01 The next layer is meaning and action
An agent can access five systems and still misunderstand the business. It can find the customer, contract and invoice, yet miss the distinction between a commercial exception and an approved standing policy. Connectivity solves access. Knowledge engineering addresses meaning. Governance determines what action is allowed.
For the CFO, these are strategic capabilities with operating costs, ownership choices and concentration risk. Their purpose is to help protect, create and orchestrate enterprise value. Trusted Finance requires financial meaning, authority and evidence to survive every handoff between people, applications and agents.
SAP describes a direction in which applications expose capabilities and a context layer connects data, process knowledge and orchestration. Salesforce documents API access to headless agents. Foundation Capital argues that decision context may become a valuable enterprise software layer. These are different kinds of evidence: vendor architecture, technical documentation and an investor thesis. None proves that systems of record disappear. SAP ↗ Salesforce ↗ Foundation Capital ↗
02 Knowledge engineering is more than document retrieval
Retrieval can find relevant material for a model. A reliable business decision also needs to know which definitions, relationships, policies and permissions apply at the time of action.
| Layer | What it makes explicit | Example in customer credit |
|---|---|---|
| Business meaning | Entities, relationships and agreed definitions. | The customer’s legal entity, group exposure and the difference between revenue and cash collected. |
| Policy and constraints | Rules, scope, limits, approvals and effective dates. | Which credit policy applies in this jurisdiction and who may approve an exception. |
| Provenance | Where information came from and how it changed. | Source receivables, disputed amounts and the transformations behind the exposure calculation. |
| Decision context | Relevant evidence, alternatives, business rationale and approvals. | Why a temporary limit was approved and the conditions attached. |
| Outcome and learning | What was executed and what subsequently happened. | Payment behavior, losses, missed sales and whether the exception should be reconsidered. |
This does not require one enormous ontology before any useful work begins. Start with the smallest meaningful domain. Use ordinary tables, rules, documents or graph structures according to the problem. Established provenance models such as W3C PROV-O provide useful concepts for recording actors, activities and sources; they do not establish that a decision was correct. W3C ↗
A context graph connects decision events and business entities. A knowledge graph usually emphasizes entities, relationships and meaning more broadly. Terminology varies, and the concepts overlap. The investment question is which missing knowledge prevents a better decision, not which fashionable label a vendor uses.
03 Capture judgment without turning history into policy
Past actions contain insight and mistakes. A customer concession approved during a supply interruption should not become the default for every renewal. A precedent may be relevant only to one legal entity or period. A recorded explanation may rationalize a decision rather than establish its real cause.
Assign knowledge owners. Distinguish authoritative policy from contextual evidence. Keep effective dates, scope and provenance. Test a proposed rule against reviewed cases before allowing it to affect execution. Record changes and retire content that no longer applies.
The enterprise should capture purposeful decision records, with appropriate permissions and retention. Andreessen Horowitz’s essay on recorded work highlights the potential value of richer context. Koko’s recommendation is narrower: collect the information needed for a legitimate business purpose, rather than treating indiscriminate recording as a prerequisite for enterprise intelligence. Andreessen Horowitz ↗
The economic test: does better context reduce material errors, repeated expert effort or missed opportunities enough to justify curation and maintenance? If not, simplify.
04 Headless SaaS changes the architecture debate
Headless SaaS separates business capabilities from a product’s own user interface. An agent or another application may invoke those capabilities through APIs and events. An agentic orchestration layer coordinates tasks and decisions across them, with policy and human intervention where required.
Trusted Finance must survive the loss of the familiar screen. A headless process still needs an authoritative financial record, tested calculations, valid approvals and a clear distinction between a proposed action and a posted transaction. Reconcile what the orchestrator requested with what the system actually executed; route discrepancies to an accountable owner.
The authoritative ledger, identity controls and transaction logic still matter. A simpler interface does not eliminate complex obligations. Nor does an open protocol make application semantics, licensing or workflow behavior interchangeable.
MCP provides a way for AI applications to connect to tools and resources; A2A specifies communication between independent agents. Both can support integration. Financial authority, definitions, acceptance criteria and reconciliation still have to be designed. Model Context Protocol ↗ A2A Protocol ↗
| Architecture choice | Why it may be attractive | What the CFO must underwrite |
|---|---|---|
| Incumbent suite and native agents | Existing data, workflows and controls may reduce integration effort. | License changes, cross-suite coverage, evidence access and dependency. |
| Enterprise platform or external orchestrator | Can coordinate work across applications and offer common tooling. | Additional control layers, duplicated charges, semantic consistency and failure recovery. |
| Enterprise-owned orchestration | Greater control over distinctive decisions, context and supplier choice. | Engineering, maintenance, evaluation, security and continuity obligations. |
| Hybrid pattern | Places commodity execution and differentiated decisions in different layers. | Clear boundaries, a common evidence model and ownership across suppliers. |
Koko does not assume that owning software always creates more value. Own or control the capabilities that matter to advantage, and negotiate access to the evidence needed to govern them. Rent commodity components where the economics and accountability work.
05 A registry is useful; authority must reach the action
Workday’s announced agent system of record, Microsoft’s Agent 365 and ServiceNow’s governance tooling illustrate competition around agent management. They also raise a practical issue: an enterprise may have several inventories claiming to be the management center. The 2025 Workday announcement is used here as evidence of direction, not a current availability guarantee. Workday ↗ Microsoft ↗ ServiceNow ↗
Map those inventories to an enterprise account of who owns each workflow, what it can do and where policy is enforced. A dashboard that reports a risk does not necessarily block an unauthorized payment. Test enforcement and recovery at the action boundary.
The CFO and control owners should be able to distinguish an agent’s recommendation, a user’s approval and the completed transaction. The same record should support cost analysis, exception investigation and value measurement without implying that all these responsibilities belong to one vendor.
06 Negotiate before the dependence becomes invisible
Require clear answers on API access and quotas; data and decision-record portability; model training and data-use rights; audit access; identity and delegation; change notification; retention; service continuity; and transition support. These are procurement requirements to assess with legal and technology teams, not assertions about rights a contract already provides.
Model costs at the expected mix of successful work, retries, escalations and human review. Compare seat, consumption and outcome-based commercial models. An outcome fee is useful only when the outcome, acceptance and responsibility for error are clearly defined. Oliver Wyman’s procurement guidance and McKinsey’s agent-economics interview both make this a relevant CFO conversation. Oliver Wyman ↗ McKinsey ↗
The move now: run one workflow through the proposed architecture, export its decision evidence, test a prohibited action and estimate the cost of changing one supplier. That exercise is more informative than a diagram promising universal orchestration.
07 What should compound
The reusable value may live in trusted definitions, evaluated decision patterns, maintained policies, a proficient workforce and an evidence record that improves future choices. A growing collection of agents is not, by itself, proof of an appreciating enterprise capability.
By 2030, aim to know which decisions depend on this layer and whether reuse improves their economics. By 2035, a broader autonomous ecosystem could make that capability strategically important. The scenario analysis also considers the future in which interoperability or economics develops more slowly.