Aldea Signals

Thesis

The Org Chart Is Getting a Second Layer

Agents are taking on ongoing work. The investment question is which parts of coordinating that work become independent businesses—and which are absorbed into the enterprise stack.

Andrew PadillaOctober 6, 20267 min read

On September 29, 2026, OpenAI introduced dots: persistent agents with their own cloud computers that can work across connected applications. It also previewed specialist dots with distinct identities, credentials and defined organizational responsibilities. Those specialist deployments are beginning with focused enterprise pilots; broader teams of dots remain part of the company’s stated direction. OpenAI

Our hypothesis is that the strongest independent opportunities will emerge where valuable work crosses systems, requires judgment and lacks a clear software owner. A product that takes responsibility for completing that work can become part of how the business operates.

Delegated work needs responsibilities, context, dependency management and exception handling. The org chart begins to acquire a second layer. The commercial question is which parts of that structure customers will buy separately—and which their existing software will provide.

The new problem is discretion, not persistence

Workflow infrastructure already addresses persistence, retries and recovery. Temporal’s agent reference architecture, for example, separates durable orchestration from individual actions and supports explicit pauses for human decisions. Temporal

What is changing is the range of work software can attempt without every step being specified in advance. n8n’s new agent functionality illustrates the distinction: agents can decide which steps to take and use predefined workflows as controlled tools. A workflow can also call an agent for a step requiring flexible reasoning. The functionality remains in preview. n8n Blog

Consider a hypothetical manufacturer preparing a customer quotation.

A standard request might follow a predictable sequence: check inventory, apply pricing rules, estimate delivery and obtain approval. A less standard request might require interpreting an ambiguous specification, asking a clarifying question, identifying a substitute component and deciding whether engineering needs to review it.

A revised specification might invalidate an earlier price. An alternative component might require a different delivery commitment. An approval might apply to an outdated proposal.

The opportunity is to let software resolve more of these variations while preserving the dependencies and controls that make the result trustworthy.

Permission to update a quotation does not establish that the quotation is ready to send.

Illustrative workflow / 01

A changed request can make earlier work stale

The agent may decide what to do next. The process still has to track which decisions need to be made again.

01

Customer request

A specification changes

02

Product choice

Recheck the component

03

Commercial terms

Recalculate price and delivery

04

Human approval

Review the current proposal

Hypothetical manufacturer quotation. A previous approval cannot release a proposal whose underlying specification, price or delivery commitment has changed.

One phrase, several markets

“Agent orchestration” can obscure the distinction between several different businesses.

Execution infrastructure keeps a process running: preserving state, handling failures and coordinating actions. Temporal’s architecture illustrates this layer. Its commercial case depends on developer adoption, production workloads and the cost of operating them. Temporal

Governance infrastructure establishes which agents exist, what they can access and how their activity is supervised. Microsoft positions Agent 365 as a control plane for observing, governing and securing agents, including those built by partners. A new entrant must offer enough beyond existing identity, security and administrative tools to justify a separate purchase. Microsoft

Business coordination software helps a team complete a process. Dust’s Vanta deployment connects departmental knowledge and agents. Dust Nexcade takes a more specialized approach, bringing together requests, rate sources, transport systems and internal rules for freight workflows, with human review points. Nexcade

These businesses may overlap, but managing agent credentials and preparing freight quotations address different buyers and budgets. The investment case starts with the problem a company owns and what the customer would otherwise use.

The independent opening is between systems

Supporting several models does not by itself justify an independent platform. OpenAI says it is working with Microsoft to integrate specialist dots with Agent 365’s governance and security controls: one provider supplies the agent, while an established platform supplies part of its management. Customers may assemble an adequate solution from capabilities they already buy. OpenAI

An independent vendor must beat that alternative after implementation, maintenance and the burden of managing another supplier. We see the more promising opening in business coordination, where a process spans applications and departments but no existing application owns the outcome.

In the manufacturer’s quotation process, customer requirements, production capacity, supplier information and approvals might live in different systems. The product’s job would be to keep them aligned as a request changes, so the business can send a quotation it can actually fulfill.

It also needs a commercial owner: someone with a budget who will sponsor deployment and remain accountable for the result. Crossing organizational boundaries does not automatically create a market.

Both horizontal platforms and specialists can pursue this opening. A horizontal platform might give teams a repeatable way to build and operate workflows. A specialist might arrive with more of the process already understood, reducing what the customer has to implement.

