To find out whether your business is ready for agents, score four things: your data, your governance, your discoverability, and your operational readiness. You can run this audit yourself in an afternoon, and you should run it twice, once for the agents you deploy inside the business, and once for the customer agents that will read your public record and act on it without asking you first. The finding that sits under the whole exercise is a plain one: agent-readiness is not a product you buy. It is a property of your record, deterministic, single-source, attributable. And it compounds, because the discipline pays off even if no agent ever visits.
The four dimensions below are not an Origin Pi invention. They line up almost one-to-one with the four pillars of Singapore's Model AI Governance Framework for Agentic AI, which the Infocomm Media Development Authority launched on 22 January 2026. That framework is voluntary guidance, not law. It is still the most useful local backbone an SME operator has, because it was written by the regulator and it says the same thing the payment rails and the search engines are now saying independently.
The audit is a data-hygiene score, not an AI project
Start by dropping the word "AI" from the exercise. Almost everything that determines your score is settled before a model is ever involved. Can you point to one current record of your products, prices, policies, and process? Does a consequential action pause for a human before it commits? Can a machine read a clean answer about your business? Does every change trace to a named person? Those are questions about how your operation is wired, not about which model you picked.
That reframing has a useful parallel. Systematising your business record for agents is the same move as going from a shoebox of receipts to a proper ledger. You do not build a ledger because an auditor is at the door. You build it because a single, current, attributable record removes operational ambiguity every day: fewer internal contradictions, faster answers from staff, cleaner handover. The machine-readability that lets a customer's agent transact correctly is a by-product of that discipline, not the point of it. That is why the score is really a maturity read on the operation.
The scored checklist
Each dimension breaks into three concrete checks. Score the table-stakes items pass or fail, and score the graded items on the scale noted. A dimension with any table-stakes failure is not ready, whatever the total.
| Dimension | Check | What good looks like | The common gap | Why it matters to an agent |
|---|---|---|---|---|
| Data | Single source of truth | Product, price, availability, and policy all served from one gateway with identical values on every surface | Price lives in one system, stock in another, policy in a document nobody syncs | The agent quotes whichever copy it reached, so contradictory copies mean contradictory answers |
| Data | State-drift latency | A change in the database appears in the public record in under a minute | Structured data was set months ago and no longer matches the live figure | A stale record sells the wrong price to every machine reader at once |
| Data | Contract completeness | The machine-readable record exposes price, tax, stock, and the policy version in force | The record carries a name and address and little else | Missing fields become guessed fields, and a guessing agent is a liability |
| Governance | Confirm step | Any outbound action that carries liability (a refund, an order, a commitment) pauses for human approval | The agent is wired to act end-to-end with no checkpoint | A non-deterministic actor committing your business at machine speed with no pause is an uninsured risk |
| Governance | Row-level access control | A customer-agent token reaches only that customer's data, tested with a second tenant's token | One broad key opens every record | Without scoped access, one agent can read across every account it should never see |
| Governance | Immutable audit trail | Agent ID, prompt ID, timestamp, and result recorded and retained at least 90 days | Actions happen with no durable record of who or why | You cannot contest or explain a decision you cannot reconstruct |
| Discoverability | Server-rendered truth | JSON-LD generated from the live database at request time, not a static file or client-only render | Static markup that drifts, or data that only appears after JavaScript runs | Machine readers need the answer present and current in the served response |
| Discoverability | Schema validity | Zero errors and zero warnings on key pages, and the markup matches what a human sees | Markup that describes content not visible on the page | Google's policy requires markup to be a true representation of the page, and mismatches break trust |
| Discoverability | Machine-readable terms | A record of allowed agent actions and rate limits, kept separate from your volatile facts | Volatile facts like price or stock stuffed into a hand-written text file | A hand-written file goes stale and is ignored for ranking, so never let it carry facts that change |
| Operational readiness | Human override rate | Over 30 days, more than one in ten confirmations actually modified or rejected the agent's proposal | Every confirmation is approved without reading | A confirm step that is always clicked approve is theatre, not governance |
| Operational readiness | Attribution per action | Every state change traces to an agent ID, a session, and a named human reviewer | Actions land with no owner attached | Accountability that survives after the fact requires a name on the record |
| Operational readiness | Anomaly review | Only deviating actions (a refund above a threshold, a price drop beyond a percentage) surface for a human | Everything or nothing is reviewed, so attention burns out | Targeted review keeps human judgment where it matters and prevents alert fatigue |
Dimension one and two: data and governance are the same discipline
The data checks and the governance checks are two ends of one problem. Data asks whether there is a single, current, attributable record. Governance asks whether an actor can be trusted to act against it. State drift is the quiet killer on the data side. Static structured data that says one price while the live database says another means you are selling at the wrong price to every agent that reads you. The fix is operational, not notational: one source of truth, and a record generated from it, re-synced whenever a fact changes.
