← Insights

The 5 Dimensions of Agent Risk: A Scoring Methodology

Ritesh Vajariya · 5 January 2026

When your security team evaluates risk for a new application or vendor, they have established methodologies. NIST's risk management framework. FAIR quantitative analysis. ISO 31000. These frameworks provide structured, repeatable processes for assessing and prioritizing risk.

No equivalent exists for AI agents.

Most organizations are either ignoring agent risk entirely or assessing it ad hoc — a quick conversation between the deploying team and a security engineer, maybe a one-page questionnaire borrowed from the application security review process. The result is inconsistent evaluation, blind spots, and an inability to compare risk across agents or prioritize governance investments.

What follows is a scoring methodology designed specifically for AI agents. It assesses risk across five dimensions that together capture the full surface area of agent-related exposure. Each agent in your inventory is scored across all five dimensions, producing a composite risk score that maps to governance tier requirements.

Updated 18 October 2026. An earlier version of this article scored each dimension 1–10 and published tier bands at 3.1, 5.1 and 7.1. The method below is the one Quontera implements — 1–5 per dimension, a composite of 5–25, bands at 10, 15 and 20 — so that scoring an agent by hand from this page and scoring it in the register produce the same number. Two of the dimension names changed with it; the section on what is no longer scored explains where that material went.

The Five Dimensions

Dimension 1: Data Access Scope — 25%

This dimension measures the breadth and sensitivity of data an agent can access.

An agent that reads public-facing FAQ content operates in a fundamentally different risk category than an agent that queries your customer database, accesses employee records, or reads financial data. The distinction isn't just about volume — it's about sensitivity classification, regulatory implications, and the consequences of unauthorized disclosure.

What can this agent read, and how sensitive is it?

ScoreDescriptor
1Public or already-published data only.
2Internal operational data. No personal data, no financials.
3Personal data for a limited population, or non-material financial data.
4Personal data at scale, material financial data, or regulated records.
5Special-category personal data, credentials, or the crown-jewel dataset.

For example: an agent that summarizes published product documentation scores 1. One that queries the sales pipeline to generate forecasts scores 2. One that reads customer support tickets containing personal details scores 3, or 4 if it reads all of them. One holding the credentials to the billing system scores 5.

Assessment questions:

  • What data sources does this agent connect to?
  • What is the classification level of the data it accesses?
  • Does the data include PII, PHI, or regulated financial information?
  • Is the data access scoped to the minimum required, or does the agent have broad read access?
  • Could the agent's access be exploited to extract data beyond its intended scope (prompt injection, conversation manipulation)?

Dimension 2: Autonomy Level — 25%

This dimension measures the degree of human oversight in the agent's operation.

Autonomy exists on a spectrum. At one end, an agent that suggests actions for human approval. At the other, a fully autonomous agent that operates continuously without human review. The appropriate level of autonomy depends on the stakes of the agent's decisions and the organization's tolerance for unsupervised action.

It carries the same weight as data access, and for the same reason: both decide how much can go wrong before anyone is in a position to stop it.

How much does it do without a human approving the step?

ScoreDescriptor
1Drafts and suggests only. A human performs every action.
2Acts only after a human approves each individual action.
3Acts within a narrow scope; a human reviews afterwards.
4Acts continuously within set bounds. Humans sample rather than review.
5Acts without bounds or routine human review, including on new cases.

For example: an agent that drafts replies for a representative to send scores 1. One that handles routine inquiries on its own and escalates anything unusual scores 3. One running unattended overnight, deciding cases nobody anticipated, scores 5.

Assessment questions:

  • Does a human review agent outputs before they're acted upon?
  • Is there real-time monitoring of agent behavior by a human?
  • What triggers escalation to human review?
  • How frequently are agent actions reviewed after the fact?
  • Could the agent operate for an extended period without anyone noticing a problem?

Dimension 3: Blast Radius — 20%

This dimension measures how far a bad action travels before somebody catches it.

It is the dimension that most often gets scored too low. The instinct is to score the agent's intended reach — one team, one workflow — when what matters is the reach of a mistake. An agent with write access to a shared customer record system has the blast radius of that system, not of the task it was built for. The question is not what the agent is for. It is what is downstream of it.

Three things usually set it: what the agent can write to, whether it can communicate outside the organization, and how many records or people one erroneous run touches before the error becomes visible.

How far does a bad action travel before someone catches it?

