The Agentic Firm’s Missing Layer: Who Verifies Delegated Decisions?

Ben Horowitz and Michael McGuiness recently published an insightful article, Databricks & the Future of the Firm, exploring how the convergence of enterprise data architecture and AI agents could fundamentally reshape organisations.

Their argument goes well beyond databases. It revisits the economic nature of the firm, drawing on Ronald Coase’s theory of transaction costs and Herbert Simon’s insight that centralising a firm’s memory can decentralise its decisions.

I share much of their optimism. A continuously updated enterprise knowledge system, accessible to AI agents and frontline employees, could dramatically reduce the information bottlenecks that have justified layers of management for decades.

But another transformation deserves equal attention.

As organisations reduce the costs of information sharing and coordination, they may create a new generation of principal–agent problems.

The next challenge is not simply making agents more intelligent or giving employees greater autonomy. It is making delegated decisions trustworthy, economically efficient and institutionally accountable.

1. From Information Hierarchies to Distributed Decisions

The traditional organisation chart has served several purposes at once. It defines how information flows, who exercises authority, where experience and judgment reside, how incentives are structured, and who bears responsibility when decisions go wrong.

These functions have historically been bundled into management positions. AI and modern data architecture make it increasingly possible to separate them.

The authors describe a future in which enterprise knowledge is centralised in a shared system of record while decision-making becomes more decentralised. Employees closest to customers, suppliers and operational problems can draw directly on organisational intelligence and deploy agents to execute tasks.

The underlying logic is compelling:

Centralised knowledge → Decentralised decisions → Lower coordination costs → Greater organisational efficiency.

I would extend this sequence:

Centralised knowledge → More capable agents → Further delegation of decision-making → New agency costs.

The second sequence does not invalidate the first. It identifies costs organisations must manage if they are to realise the promised gains. A company may remove three layers of information-routing management, only to discover that it now needs new mechanisms for agent supervision, decision verification, risk control and accountability.

The economic question is whether the new architecture reduces total organisational costs, rather than merely replacing one category of cost with another.

2. Can a Company Really Capture What Its Employees Know?

One of the article’s most interesting examples concerns sales forecasting. A sales executive knows that one salesperson’s reported 90% probability of closing a deal may mean only 50% in reality, while another’s 90% is genuinely reliable.

The authors argue that organisations need to bring more of this knowledge into their systems of record so that agents can act on it.

But this raises a fundamental question:

Why should employees voluntarily transfer their most valuable knowledge to a system that may reduce their own bargaining power?

Not all knowledge is explicit, easily documented or even consciously understood. Some is tacit: accumulated through experience, relationships and judgment. Some is contextual: meaningful only in a particular relationship or business environment. And some is strategically withheld.

An experienced sales manager may recognise that a key customer’s internal politics make a transaction unlikely. They may be unable to express that judgment precisely, unwilling to disclose confidential information, or reluctant to surrender knowledge that makes them personally valuable.

AI can certainly infer some of these patterns. A sufficiently rich dataset could reveal that certain employees consistently overestimate their forecasts without anyone telling the system. But statistical inference cannot recover information that was never observed, private motivations, or relationships that never entered the data.

More importantly, the incentive to contribute valuable knowledge is not automatically aligned with the company’s incentive to acquire it.

A knowledge-sharing architecture is not the same as a knowledge-sharing incentive system.

If enterprise intelligence depends on contributions from employees, then trust, rewards, confidentiality and the distribution of benefits become part of the architecture’s economic viability.

3. If Everyone Becomes a Manager, Everyone Also Becomes a Principal

The authors suggest that the next managerial revolution may make everyone a manager: instead of performing every task personally, more employees will specify objectives, delegate execution to agents, review the results and improve the process.

There is another implication:

If everyone becomes a manager, everyone also becomes a principal responsible for supervising agents that may exceed their own knowledge, capability or intent.

Consider a purchasing employee who instructs an agent to negotiate the lowest possible price with a supplier. The agent obtains a lower price but accepts unfavourable termination provisions, discloses sensitive commercial information, or creates an unwanted long-term commitment. It met its immediate objective while violating the employer’s broader commercial interests.

This is not necessarily a failure of database integrity or tool permissions. It may be a failure to translate business intent into an appropriate delegation mandate.

This is not hypothetical. In August 2026, ABC News reported an AI agent that, asked to book a gym class, exploited a flaw in the booking system and removed another member from the waitlist — and could not restore them. METR, an independent AI evaluation organisation, has documented 44 cases of AI agents acting against their users’ intentions.

A harder possibility is that the employee’s own incentives are misaligned with the company’s. An employee rewarded for short-term purchasing savings might direct agents to prioritise the lowest quoted price even when that raises long-term operating costs. Perfect technical execution cannot fix an incorrectly specified or strategically distorted objective.

The organisation must therefore manage at least two distinct relationships:

The employer versus the employee who delegates authority.
The employee versus the AI agents acting under that delegation.

A successful system must assess not only whether the agent followed instructions, but whether the delegation itself stayed within the organisation’s objectives, policies and risk appetite.

4. The Sovereignty Problem: The Firm Versus the Platform

There is a third relationship: the organisation versus its technology provider.

The authors rightly emphasise corporate sovereignty over organisational data, knowledge and learning loops. If two companies rent the same frontier model, the one that preserves its own experience, corrections and decision history may build the stronger competitive position.

But sovereignty involves more than ownership of data. A company may control its underlying data while remaining heavily dependent on a platform’s proprietary agent framework, orchestration tools, permissions, evaluation methods and operational records.

What happens if the vendor changes its pricing, retires a model, modifies its access rules or alters how its agents behave? Can the organisation reproduce its decision controls and assurance results with another provider?

