Human capital and token capital: why enterprise AI value depends on both

For years, enterprises measured competitive advantage with a familiar set of assets: capital, intellectual property, technology, people. Build enough of each and deploy them well, and the advantage held.

AI hasn't changed which assets matter. It has changed how fast an advantage built on the wrong balance of them turns into a liability.

The clearest way to see the shift is as two distinct forms of capital every enterprise now has to build: token capital and human capital. Token capital is the proprietary AI capability a company builds and owns: its systems, models, and workflows. Human capital is the knowledge, judgment, relationships, and pattern recognition of its people. Neither shrinks as the other grows. If anything, human capital gets more valuable as token capital scales, because agents without human direction just produce output faster, not better.

The organizations winning with AI have already figured this out. They've stopped treating humans and agents as competing resources and started treating them as two forms of capital that compound together.

Microsoft CEO Satya Nadella named this shift well in his June 2026 essay on the subject: without human direction, he wrote, you just have compute running in circles. That framing is right, and it's spreading fast for a reason. What it doesn't answer is what happens when the direction itself breaks down at scale, when nobody can tell whether the human judgment that's supposed to be steering the system is actually reaching it. That's the part worth working through.

Why human-in-the-loop still beats full AI automation

Most people assume human judgment is a fallback: something you lean on when an agent gets stuck or makes an obvious mistake. That's backwards. Judgment isn't damage control. It's a separate skill, and it's one AI can't do yet. Humans are good at finding the gaps agents can't see:

• They sense when a situation has drifted outside the policy an agent was designed for.
• They carry the unspoken context of a customer relationship that no training run captured.
• They know when the technically correct answer is the wrong answer for this person, in this moment, given the history and the thing that happened on last quarter's call that nobody wrote down.

What makes people valuable here isn't the tasks a model has already learned to do well. It's their ongoing ability to spot what the model hasn't learned yet.

McKinsey research points the same direction: companies that combine human and AI workflows outperform those chasing full automation in complex, judgment-heavy work. That contextual judgment isn't something to minimize. It's something to capture, by turning every judgment call into a structured signal that makes the agents working alongside that person better over time.

What is token capital, and how does it show up in AI agents?

Token capital covers everything from a single AI-assisted search to a fully autonomous workflow, but agents are where it gets risky fastest, because they're the only form of it that acts on its own without waiting for a person to approve each step. That's what most organizations picture when they think about their "AI program" anyway: the agents, the processing power, the throughput, the ability to run a workflow end to end with no human involved. It's what lets a company handle ten thousand routine requests at 2 a.m. with the same consistency as the first one of the day. No fatigue, no boredom, no shortcuts under deadline pressure.

But an agent has no built-in way to check itself against reality. It optimizes for the objective it was given at the moment it was configured. It doesn't know when that objective has gone stale.

Why most enterprise AI agent projects fail to deliver ROI

A fully autonomous deployment can look excellent from the outside: agents running end to end, human touchpoints minimized, throughput high, dashboards green. For a while, everything holds. Then something shifts, gradually at first and then all at once:

• A policy changes and the agents don't know yet.
• A regulation takes effect and nobody can demonstrate what controls were active during the period it covers.
• A customer segment starts behaving differently, and the assumptions baked into an agent three months ago no longer reflect reality.

The agent is still fast, still confident, still producing outputs that look correct on the surface. It's operating on assumptions that have gone wrong, and nothing in the system is built to catch that on its own.

Nobody notices right away, because everything downstream still looks fine: tickets close, the dashboard stays green, the agent reports success. By the time someone does notice, the agent has made the same bad call hundreds or thousands of times, and the work isn't just fixing the model anymore. It's figuring out how far back the damage goes.

The numbers back this up:

• Only 15% of AI decision-makers report a measurable EBITDA lift from AI in the past 12 months (Forrester), even as global agentic AI spending is projected to reach roughly $202 billion in 2026 (Gartner).
• Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls, not because the underlying technology failed (Gartner).

This isn't a technology failure so much as a design failure: the result of treating autonomy as the destination rather than a calibrated mode within a system that keeps human judgment informed and involved.

How human feedback makes AI agents better over time

Every interaction in a well-designed human-and-agent system generates something valuable: a structured record of how your business actually makes decisions. Not generic decisions, your decisions, shaped by your policies, your risk appetite, your customer relationships, and the accumulated judgment of your most experienced people.

• When a human reviews what an agent produced and approves it unchanged, the system learns what good looks like in your domain.
• When a human edits before sending, the gap between the agent's draft and the final output is a lesson.
• When a human escalates a case rather than letting the agent proceed, the trigger condition becomes known and can be encoded.

Over thousands of interactions, these signals accumulate into institutional knowledge a competitor can't easily replicate. Not because they lack access to the same model (they likely have it), but because they don't have your people's judgment applied to your situations, measured against your definition of good, captured in a system that keeps learning from it.

For most of business history, human capital was real but nearly impossible to represent as a hard asset. It lived in people's heads, in habits and relationships, not in any system that could capture and carry it forward. A system that captures the trace of every human judgment call, and feeds it back into the agents working alongside that person, makes that tacit knowledge visible and durable for the first time. Token capital deployed at scale is valuable. The institutional knowledge encoded in every human-and-agent interaction is the asset that endures, and it's what widens the gap between your program and everyone else's.

