Tokenomics of an Agentic SOC: When the Failure Shows Up as a Bill, Not a Backlog

When a human-staffed SOC has an alert spike, you see it. The backlog grows, SLAs slip, and someone escalates. When a monolithic agent experiences the same spike, nothing in the queue looks wrong. Every alert is processed at full cost, and without proper cost monitoring, the failure arrives later as an invoice.

That was the argument we brought to InfoSec Nashville on September 18, and it is the part of the agentic SOC conversation most security leaders have not budgeted for. Tokenomics is the cost of the tokens an AI agent consumes to investigate each alert, and in an agentic SOC it becomes a line item that needs the same scrutiny as headcount and licensing. Full autonomy is the pitch. An AI-augmented SOC is the operating model.

What Is an Agentic SOC?

An agentic SOC uses AI agents that reason, leverage multiple tools, and act within a defined scope to handle security operations work. It is not the same as more automation. Automation is efficient, and agentic systems are adaptive and work to deliver an outcome. Three properties define the term:

• Autonomous reasoning: Interpreting ambiguous, novel situations instead of matching fixed conditions.

• Tool use: Querying systems, pulling enrichment, and executing actions on its own initiative.

• Bounded action: Operating within an explicitly scoped authority, not unlimited autonomy.

Arctiq’s SOC operations delivery runs from Reactive (manual triage, siloed tools) to Assisted (SOAR playbooks) to AI-Augmented (copilots with a human in the loop) to Agentic (bounded autonomy, oversight by exception). Crawl, walk, then run. It is not a leap into a fully autonomous model.

Why Is the Traditional SOC Under Pressure?

Alert volume, tool sprawl, and analyst burnout are growing faster than headcount can absorb. The numbers:

• Alert volume: Forrester research puts the average SOC team at roughly 11,000 alerts a day, with only about 22 warranting investigation.

• Tool sprawl: An average enterprise stack runs 47 point tools, according to Ponemon survey data cited in CISO MindMap 2026.

• Burnout: Tines reports that 71% of analysts experience burnout and 64% are considering leaving within a year.

• Dwell time: Mandiant’s M-Trends 2026 shows median dwell time rose to 14 days in 2025, up from 11 in 2024.

The reversal in Dwell time is the takeaway. Dwell time has been falling for nearly a decade. The increase in dwell time underscores the challenge of scaling telemetry, tooling, and staffing in the environments they are meant to protect.

Standing still- the cost of doing nothing has a price tag too. IBM and Ponemon’s Cost of a Data Breach Report 2026 puts the 2025 global average cost of a breach at $4.99M, a record high and up 12% year over year. Organizations with extensive use of AI and automation saved an average of $1.93M per breach compared with those not leveraging automation.

Can a Fully Autonomous SOC Work?

No, not on its own. In an Autonomous SOC build, we’ve identified three structural ceilings, and none of them come down to how capable the model is. The pitch is appealing: alerts in, agents triage, correlate, investigate, and close, and humans watch a dashboard. The argument for an Autonomous SOC lands because burnout is disguised as a headcount problem, alert growth outpaces budget, and the hype cycle rewards the boldest claim. None of those reasons address whether a fully autonomous SOC is viable.

Ceiling 1: Capability

Agents are pattern-matchers. They do well on known TTPs and campaign IOCs, repetitive triage steps, and correlation against last quarter’s map. They stumble with novel incidents, attacker behavior outside their training, and business context the model hasn’t accounted for.

This is not a “wait for a better model” problem. Threats keep evolving, and environments change with new tooling, new architecture, and new business context. That guarantees a steady stream of genuinely new alerts. A human in the loop is there to reason about what is actually new, not to classify against what has already been seen.

Ceiling 2: Accountability

High-impact security decisions, including containment, customer notification, and regulatory disclosure, require clear human accountability and appropriate oversight. Breach notification obligations, cyber insurance requirements, and board governance reinforce the need for defined decision-making authority. Even a hypothetically perfect model does not absorb liability.

Ceiling 3: Economics

A monolithic, single-loop agentic architecture offers limited opportunities to control spending at individual investigation steps. The next two sections cover what drives that cost and why it can be hard to see coming.

What Drives the Cost of an Agentic SOC?

Four factors drive cost per alert, and how much data is in the context window can be one of the largest.

