Gartner's AI Central Bank is a control mechanism for enterprise AI, but only if every decision cites its evidence

On 19 October this year, Gartner opened its IT Symposium in Orlando with two numbers that describe most large enterprises right now.

In its Q2 2026 survey of more than 2,000 CIOs and technology executives, 71% expect to deploy agentic AI within the next 12 months, and 21% have governance for it in place.

The keynote added a third figure, 86% of CIOs say AI risk is growing faster than AI value in their organization.

Gartner’s answer is an “AI Central Bank”, a single control mechanism for all the AI in an enterprise.

In the words of Daryl Plummer, Distinguished VP Analyst and Gartner Fellow, it “creates accountability, governs technology, manages economics, and creates shared intelligence.”

Plummer also said what it will not do, which is “eliminate risk” or “drive value.” Read that as a definition of a control mechanism.

It accounts for the value and the risk of the AI it governs; the value and the risk themselves come from the decisions the AI makes. That is the part the keynote leaves out.

A control mechanism can only control what it can see, and nowhere in Gartner’s description does it say what an automated decision has to show so the mechanism can govern it.

This piece answers that. It explains what each decision has to cite, why monitoring, logs and audits do not supply it, and how to check whether your own agents do before the mechanism is built.

A control mechanism needs each decision to cite its evidence, not just to log its activity

Two kinds of information come out of an AI agent, and only one of them can be governed.

The first is an activity log. It says which agent acted, which tool it called, when, and what came back. Almost every agent in production produces this today.

The second is a decision that cites its evidence. For a claim, a loan, a KYC file or a supplier invoice, it names the clause in the policy or contract that decided the case, the version of that clause in force that day, the document or data point that satisfies each condition, whether the case went to a person and why, and what has changed since.

Think of these as the citations under a decision, in the same way a footnote shows where a statement came from.

The difference decides whether Gartner’s control mechanism has anything to work with. An activity log can tell you that a claim was denied at 14:02 by a given agent.

It cannot tell you whether the same claim would be denied tomorrow, or what to say to the customer who asks why.

The gap Gartner measured, 71% of CIOs deploying agents against 21% able to govern them, is a gap in what each decision cites, not a gap in committees.

For reversible work, such as sorting incoming documents, ranking leads or drafting a reply a person reads first, an activity log is enough, because a mistake is undone at no cost to anyone else.

For decisions that move money or change someone’s rights, a paid or denied claim, an approved loan, a cleared KYC file, a released supplier payment, the citations are what the auditor, the regulator and the customer will ask for.

Why monitoring, logs and audits after the fact do not solve the problem

Most enterprises are already governing agents, just not at the decision itself in a deterministic, evidence-backed way.

Three substitutes carry most of the load today, and each one fails for a reason somebody else has already documented.

Monitoring catches the problem after the decision has taken effect, and the fix is to switch the system off.

EY surveyed 202 senior AI executives at US companies with at least $1 billion in revenue in May and June 2026. Among those that had run a formal AI assurance review, 29% went on to pause a quarter or more of their AI systems and 25% stopped a quarter or more outright.

Monitoring protects the enterprise, but only by removing the AI it reviewed.

Logs only prove activity, so auditors have given up explaining individual AI decisions.

ISACA, the professional association for IT auditors, published guidance for auditing agentic workflows on 11 September that tells auditors to test averages over a period instead of individual transactions, because “output variance is a property of the system rather than an anomaly that sampling can isolate.”

In plain terms, the profession that signs off on your controls has accepted that a single automated decision may not be explainable, and now audits the pattern instead.

The same guidance states that “the absence of adequate logging is itself a finding.”

Asking the model why it decided does not produce a citation.

OpenAI’s system card for GPT-6 Astra, published on 3 September, states that the model’s “monitorability has decreased relative to GPT-5.6 Sol” and that it is “less likely to include incriminating information in its CoT”, the chain of thought it writes while reasoning.

OpenAI notes that these findings come largely from adversarial tests, but the practical point holds. A written explanation the vendor itself does not treat as reliable cannot serve as evidence.

All three give the control mechanism the same thing, a description of activity. None gives it the reason.

Five citations every automated decision should carry

Check the five citations against the three agents closest to production.

Take one recent decision each agent made and ask whether the system can produce all five without a person reconstructing them.

Suppose the agent decides whether a supplier invoice matches its contract. The check passes if the system can name the contract clause, the version in force on the invoice date, the line items that satisfy it, and why the case did or did not go to a person.

An agent that produces all five citations can run beside the team today, with its decisions compared against theirs before it is given authority.

An agent that produces none stays on reversible work until it does. No agent is switched off, and no control mechanism is built on decisions that cannot show their evidence.

How Otera makes every decision carry its citations

The first citation decides the rest. A decision can only cite the clause that decided it if it was made by logic that knows that clause, which is not how a language model decides.

Otera’s belief is simple. No probabilistic system, human or AI, can run a regulated operation. AI formalizes rules into deterministic enforceable code, code decides at scale.

In practice the model does the reading. It takes your policies, procedures and agreements and formalizes them into decision logic that a business reviewer checks before go-live.

That logic then makes each decision deterministically, from evidence to outcome along a path a reviewer can read, and attaches the five citations as it does so. Uncertain cases go to a person before anything takes effect.

A control mechanism like Gartner’s AI Central Bank then queries the citations rather than reconstructing them, which is our reading of what the proposal needs rather than Gartner’s own.

HFS Research’s case study of Otera’s work with Allianz Partners describes the same principle from the outside, “ensuring every action can be explained, traced, and reversed if required.”

Gartner has given CIOs a control mechanism at the moment most of them are about to need one. It will govern exactly what the decisions bring.

Decisions that bring an activity log give the enterprise a clear view of actions it cannot explain.

Decisions that bring their five citations give it a control mechanism that can answer the auditor, the regulator and the customer in one query, and the time to find out which kind your agents produce is before the first one is connected.