The demand side of the agent economy

The AI governance landscape addresses several different questions

AI governance is a broad category. To understand the landscape, it helps to separate the questions buyers are trying to answer. These categories overlap; a single platform may address several of them.

Building and orchestration: how do we create, test, and deploy an agent? Salesforce’s Agentforce documentation, for example, describes building agents and connecting them to actions and business data. This is part of the capacity side of the market: making AI able to perform work.

Identity and access: who or what is acting, and which resources may it use? Microsoft Entra Agent ID provides identity foundations for agents, alongside governance capabilities. An agent identity is an important control surface; a useful assignment also needs a description of its intended work.

Discovery and security: which AI assets exist, including ones the organization has not approved? Noma describes discovering agents and other AI assets and maintaining an inventory. This addresses visibility into the environment and the security team's need to understand what is operating.

Risk and assurance: how are AI systems evaluated, documented, and governed over their lifecycle? IBM watsonx.governance addresses governance and risk across AI assets. This work extends beyond agent construction into organizational oversight.

Operational context: how does the business run, and what information helps AI act usefully within it? Celonis explicitly positions its process intelligence and context model around business operations and AI. This is an adjacent space that matters when assessing DutyGraph's position.

References: Salesforce: Design and Implement Agents · Microsoft: Entra Agent ID · Noma: AI asset discovery · IBM: watsonx.governance · Celonis: Context Model.

Agent capacity needs an equally clear definition of demand

We use “supply side” as shorthand for the tools that make agents available and operable: models, builders, orchestration, identities, access controls, and security. Those tools serve different purposes, and governance itself is broader than this metaphor.

The demand side asks what the business needs accomplished. Which task contributes to an outcome? What starts it? What does good completion look like? Who needs the result? What judgment or authority must stay with a person?

A buyer's interest in AI is commercial demand. Here, we mean something more operational: a reviewed description of work that can become a candidate assignment. A pile of vague automation ideas does not provide that.

An agent can be correctly authenticated and still be doing poorly specified work. A beautifully specified task can also be unsafe to execute without appropriate controls. Productive delegation needs both.

The structured work unit is the bridge

A role such as procurement manager contains several duties. A duty such as maintaining supplier records contains multiple tasks. Checking whether documents are present, preparing a record draft, resolving contradictions, and routing final approval have different inputs, risks, and decision requirements.

A useful work unit names its responsible human and parent duty. It describes the trigger, incoming information and source, ordered actions, software, expected output, recipient, and exceptions. It preserves the evidence behind the description.

For a delegation proposal, that description also needs a review boundary and a way to judge success. What may the agent prepare? What may it change? When must it stop and ask? How will a person know that the result is correct?

This structure gives a builder a better specification, a reviewer a clearer scope, and an advisor something concrete to discuss. It does not, by itself, prove that an agent will perform the task reliably. That requires evaluation.

From “automate procurement” to an assignment someone can review

Consider a fictional distributor whose procurement team maintains supplier records. An interview reveals six tasks: check the incoming packet, prepare the record draft, resolve missing information, prepare the Finance handoff, summarize open handoffs, and route final approval.

Document checking might be a candidate for AI assistance. The input is a supplier folder and an approved checklist. The actions are to compare the files with the checklist and identify missing or unreadable material. The output is a draft checklist with source references. A named person reviews it.

That assignment does not include changing bank details, approving a supplier, or activating a record in an enterprise system. Those actions require different authority and evaluation.

The same discovery may reveal that nobody knows who gives final supplier approval. Adding an agent to chase the request would leave the ownership problem unresolved. The immediate productive action is a leadership decision.

This is why the upstream work matters. It can identify a suitable automation candidate, narrow an overbroad request, or show that the next improvement needs a human decision. Each is a useful outcome.

Where DutyGraph enters the picture

DutyGraph starts with an advisor and a company. Public research supplies initial context. Leadership explains departments, responsibilities, goals, and problems. The people doing the work then describe their tasks, including the exceptions and handoffs that a job title rarely captures.

AI prepares granular task descriptions from those accounts. Participants review and edit them to confirm their understanding. The advisor can inspect the returned cards and follow connections in the company work map.

That review is deliberately specific: a participant confirms what they understand their work to be. Conflicting accounts can still exist, ownership can still need resolution, and an automation proposal still needs separate authorization.

Our intended position is upstream of selecting and governing an agent: helping the company establish the work it wants to delegate and the human responsibility behind it. The result should make the next conversation with an implementation or governance team more concrete.

A blue-ocean direction, with a testable distinction

The opportunity we see is to compete on the clarity of the assignment and the value of discovery. Can a company leave the process knowing which work matters, where a handoff fails, and which bounded tasks deserve an AI pilot?

That is a blue-ocean direction for our product strategy. It is not evidence of an empty market. Process intelligence, task mining, business architecture, workflow design, and consulting already address parts of this problem. Agent platforms also incorporate business context.

Celonis's stated focus on operational context is a useful reminder of that overlap. DutyGraph must earn its position through its particular entry point: advisor-led discovery, employee-reviewed task accounts, and a traceable connection from duty to proposed delegation.

The distinction needs to hold up in practice. Less effort to reach a usable task specification, fewer corrections, and better decisions about what to automate would be meaningful evidence. Describing a new category is only the starting point.

References: Celonis: Process Intelligence and effective AI agents.

Better discovery can make the surrounding ecosystem more useful

For an advisor, the work map creates a basis for recommending an intervention. The answer might be AI assistance, a clearer responsibility, a workflow change, or further investigation. Software consolidation could be worth exploring when the evidence supports it.

For an agent builder, reviewed work units can become candidate specifications and evaluation scenarios. For an identity or access team, the same record can explain why a scope is requested and who is responsible for the underlying work. For a risk reviewer, it can preserve the intended boundary and supporting account.

These are complementary roles. DutyGraph's proposed governance direction is to connect a task and its responsible person to a manifest reviewed against effective access and policy. Actual provisioning and enforcement require real integrations; a task description cannot grant permission.

Our current advisor pilot supports discovery, reviewed task descriptions, work maps, and strategy drafts. The governance examples are fictional demonstrations, and the pilot does not establish live IAM provisioning or customer-system agent execution.

Start with the work you need done

Before choosing another agent, choose one business outcome and ask the people involved to explain the work that contributes to it. Capture the inputs, actions, outputs, systems, and dependencies. Review the account together. Then decide which task is worth testing with AI.

The agent economy needs capable technology and trustworthy controls. It also needs assignments grounded in the reality of a business.

That is the demand side DutyGraph aims to make visible: structured work that people understand, can review, and can deliberately choose to delegate.