1. Context window: How much log and telemetry data is loaded into the prompt per investigation.

2. Tool-call iterations: The loops, retries, and escalation chains an agent runs before it reaches a verdict.

3. Model tier: A cheap, fast model versus a frontier model, and critically, who decides which one runs.

4. Orchestration overhead: The coordination cost between agents, tools, and the SOAR and SIEM layer.

Why Does Token Spend Spike When Alert Volume Rises?

In a monolithic agent loop, token spend can track alert volume closely. If the same investigation process runs for every alert, tripling the volume can roughly triple consumption, even when many alerts are routine.

An alert spike in a human-staffed SOC typically becomes visible through declining output, growing backlogs, and slipping SLAs. An agentic system continues processing without those same warning signs, while consumption-based costs accumulate. In an agentic SOC, without monitoring, spending limits, and appropriate controls, the financial impact may not become apparent until later.

Headcount grows in discrete hiring decisions, while token spend can grow linearly or worse, exponentially. These are fundamentally different cost curves. Model your token spend curve against your alert volume curve before you commit budget, not after.

How Does a Skills-Based Architecture Reduce Token Costs?

It routes the routine alerts and escalates the exception. Instead of one opaque loop, investigations are broken into inspectable, step-by-step skills:

1. A small, fast model classifies against known IOCs and CVE signatures first, allowing routine cases to be handled at a lower cost.
   
   
2. Genuine misses and high-risk cases escalate to a frontier model, reducing the number of investigations that require more expensive reasoning.
   
   
3. Cost can track the complexity of each alert instead of raw volume, so token spend can grow more slowly than alert volume.

The same granularity lets a human intervene mid-investigation, supporting the accountability and oversight requirements discussed earlier. One design decision helps address two problems. A single opaque loop offers fewer checkpoints for analysts and fewer opportunities to apply cheaper models to routine work, which is why “fully agentic” and “skills-based agentic” are not the same product, even when marketing uses the phrases interchangeably.

What Should Humans Own in an AI-Augmented SOC?

Humans should retain oversight of decisions that are novel, irreversible, or legally consequential. The dividing line is reversibility and risk. The question everyone should be asking: Is the action hard to undo, or does it cross a trust boundary?

• Agents own: Enrichment, correlation against known patterns, first-pass triage, and other repetitive, well-specified work within defined permissions.

• Humans own: Novel judgment, escalation decisions, legal and customer exposure, and high-impact actions such as isolating a host, disabling an account, notifying a customer, or making a regulatory disclosure. These decisions require human sign-off.

Build-in this decision gate during implementation. Re-engineering a process after a deployment is stable can be disruptive and costly.

What Does Agentic SOC Governance Require?

Agents are non-human identities and need the same lifecycle discipline as human identities: provision, enforce, rotate, and deprovision. Left unmanaged, agent sprawl lead to shadow IT, with ownership today that is: “nobody yet.” It also broadens your attack surface via: prompt injection, agent hijacking, data poisoning, and automation complacency.

We recommend borrowing model risk management practices from financial services and requiring a tested kill switch for every autonomous agent in production. “The model decided” is not an acceptable answer to a regulator or a board. For more on the security side, see Joe’s posts on preparing security programs for the agentic AI era and AI governance.

Governance is a function of data. Normalized, correlated, queryable telemetry is usually the real bottleneck, not the model. Tim Tipton covers this in The Agentic SOC Has a Prerequisite Nobody Is Selling.

Which Metrics Should Security Leaders Track?

Track operational, financial, and accuracy metrics:

• MTTD and MTTR shift
• Analyst capacity multiplier
• False-positive reduction
• Cost per alert and cost per incident resolved, fully loaded and including token spend
• Agent accuracy and precision, tracked continuously
• Percentage of Tier-1 and Tier-2 volume handled without escalation
• Analyst retention
• Token spend per case, as a trend line you review regularly

What to Do Monday

For practitioners: Make cost per alert resolved a real KPI, not just MTTR or accuracy. Then map your workflow to the three ceilings. Where does judgment actually sit today, and does your tooling respect that boundary, or paper over it?

For leadership: Evaluate “full autonomy” claims against the three ceilings and ask which one the vendor claims to have solved, and how. Then apply a procurement test: Is the investigative logic portable outside that runtime, or are you locked to their roadmap and pricing?

The goal is not fewer humans. It is better-placed humans. An AI-augmented SOC is the model that is still solvent after the alert storm, and still standing after the incident review.

Taking the Next Step

Whether you are modeling agentic SOC token costs for the first time or pressure-testing a vendor’s full-autonomy claims, Arctiq’s cybersecurity team is glad to talk through where your SOC stands. Connect with us to get started.