How to govern AI agents at scale (without slowing deployment)

That compounding effect is powerful, but it doesn't survive scale on its own. Here's what most organizations discover somewhere between month six and month eighteen: the collaboration itself was never the problem. Human judgment working alongside AI agents is the right architecture. The problem shows up when that collaboration runs on informal coordination instead of structural governance.

At small scale, informal works. Someone reviews outputs, someone escalates when things feel off, and the loop closes through conversation and proximity. It's imperfect but manageable. Then the program grows. At fifty agents across multiple teams and geographies, informal starts to crack:

• Policies diverge between teams without anyone deciding they should.
• Accountability for a given decision becomes genuinely unclear.
• An audit request takes days to reconstruct who reviewed what and which guardrails were active.

At two hundred agents, that gap between the right idea and the right infrastructure becomes a serious liability.

BCG's research on responsible AI backs this up: enterprises that build governance into their AI architecture from the start, instead of bolting it on as an oversight layer later, get better outcomes on both performance and compliance (BCG). Governance isn't a constraint on the compounding described above. It's what makes the compounding possible at scale.

The layer that holds all of this together already has a name in the systems that do it well: the harness, not the model. It governs which models, which data, and which tools are in the loop, and it keeps a continuous feedback loop running across all three. The model underneath is swappable, and it will get swapped as better ones ship. The harness is what stays constant, and it's where the durable advantage actually compounds.

How Artemis delivers human-in-the-loop AI governance

Artemis is built on the idea running through this whole framework: humans and agents aren't competing for the same work, they're partners in a system. The platform's job is to make that partnership explicit, governed, and continuously improving. Artemis is that harness: the layer that decides what's allowed to run on its own, and what has to stop for a person.

1. Every decision is traceable: Artemis captures a reasoning-aware trace of every agent decision: what was proposed, what a human decided to do with it, what happened next. When a compliance officer asks what controls were active on a specific interaction from weeks ago, the answer is immediate and complete.
2. Human-in-the-loop from the start: Approval gates, review steps, human ownership transitions, and supervisor copilot behavior are declared directly in the Agent Blueprint Language, the same typed specification that defines how an agent reasons and what it's not allowed to do.
3. Autonomy calibrated by risk, not treated as a binary switch: Policy gates, budget limits, entitlement checks, and human-in-the-loop triggers are all declared in the agent specification and enforced at runtime, so the agent knows when to proceed and when to stop, every time.
4. Every human action improves the system: Every override, correction, approval, and rejection is a structured signal that feeds the eval suite, which feeds Arch's analysis, which proposes specific, reviewable improvements for engineers to approve. The institutional knowledge your people carry every day doesn't stay locked inside individuals. It gets encoded into a system that carries it forward regardless of who's in the room on a given day.
5. One governance surface across every agent, team, and channel: Identity, policy, audit trail, and accountability stay consistent across the entire platform, so the controls built for the first agent apply automatically to the two hundredth.

How to tell if your AI program is governed or just automated

The question isn't whether you have AI agents deployed. Most organizations do. The question is whether the people in your organization are genuinely in the loop, contributing judgment that makes the system better, or simply watching from a distance and intervening when something goes visibly wrong.

• Can you trace every consequential decision an agent made, including who reviewed it and when?
• Do your agents know reliably when to stop rather than proceed?
• Does the judgment your people apply every day feed back into how your agents improve?
• Is your governance consistent across every agent on the platform, or rebuilt from scratch by each team?

If those questions surface gaps, you're not behind on AI. You're behind on architecture, and that's a solvable problem if you start with the right foundation.

That's what Artemis was built to be: the layer that lets human capital and token capital compound instead of compete. If you want a real answer instead of a gut check, the same six questions we've been asking 400+ AI leaders are now a two-minute scored benchmark.

[Take the AI Governance Benchmark →] (interactive tool: ai-governance-benchmark-tool.html)

FAQs

What is the difference between human capital and token capital in enterprise AI?

Token capital is the proprietary AI capability a company builds and owns, its agents, models, and workflows. Human capital is the judgment, relationships, and pattern recognition of its people. They aren't substitutes: agents running without human direction just produce output faster, not better, so the organizations getting the most value from AI invest in both and build systems that let them compound together.

Why do most enterprise AI agent projects fail to show ROI?

Only 15% of AI decision-makers report a measurable EBITDA lift from AI in the past 12 months, according to Forrester, and Gartner predicts more than 40% of agentic AI projects will be canceled by 2027. The reason is rarely the model itself. It's usually a governance gap: agents keep running on assumptions that have gone stale, and there's no structured way for a human to catch it before it compounds.

What does "human-in-the-loop" mean for AI agents?

Human-in-the-loop means an agent is built to pause and hand a decision to a person at defined points, such as approvals, escalations, or high-risk actions, instead of running fully autonomously from start to finish. In a well-governed system, those stopping points are declared directly in the agent's specification and enforced at runtime, so the boundary doesn't depend on the agent's own read of the situation.

How should enterprises govern AI agents at scale?

Governance needs to be built into the AI architecture from the start rather than added on later. BCG research finds enterprises that embed governance structurally, instead of treating it as an oversight layer, get better outcomes on both performance and compliance. In practice, that means one consistent policy and audit layer across every agent and team, so the controls built for the first agent still hold at the two hundredth.

What is an AI "harness," and how is it different from the model?

The harness is the governance layer around a model that decides which models, data, and tools it can use, and when a human has to step in. The model underneath is swappable and will change as better ones ship; the harness is what stays constant, which is why it's where durable competitive advantage in enterprise AI actually builds up.