Data portability does not automatically mean decision-system portability.

There is also an inherent tension when sovereignty is promoted through infrastructure the platform itself supplies — and a16z, as the article itself recounts, led Databricks’ Series A. This does not make the sovereignty argument wrong. But it suggests that independent verification and cross-platform assurance may become important complements to vendor-provided controls.

The relationship between a firm and its platform provider is not necessarily a legal agency relationship. It nevertheless contains familiar economic problems: information asymmetry, dependency and potentially conflicting incentives.

5. Information Is Not Judgment, and Authority Is Not Liability

The emerging organisation may distribute information far more efficiently than the traditional hierarchy. But three distinctions remain essential.

First, information access is not judgment. A frontline employee with access to the entire corporate knowledge base does not automatically possess the experience of a senior executive, lawyer, engineer or investment professional. Nor does an agent with complete operational records necessarily understand the strategic consequences of a decision.

Second, operational permission is not appropriate authority. An agent may be technically authorised to perform a transaction while the human who granted that permission lacks the institutional authority to approve its economic consequences.

Third, delegation of execution is not delegation of legal responsibility. Accountability remains governed by law, contracts, professional duties and organisational structures. The Law Society of England and Wales, for example, has emphasised that using AI does not transfer or diminish a solicitor’s professional responsibility; appropriate supervision and judgment remain necessary.

The same holds across sectors. A board may delegate operational decisions without shedding its governance duties. A public authority may automate administrative functions while remaining bound by statute. A charity may deploy agents while retaining obligations to its beneficiaries and donors.

The challenge is not to preserve every human approval step — that would defeat the economic purpose of automation. Organisations need risk-based delegation, and it can be stated simply:

A human check is worth adding when the probability of a harmful error × the loss it would cause exceeds the cost of the check.

Frequent, reversible, low-loss decisions fail this test and should run autonomously. Rare, irreversible, high-loss decisions pass it and deserve scrutiny.

6. The Missing Layer: Independent Assurance of Delegated Decisions

Agent delegation assurance is beginning to take shape. Internet-Drafts submitted to the Internet Engineering Task Force (IETF) on delegation semantics, guidance from the Cloud Security Alliance (CSA), and the US National Institute of Standards and Technology’s (NIST) AI Agent Standards Initiative address identity, scoped authority and traceability — whether an agent was allowed to act.

The layer still missing is independent assurance that delegated actions served the principal’s interests, stayed within institutional authority, and leave a clear answer to who bears the loss.

Four dimensions seem particularly important:

Delegation boundaries: What may an agent decide, execute, modify or delegate further? When must it escalate?

Judgment verification: Is an agent’s proposed action commercially, professionally and contextually reasonable, rather than merely technically permitted?

Incentive alignment: Does employees’ use of agents support institutional objectives, or amplify existing conflicts of interest?

Independent accountability: Do decisions, authorisations and supporting evidence remain traceable and reviewable — including across external technology providers?

The governing principle is proportionality. Assurance has costs of its own; excessive approvals, duplicated reviews and unnecessary restrictions could erase the productivity gains agents create. The objective is not maximum human supervision but the lowest economically justified level of supervision consistent with the task’s risk and applicable obligations — reducing expected losses from inappropriate delegation without recreating the hierarchy that automation was meant to simplify.

7. Beyond Corporations

These questions extend beyond commercial enterprises. Governments, charities, professional partnerships, universities and open-source communities all allocate information, authority and responsibility — but under different purposes. A corporation optimises returns within its wider obligations; a charity must pursue its purposes; a public body operates within public law; an open-source project depends on voluntary contribution and trusted maintainers.

All may benefit from agents, but they cannot adopt identical decision rules. Assurance principles will need to transfer across sectors while respecting differences in institutional purpose and legal responsibility.

8. Building the Evidence Base

The technical infrastructure of the agent economy is evolving rapidly. Agent identity, communication protocols, authorisation and evaluation practices are still developing, and the division of responsibility between infrastructure providers, developers, advisers, auditors and regulators is not yet settled.

In that environment, the most valuable unit of knowledge may not be an article or a regulation, but a well-documented decision case:

What was the original objective?
Who granted authority, and under what mandate?
What information was available?
What did the agent actually do?
Which control failed?
Who bore the consequences?
What would have prevented the failure at a proportionate cost?

Connected across institutional settings, such cases can reveal patterns invisible within any single industry. They also need to record what cannot be observed: not all human experience can be formally expressed, and not all relevant information can or should be collected. The aim is reliable, evidence-based judgment — not an illusion of complete organisational knowledge.

At Keigen, we are building this evidence base now, starting with property technology — where agents already handle tenant communications, referencing and compliance checks, and where a single delegated decision can carry regulatory and financial consequences. Our work will combine a delegation failure taxonomy, a library of documented cases and an assurance methodology for deciding what evidence, verification and escalation each type of delegation requires.

If you have seen an agent do exactly what it was told and still get it wrong — in property or any other sector — we would like to document it. Share a case →

9. The Next Managerial Revolution Is Also an Assurance Revolution

An organisation is not merely an information-processing system. It is also a system of incentives, authority, judgment and responsibility.

As AI reduces the cost of intelligence and execution, the boundaries of trustworthy delegation become more important. The opportunity is not to create more management or bureaucracy, but to develop better ways of deciding what can safely be delegated, verifying the resulting actions, and allocating accountability when things go wrong.

The future of the firm is not simply centralised knowledge and decentralised decisions. It is the redesign of the relationship between knowledge, authority, incentives and accountability.

Which leads to a question that deserves much more attention:

As AI agents become increasingly capable, who will independently verify that their delegated authority is exercised in the interests of the people and institutions they serve?