningless.

Do not use it before you know what the workflow is. Scoring an idea produces a score for an idea. Build the specification first, with a named trigger, a named source of truth and a named approver, then score the thing you actually intend to run. The workflow planner and the workflow builder both produce that artefact, and both take less than half an hour.

And if the score comes back High, the next step is not a different tool, it is a smaller workflow. Almost every High workflow can be split into a low-risk part the agent does and a high-risk part a person keeps. Drafting the message rather than sending it, proposing the refund rather than issuing it, flagging the record rather than changing it. That split is free, it usually captures most of the time saving, and it is the single most reliable risk control on this page. If you want it built that way from the start, the done-for-you route ships approval gates and logging as standard.

At a glance

Frameworks and rules relevant to scoring AI agent risk (verified July 2026)

Framework or ruleDateWhy it matters for agents
NIST AI RMF 1.0January 2023Govern-Map-Measure-Manage structure for AI risk programs
NIST Generative AI Profile (AI-600-1)July 202412 gen AI risk areas, including confabulation
OWASP Top 10 for LLM Applications2025 editionPrompt injection (LLM01) and Excessive Agency (LLM06)
OWASP Agentic AI Threats and MitigationsFebruary 2025Agent-specific threat taxonomy: memory, tools, autonomy
OWASP Top 10 for Agentic ApplicationsDecember 2025Dedicated agentic risk list from 100+ contributors
EU AI Act full applicationAugust 2, 2026High-risk obligations and penalties apply

What mistakes should you avoid?

  • Treating all agent actions as equal risk.
  • Allowing customer-facing automation before review metrics exist.
  • Hiding failures instead of logging them.
  • Forgetting privacy and permission boundaries.

FAQ

How is the AI agent risk score actually calculated?
Five factors are each answered on a five-point scale stored as 0 to 4, then weighted: reversibility 1.3, data sensitivity 1.2, autonomy 1.1, who sees the output 1.0, and volume 0.9. The weighted answers are summed and divided by the maximum possible total of 22, then expressed out of 100. Sixty or above is High, 33 to 59 is Medium, and below 33 is Low. The weights are an editorial judgement rather than a measurement, chosen so that irreversibility and sensitive data pull hardest.
What risks does this scorer deliberately not model?
Several important ones. It does not model regulatory exposure, so a workflow touching hiring, credit, insurance, health or education can score Low while carrying obligations the tool never mentions. It does not model data residency or which subprocessors handle your data. It does not model vendor insolvency or acquisition, or model deprecation, where a provider retires the model your agent depends on and behaviour changes on a date you did not choose. It also does not model prompt injection, memory poisoning or tool misuse. Treat the score as a blast-radius ranking, not a risk assessment.
Does a Low score mean the workflow is safe to run autonomously?
No. Low means the five questions asked all pointed the same way: small blast radius, cheap mistakes, internal audience, easy to undo. It does not mean nothing can go wrong, and there is no threshold in this model below which harm stops being possible. Even at Low the tool insists on three controls: log every input and output, keep a kill switch that pauses the agent, and review a weekly sample of outputs. Those exist precisely because the sampling is how you find the failure mode the questionnaire never asked about.
When can an agent act without human approval?
When the action is reversible, the audience is internal or the output is checked before it reaches anyone external, the data involved is not sensitive, the workflow has been running under supervision long enough to produce an error rate you can quote, and there is logging that would let you find and undo a bad run. Missing any one of those means keeping the gate. Autonomy is something a workflow earns from evidence, which is why the sensible sequence is draft mode, then approval, then sampled review, then unattended.
Why did changing one answer move me from Medium to High?
Because with only five factors the bands are sensitive, and autonomy carries a weight of 1.1 across a range of four steps. Moving from acts with approval to fully autonomous adds 2.2 points of weighted score, which is ten points on the 100 scale and enough to cross the 60 boundary on its own. That sensitivity is intentional. Removing the human from the loop genuinely is the largest single change you can make to a workflow's risk profile, and the model is built so that the answer you are most tempted to fudge is also the one that moves the result most.
Does human review remove the risk?
It reduces it and changes its shape. A reviewer catches the obvious errors immediately, which is why a workflow whose mistakes are visible in seconds is so much safer than one whose mistakes look plausible. What review does not fix is the confident, well-formatted, wrong output that a busy reviewer approves without checking the source, or the gradual decay where a 100 percent approval rate quietly becomes a rubber stamp. Give reviewers criteria, source visibility and a sample they can genuinely inspect, and track edit rate as a number rather than a feeling.
Should I rescore a workflow after it launches?
Yes, and most teams never do. The score describes the workflow as designed, and workflows drift: approval steps get skipped under pressure, volume climbs, someone connects a second system, or the agent gets write access it did not have at launch. Any one of those changes an input to the calculation. Put a rescore on the calendar at 30 days and again whenever the agent gains a new tool or a new permission, because permissions rather than intelligence decide what a failure can damage.
Is this score enough to satisfy a client, auditor or insurer?
No, and it should not be presented that way. It is a ten-minute heuristic with five questions and weights we chose, not a documented risk assessment against a named framework. Its legitimate use is internal: ranking your own workflows, choosing a starting control set, and structuring the conversation with whoever carries the professional or legal responsibility. If a third party needs assurance, they will want your control evidence - logs, approvers, escalation rules, incident history - and this tool is at best the thing that told you to collect it.

Sources & further reading

Was this page helpful?