The reasoning must be logged in real time: why AI agents' after-the-fact explanations fall short
Within a single week in mid-October 2026, a series of signals emerged that accountability for AI agents has moved from theoretical concern to practical governance problem: IBM presented survey data showing that three quarters of surveyed EMEA leaders expect an agent to complete a full business task and reach a decision without human intervention by 2027, the AICPA released a 51-page guidance document on AI in financial reporting and auditing, and trade media described the control gap in agentic payments. Against this landscape stands a concrete framework proposal — "auditable intelligence," put forward by practitioner Narasimha Reddy Annapareddy — which argues that agent accountability must be solved as an engineering problem, not as a legal or ethical discourse. This analysis works through the three technical requirements in the framework, sets them against the broader picture from that same week, and points to what remains unresolved.
The week that put accountability on the agenda
At IBM Think on Tour EMEA on 14 October 2026, the company presented research with two striking figures: Three quarters of the surveyed leaders across EMEA expected that by 2027 an agent would carry out a complete business task and make a decision without human intervention. At the same time, 80 percent said that at least one AI-related problem had already reached the board or senior leadership (TechTarget).
The latter figure may be the most significant: AI issues are no longer something handled in the IT department. They land where senior leadership must answer for them. It is worth noting that the figures come from IBM's own research, presented at an IBM event, and that methodology or sample size is not given in the coverage. They should therefore be read as the company's picture of market sentiment, not as an independent measurement.
The same week, on 13 October, the AICPA released a 51-page "practice aid" on AI, developed over several years, with guidance for two main areas: how businesses use AI in financial reporting and controls, and how auditors use AI in the audit process (CFO Dive via Yahoo Finance). And on 16 October, PYMNTS reported on a control gap in agentic payments, with analysis from Troy Mann, SVP of product management at FIS Global (PYMNTS).
The same week, CNBC also reported on a third group affected by the same question: the independent safety evaluators. More on that further down.
The simultaneity is not coincidental. Agent deployments are now moving from pilot projects into actual workflows in finance, procurement and compliance — precisely the areas where documentation, traceability and audit trails are already mandated by law or expected by industry. That is why the accountability question is surfacing simultaneously in the banking sector, the audit profession and among software vendors.
The framework: "auditable intelligence"
The most concrete answer to the question came in a TechTimes profile on 16 October 2026 of Narasimha Reddy Annapareddy, who has concentrated much of his recent work around a principle he calls "auditable intelligence." The requirement is that every consequential action a company's AI agent takes must be tied to a verifiable data source, a logged chain of reasoning, and a decision point that human reviewers can reach — and where they can inspect, contest and reverse the action (TechTimes).
Annapareddy breaks this down into three technical requirements, which according to the profile must each be fulfilled "by design" rather than through policy:
1. Source traceability. Every conclusion an agent reaches must be traceable back to the specific data source it drew on — a record in the ERP system, a line in a policy document, a transaction in the historical log. Without this, every agent decision is a black box that can only be discussed in abstract terms.
2. Reasoning documentation at runtime. The reasoning must be documented as it happens — during inference, not reconstructed after the fact. This is perhaps the framework's most important technical point: An after-the-fact explanation of why a model did something is neither credible nor complete, because the model's actual runtime context cannot be recreated retrospectively.
3. Human override points. Reviewers must have access to decision points where they can inspect, contest and reverse the agent's actions. Traceability without the ability to act is merely surveillance.
According to the profile, Annapareddy frames the accountability problem in enterprise AI not primarily as a legal or ethical question, although it has consequences for both — but as an engineering problem: How do you design an AI agent system so that every consequential action is logged, attributed and available for review in a form that satisfies the organization's governance requirements? (TechTimes).
It is important to be precise here: This framework is known through a profile article, not through a technical paper, a code project or independent verification. It is Annapareddy's claim about how accountability should be built, not an established or tested method. But as a starting point for debate, it is useful precisely because it converts an abstract ethical discussion into three concrete requirements that can be evaluated.