On the governance side, the protocol designers already wrote your checklist for you. The Model Context Protocol specification makes consent normative: a user must explicitly consent before an agent invokes any tool, tool descriptions are untrusted unless they come from a trusted server, token passthrough is forbidden, and authorization is built on OAuth 2.1 with audience-bound tokens. That is the confirm step and row-level access control, written into the standard rather than bolted on. Under GDPR Article 22, where a decision is based on contract necessity or explicit consent, a person has the right to obtain human intervention, to express a point of view, and to contest a solely automated decision that produces legal or similarly significant effects. The confirm step is that safeguard implemented as a runtime control, and the audit trail is what lets a decision be contested after the fact.
One honest caveat belongs inside governance, not in a separate box. A flawless record is a double-edged asset. The same clean structured data that lets a customer's agent transact correctly also makes it easier for a competitor's agent to scrape your pricing and undercut you, and a poisoned product description can carry a prompt injection into a customer's reading agent. Readiness is therefore a business decision about what you expose, not only a technical task. Scope, access control, and the audit trail are where you make that decision deliberately.
Dimension three: discoverability, and why structured data beats a hand-written file
Discoverability is where most SMEs will quietly fail, and it is also where the market's honesty matters. Be precise about what structured data does and does not buy you. Google states on the record, in its AI features documentation updated 10 December 2025, that "you don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." So do not sell schema.org as an AI-ranking trick. It is hygiene for correctness and for rich results, and that is reason enough to get it right.
For the facts agents get wrong most often, the boring standard already has typed fields. OpeningHoursSpecification carries dayOfWeek, opens, closes, validFrom, and validThrough, and Offer.price carries an exact figure. The rule that makes these trustworthy is Google's structured-data policy: the markup must be a true representation of the visible page. That is why a hand-written text file is the wrong instrument for anything that changes. Its adoption is low, Google ignores it for ranking, and a file you edit by hand drifts the moment a price moves. Use it, if at all, to point an agent at documentation. Never put a price or a stock count in it. The correct home for a volatile fact is server-rendered structured data generated from the live database, so the served answer is current by construction.
Dimension four: operational readiness is whether the human is real
The fourth dimension scores the one thing you cannot fake with markup: whether the reviewer is a real reviewer or a rubber stamp. The IMDA framework's second pillar asks for meaningful human accountability through significant approval checkpoints, and the failure mode it guards against is automation bias, the drift where a human approves whatever the machine proposes. The practical measure is the human override rate over 30 days. If none of your confirmations ever modified or rejected the agent's proposal, alert fatigue has already hollowed out the control, and your true score is zero no matter what the interface shows. Anomaly-scoring is the safety net, surfacing only deviating actions so attention lands where it belongs. And every state change must trace to an agent, a session, and a named human, so attribution survives the moment.
Where Origin Pi stands
Our read is that the signal is not any single announcement. It is convergence. Singapore's IMDA framework, Google's Universal Commerce Protocol announced 11 January 2026, the Agent Payments Protocol it is compatible with, and the card networks behind them all ship the same primitives: bounded authority, a confirm step before a liability-bearing action, an audit trail, and the business staying merchant of record. When the standards body, the payment rails, and the search engine arrive at the same posture without coordinating, that is not a trend. It is structure.
Two shortcuts the hype invites, we refuse. Schema.org is hygiene, not a ranking trick, because Google says so on the record. And the audit is accountability, not a legal shield. The Moffatt v. Air Canada decision covers an operator's own chatbot. When a customer's own agent acts on your valid data, the liability is legally untested, so no checklist can promise it makes you agent-proof. The regulatory runway, meanwhile, is longer than the earlier reading suggested: the EU AI Act's high-risk obligations were deferred to 2 December 2027 and 2 August 2028 under the simplification package agreed in May 2026. Its general-purpose AI duties are already in force from August 2025, and its transparency obligations apply from August 2026. More runway is not less obligation.
That is why we treat agent-readiness as a property of your record rather than a tool you buy. Properties compound. Tools depreciate. The checklist above is deliberately self-contained and scorable so you can run it unaided. Cerebrum, the governed record we are building with a confirm step, row-level access control, and an immutable audit trail, is the system that operationalises it. It is in development and coming soon. The honest position is that you do not need us to start: score yourself against the table, fix the table-stakes failures first, and the record you build pays off whether or not an agent ever arrives.
