Your SOP needs an ontology before an AI agent can run it

Your SOP says that on suspected fraud, the account is frozen and compliance is informed. Every operator on your team knows what that means at your company: outgoing transfers blocked, the card left active, a case reference attached. An AI agent reading the same line knows none of it. What sits between the two is an ontology, a shared definition of the records your procedure touches and of every action allowed on them, written once so that people, apps and AI agents all read the same thing.

Your SOP was written for auditors and trained operators

A regulated SOP describes what must happen and who controls it, at the level an auditor reads. It was never meant to describe each action at the level a machine executes. Operators fill that gap with habit, with what they know about the customer, and with the judgment to escalate when something looks off.

An AI agent has none of that. It reads "freeze" at face value and applies the same reading to every case, so one misunderstanding becomes the way the procedure runs.

AI agents, your apps and your team now act on the same records

For years, the only one acting on a customer's account from the backoffice was a trained operator clicking a button. That operator now shares the records with three newcomers, and all three want to write.

In a June 2026 Retool survey of about 300 CTOs, CISOs and CIOs, 93% said they were concerned about vibe-coded apps running in production, and only 8% described their governance as strong (Retool via Business Wire, 2026). Each of these newcomers writes its own version of an action, unless that action already exists and is defined once.

What an ontology is, and what it does for your SOP

Microsoft describes an ontology as "a shared, machine-understandable representation of your business", covering the types of records, their relationships and the rules that apply to them (Microsoft, Fabric IQ). Palantir adds the part that matters for an SOP. If the data elements are "the nouns" of the enterprise, "then the actions can be considered 'the verbs'", and the ontology is "designed to represent the decisions in an enterprise, not simply the data" (Palantir).

In a regulated backoffice, the ontology does three jobs. It gives the AI agent a closed list of actions to choose from, each described with when to use it. It gives your systems a contract to enforce, whatever the AI agent decided. It lets an auditor follow each step of the SOP to the exact action that ran, with its inputs and approvals. Your SOP stays the document your auditors read, and the ontology becomes the version an AI agent can execute.

Most ontologies stop at the data

Most initiatives model what things are, and few define what you may do to them.

ServiceNow, which opened its "system of action" to any AI agent in May 2026, runs each action "identity-verified, permission-scoped" and traceable through audit records.

Palantir and ServiceNow prove the model works. Neither ships the actions of regulated money operations, such as freezing an account, refunding a card payment or requesting a KYC document, and that part of the ontology is yours to write.

An AI agent that guesses an action repeats the guess on every case

When an action has no definition, the AI agent fills the gap itself. Grama Chethan measured this on tool calls across six industrial configurations: free-form parameters produced invented domain identifiers 43% of the time, constrained parameters 0% (Chethan, 2026). Sergiy Lunyakin found that retrieval grounded in defined business entities pulled no evidence from outside the question's scope, against 16.8% and 20.2% for keyword-based methods (Lunyakin, 2026).

Both are recent preprints on small samples. Lunyakin draws the conclusion that matters for a regulated team: "auditability, not accuracy, is the axis on which ontology-grounded retrieval earns its cost." In a freeze request, an invented identifier means the wrong account frozen.

An action needs a meaning and a contract

Each action in the ontology has two halves: a meaning that tells an AI agent when to choose it, and a contract that makes it run the same way every time. Here are both for freezing an account.

Your ops team writes the meaning, in the first two rows. Anthropic's engineers found that "even small refinements to tool descriptions can yield dramatic improvements" in how AI agents pick tools (Anthropic, 2025). Your tech team writes the contract in code, where an AI agent cannot talk its way around it, and what "freeze" does inside your core banking system stays yours.

One definition per action, shared by every SOP and every tool

A definition only protects you if it exists once. When the fraud SOP freezes transfers and the card while the AML SOP freezes transfers only, your company has two meanings for one word, one per AI agent that reads them.

Treat actions like an API. The second SOP calls the accounts.freeze the first one defined, and any change ships as a new version, so a workflow built on version two never runs version three by accident. → Know Your Agent: the missing Compliance Layer

The number of definitions you have for "freeze an account" is the quickest test of whether your SOPs are ready for an AI agent, and the only right answer is one.

The Forest backoffice defines each action once for every surface

In the Forest backoffice, your tech team declares each action once in code, with its description, typed form and business logic, on top of the data model Forest reads from your databases. That is the action layer of your ontology, and every surface calls it under the same roles.

Apps your ops team builds go through the Forest Gateway, where "every call reaches your back-end as the person who made it" (Forest docs, Gateway). In the Zendesk app, AI-suggested actions that match no real action are dropped before they reach a button (Forest docs, Zendesk app). An AI agent connected through the Forest MCP server acts with the permissions of its Forest user, and its actions are logged like any action taken in the UI (Forest docs, MCP server).

Forest does not ship ready-made actions for your core database, since what freezing an account means at your company cannot be inferred from a schema. Start with the SOP your team runs most often, list its actions, and count how many definitions each one has today.

Sources

• Retool, Business Wire press release, June 17, 2026

• Capco, Educating AI in banking with ontologies (FIBO), April 8, 2026

• BIAN Service Landscape

• Tableau, The agentic future demands an open semantic layer (OSI), September 23, 2025

• Microsoft, Fabric IQ overview

• Akshay Krishnaswamy, Palantir, Connecting AI to decisions with the Palantir Ontology

• ServiceNow, press release, May 5, 2026

• Grama Chethan, The Semantic Training Gap, May 11, 2026

• Sergiy Lunyakin, Auditable by Construction, August 21, 2026

• Anthropic, Writing effective tools for agents, September 11, 2025

• Forest docs: Gateway, Zendesk app, MCP server