Measure the work, not the agents

Google Research’s January 2026 evaluation of 180 agent configurations found that performance depended on the relationship between the architecture and the task. Multi-agent coordination helped on some parallelizable work and reduced performance on some sequential tasks. Adding agents was not a universal improvement. Google Research

In its June 2025 account of building a multi-agent research system, Anthropic reported substantially higher token consumption than ordinary chat interactions. The task must justify the extra resources; those observations concern its particular systems, not a universal cost benchmark. Anthropic

The best coordination software should sometimes choose fewer agents—or a conventional workflow.

A commercial example comes from Vanta. In a Dust case study, Vanta describes using departmental agents covering finance, compliance and customer feedback to prepare quarterly business reviews. It reports saving roughly two hours per week for each of 200 representatives, or approximately 400 hours per week in total. Dust

This is a defined process with a reported time saving. It remains a vendor-published customer account without a complete accounting of implementation, maintenance and review costs, or evidence that Dust is the only way to achieve the result.

We would measure the fully loaded cost of a completed, accepted workflow: model and infrastructure usage, implementation, ongoing support, human review and rework. Time saved in one department is less valuable if a comparable burden appears elsewhere.

Implementation services can help a customer adopt the product. But if each deployment requires the same process documentation, data cleanup and bespoke engineering, growth may depend on adding delivery staff at the same rate as customers.

The investment measure / 02

Count the work the customer accepts

An agent count or gross time-saving claim leaves out what it takes to produce a usable result.

The output

A completed and accepted workflow

Account for the full cost

Model & infrastructure
Implementation
Human review & rework
Ongoing support
Conceptual accounting, not a measured benchmark. The commercial test is whether the full cost per accepted outcome falls as the product is deployed and used again.

The advantage has to improve with use

Conversation history, execution logs and stored preferences can make a product harder to replace while also preserving mistakes and outdated assumptions.

In the hypothetical quotation workflow, reviewers might repeatedly flag a specification change as requiring engineering approval. Useful learning would identify similar cases earlier, route them correctly and reduce unnecessary escalations without missing important exceptions. That requires recording what was corrected and whether the eventual outcome was acceptable.

The test is whether accumulated operating experience reduces the cost and difficulty of completing the next unit of work.

That improvement need not come from training a proprietary model. It could come from better routing, clearer business rules, more reliable evaluations or a better division between fixed workflows and model-driven decisions.

Three advantages matter here: customer-specific operating knowledge can create switching costs; reusable deployment methods can improve the vendor’s delivery economics; and cross-customer learning can improve the product itself.

One does not establish the others. A company can become deeply embedded in an account while learning little that makes the next deployment easier. It can also develop reusable capabilities without pooling sensitive customer information.

What would strengthen—or weaken—the thesis

We would read annual recurring revenue growth and net revenue retention alongside the work customers delegate. Expansion into more valuable processes, less correction at comparable complexity and faster subsequent deployments would strengthen the thesis. Higher model consumption alone would not.

We would become less positive if customers mainly used the product as an interchangeable interface to models; if implementation staffing grew in proportion to revenue; if human review remained high without a corresponding increase in task complexity; or if customers could recreate the useful workflows in their existing software with little loss.

Improving models pose a further test: do they let the product take on more work at a lower cost, or eliminate much of what it charges for? The most attractive companies should benefit from model progress without depending on today’s limitations for their relevance.

Some may ultimately become part of larger platforms. ServiceNow’s completed acquisition of Moveworks in December 2025 brought together an enterprise assistant, search and reasoning capabilities with an established workflow platform. It demonstrates strategic interest in combining these functions—not that every company in the category will earn an attractive acquisition outcome. ServiceNow Newsroom

Acquisition can be an outcome. It should not be the premise.

The org chart metaphor has a limit. Enterprises need a clearer division of responsibility and less effort moving work between people and systems. Human accountability remains when execution is delegated.

The valuable second layer is the one customers trust with more work—and have a durable reason to keep paying for.

Sources & further reading

  1. 01Introducing dots— openai.com
  2. 02AI Agent Reference Architecture— go.temporal.io
  3. 03Introducing n8n Agents— blog.n8n.io
  4. 04Microsoft Agent 365, now generally available— Microsoft
  5. 05How Vanta's GTM team saves thousands of hours annually with Dust— dust.tt
  6. 06Freight quote automation for freight forwarders— nexcade.ai
  7. 07Towards a science of scaling agent systems— research.google
  8. 08How we built our multi-agent research system— anthropic.com
  9. 09ServiceNow completes acquisition of Moveworks— newsroom.servicenow.com

Authored and reviewed by Andrew Padilla.

From Aldea

Occasional research and notes
from Aldea.

No noise—a few pieces a year.

Aldea© 2026 Aldea Ventures · Barcelona