ScoreDescriptor
1One record, one user, caught immediately.
2A single team or workflow. Contained within a day.
3A department, or a customer-visible action affecting a few accounts.
4A whole business line, or many customers at once.
5The whole organization, the customer base, or a market-facing system.

For example: an agent that files one document into one user's folder scores 1. One that updates records in a CRM four teams work from scores 3. One that posts to your public status page, or emails your customer base, scores 5.

Assessment questions:

  • What systems can this agent write to, and who else depends on those systems?
  • Can the agent send external communications (email, chat, API calls to third parties)?
  • How many records, customers or transactions does a single erroneous run touch?
  • How quickly would an error become visible to somebody who could stop it?
  • Are there hard limits on the magnitude or frequency of the actions the agent can take?

Dimension 4: Reversibility — 15%

This dimension measures whether the action can be undone, and how quickly.

It is deliberately separate from blast radius, because the two fail differently. An agent that miscategorizes ten thousand internal documents has a wide blast radius and a trivial remedy. An agent that transfers funds to one wrong account has a narrow blast radius and no remedy at all. One dimension cannot express both, and collapsing them is how an irreversible action in a small system gets scored as low risk.

Score the worst realistic case, not the average one.

Can the action be undone, and how quickly?

ScoreDescriptor
1Trivially reversible. Undo is a button.
2Reversible within hours by the owning team.
3Reversible with effort, coordination, or a restore.
4Largely irreversible. Remediation, not reversal.
5Irreversible. Funds moved, data disclosed, or a statement published.

For example: a miscategorized internal document scores 1. A record overwritten in a system with point-in-time restore scores 3. A refund issued, a disclosure made or a statement published scores 5.

Assessment questions:

  • If this agent takes a wrong action, what is the actual undo procedure?
  • Who performs it, and how long does it take?
  • Does reversal depend on a backup, and when was the restore last tested?
  • Has anything already left the organization — a payment, a disclosure, a published statement?
  • Is the remedy a reversal, or an apology and a remediation program?

Dimension 5: Regulatory Exposure — 15%

This dimension measures the regulatory and legal landscape the agent operates within.

Agents operating in regulated industries or handling regulated data face a higher governance bar. The scoring reflects not just current regulatory requirements but the trajectory of AI-specific regulation, which is moving quickly toward mandatory governance requirements for autonomous systems.

Which obligations attach to what it touches?

ScoreDescriptor
1No specific regime applies.
2General data protection duties only.
3Sector guidance applies, or a documented internal control does.
4A binding regime applies directly (financial, health, employment).
5High-risk classification, or a regime with personal liability attached.

For example: an agent helping employees brainstorm marketing copy scores 1. One processing customer data in a retail context scores 2. One assisting with insurance claims or medical triage scores 4, or 5 where the regime attaches personal liability.

Assessment questions:

  • What sector-specific regulations apply to this agent's operations?
  • Does the agent handle data covered by GDPR, HIPAA, SOX, CCPA, or other privacy/security regulations?
  • Is the agent's domain subject to AI-specific regulation (EU AI Act high-risk categories)?
  • Has the relevant sector regulator issued AI governance guidance?
  • Would the organization need to explain this agent's governance to a regulator?

What This Deliberately Does Not Score

An earlier version of this method scored a sixth idea as a dimension of its own: action scope — what the agent can write to, send, approve or execute. It is not scored here, and that is a considered position rather than an omission.

Action scope is not a judgement. It is an inventory fact, and it should be recorded rather than rated. In Quontera it is held as a permission record per system the agent can reach, carrying the action, the system and its sensitivity, and it is what decides which governance controls apply to the agent in the first place. Scoring it 1–5 would throw that structure away and replace it with somebody's impression of it.

What action scope does feed is two of the five dimensions above. The breadth of what an agent can write to is most of its blast radius. Whether those writes can be undone is most of its reversibility. Rating it separately double-counted it, which is why an agent with broad write access used to score high twice for one underlying fact.

The Composite Risk Score

Each dimension is scored 1–5. The five scores are weighted, then rescaled to a composite of 5–25.