Why "by design, not after the fact"
The central rationale in the framework is that documentation must be constructed at runtime, not reconstructed afterwards. After-the-fact explanation is weak for two reasons: It cannot be validated against the actual runtime context (what data the model saw, what instructions it received, what intermediate steps it took), and it is written by the actor with an interest in presenting the action in a favorable light.
The same message came from an unexpected corner that same week. IBM executives argued at the Think event that governance must be built in from the start, not added afterwards. The company also introduced Sovereign Core, an open source platform which IBM itself says will give organizations greater control over the deployment and operation of AI environments (TechTarget). The same caveat applies here: This is IBM's claim about its own platform, presented at IBM's event.
That a major vendor and an independent practitioner independently arrive at the formulation "by design" is a signal that the industry has recognized something: After-the-fact compliance — where you write policies meant to cover systems that were not built for them — does not hold when agents act autonomously and at machine speed.
Same problem, other sectors
The same concerns are surfacing in other industries, with different terminology.
Payments. The PYMNTS report describes a structural problem in agentic payments: Merchants know what the customer asked for. AI platforms know the instructions given to the agent. Banks know the account, payment credential and transaction history. These signals do not necessarily travel together, and much of the information needed to establish the agent's legitimacy lies outside the issuer's traditional line of sight — according to Troy Mann at FIS Global (PYMNTS). This is in practice the same problem as Annapareddy's source traceability, translated to financial infrastructure: You need a way to authenticate a machine's authority, and no universal mechanism for doing so exists today.
Audit. The AICPA's 51-page guidance shows that the audit profession recognizes that AI in financial reporting and in the audit process needs standardized guidance — not ad hoc judgments (CFO Dive via Yahoo Finance). Several years in development suggest the profession has been working on the problem long before agentic AI became widely available; the documents are thus not a reaction to an acute crisis, but preparation for a profession expecting extensive AI use.
Independent evaluators. CNBC reported on 11 October that without federal regulatory pressure, small nonprofit third-party organizations — METR, Apollo Research and Transluce — have gained outsized importance as de facto safety evaluators. Anthropic chief Dario Amodei has promised to build independent evaluators into the company, something OpenAI chief Sam Altman quickly endorsed (CNBC). Suresh Venkatasubramanian at Brown University has raised concerns about the funding of these organizations — a question that becomes more pressing when they function as the industry's reputational infrastructure.
The pattern across these cases is striking: Regardless of sector and terminology, everyone converges on the same type of solution — traceability, documentation at runtime and independent control points. None of the actors proposes that policy documents alone are sufficient.
What remains unresolved
It is important to be honest about the evidence base here.
Annapareddy's framework is known exclusively through a profile article. There is no technical paper, code repository or independent confirmation of how the three requirements are implemented in practice, what costs or performance impacts they entail, or how well they work in production environments. The presentation in the profile sets out Annapareddy's own claims, and should be read as such.
The IBM figures are the company's own research, presented at IBM's marketing event, with no stated methodology or sample size. They probably describe a real sentiment among the leaders surveyed, but they say nothing about how quickly agent autonomy will actually be deployed in practice.
For third-party evaluators, the funding, access and reporting structures are unresolved. Without a stable foundation, the industry risks building safety controls on organizations dependent on volunteer effort and project funding.
And for the payments sector, no concrete mechanism has yet been described for how the signals from merchants, AI platforms and banks will be coordinated — only that the problem exists.
From policy to infrastructure
The common denominator across IBM's operating model message, the AICPA's guidance, the payments sector's control gap and the independent evaluators is the same: Accountability for AI agents is shifting from policy documents to infrastructure. The question is no longer just "which rules should apply?" but "which properties must the systems have for the rules to be enforceable at all?"
Annapareddy's three requirements — source traceability, reasoning logging at inference and available override points — give a concrete starting point for organizations considering agent deployments. But they are a framework proposal, not a standard, and the evidence base for both the framework and the surrounding measures is still thin. The next year will show whether the industry manages to convert these principles into testable implementations — or whether accountability remains, in essence, well-formulated intentions.