DimensionWeightRationale
Data Access Scope25%Data exposure is the most common source of material harm
Autonomy Level25%Higher autonomy means less opportunity to catch errors
Blast Radius20%How far a mistake travels before anyone can stop it
Reversibility15%Whether the mistake can be undone at all
Regulatory Exposure15%Reflects regulatory and legal risk multiplier

Composite = 5 × [(Data × 0.25) + (Autonomy × 0.25) + (Blast × 0.20) + (Reversibility × 0.15) + (Regulatory × 0.15)]

All ones give 5. All fives give 25. Nothing else can land outside that range, which is what makes two agents' scores comparable and what makes the tier bands below mean something fixed.

Three details matter more than they look:

Round once, at the end. Rounding each weighted term before summing would let a 15% dimension move the result as hard as a 25% one, which quietly discards the weights you just argued about. Round the composite, half-up, and nothing else.

Rationale is required at 4 and 5. The high end is where a score gets challenged — by an approver, by an auditor, by a director a year later — so it is where the reasoning has to be on the record. A 4 with no stated reason is not a score, it is an assertion.

Record the weights with the score. If you ever change the weighting, every historical score becomes unreproducible unless the weights that produced it were stored alongside it. Re-scoring should write a new version rather than overwrite the old one, for the same reason.

A worked example

An agent that reads regulated customer records (data access 4), acts continuously within set bounds while humans sample its output (autonomy 4), touches a few customer accounts per run (blast radius 3), moves money it cannot claw back (reversibility 5), and operates under a binding financial regime (regulatory exposure 4):

(4 × 0.25) + (4 × 0.25) + (3 × 0.20) + (5 × 0.15) + (4 × 0.15) = 3.95, and 3.95 × 5 = 19.75, which rounds to 20 — High, at the very top of the band.

Note what that example also shows: the agent's lowest score is its blast radius, and it is still High, because an irreversible action under a binding regime does not become safe by affecting fewer people. A method that averaged nothing but breadth would have scored it Medium.

One property worth knowing

All threes — a genuinely middling agent on every dimension — produces exactly 15, which is the Medium ceiling. So any single dimension rising from 3 to 4 moves that agent to High, including the two weighted at 15%. This is a property of the bands, not an accident: a middling agent is one step away from the tier that requires a governance review, which is roughly where an honest method should put it.

Governance Tier Mapping

The composite maps to a governance tier. The tier sets the review cadence and the length of the approval path — and nothing else.

TierCompositeReview cadenceApproval path
Low5–10AnnualOwner submission, then an independent security review
Medium11–15QuarterlyAdds a compliance review
High16–20MonthlyAdds a governance review
Critical21–25ContinuousAdds an executive sign-off

So a Low-tier agent passes two reviews and a Critical-tier agent passes five, every one of them after the owner's own submission carried out by somebody other than the person who asked for it. "Continuous" is a posture rather than a date, but an overdue list still has to sort by something, so a Critical agent is re-checked weekly.

What the Score Does Not Decide

A composite is an input to a decision, never the decision.

A prohibited action blocks a deployment at Low tier exactly as it does at Critical. A mandatory control that has not been satisfied is not waived by a score of 6. There is no arithmetic by which a well-behaved risk assessment buys an exception, and a method that allowed one would be a method for laundering the exception rather than for measuring the risk.

This is worth stating explicitly inside your own program, because it is the first thing that gets quietly negotiated when a deployment is late.

Putting It Into Practice

Score every agent in your inventory. Build a risk heat map. Prioritize governance investments on your highest-scoring agents. Use the scores to have evidence-based conversations about which agents need more controls, which agents might need to be decommissioned, and where to allocate security and compliance resources.

The methodology is designed to be repeatable and comparable — you can track an agent's risk score over time as its capabilities, data access, or operating context change. And you can use aggregate scores to give leadership and the board a portfolio-level view of agent risk.


This methodology is implemented, not just described. Quontera scores every agent across these five dimensions, with these weights and these bands, stores each assessment as an immutable version with the weights that were in force, and sets the review cadence from the resulting tier. Re-score an agent next quarter and this quarter's score stays readable. See the register →

The score is an input, never the decision. A low score does not authorize a deployment, and it cannot waive a mandatory control or a prohibited action. The tier decides how many approval stages apply — nothing more. Pricing and plans →


Ritesh Vajariya is the CEO of AI Guru and founder of Quontera (formerly AgentGuru). Previously AWS Principal ($700M+ AI revenue), BloombergGPT Architect, and Cerebras Global Strategy Lead. He has trained 35,000+ professionals and built products serving 50,000+ users.


More on agent governance