The certification is a set of statements about the officer's own knowledge and the officer's own acts. Based on their knowledge, the report contains no untrue statement of a material fact and no material omission, and the financial information fairly presents the issuer's condition and results. They are responsible for establishing and maintaining internal controls; they "have designed such internal controls to ensure that material information relating to the issuer and its consolidated subsidiaries is made known to such officers by others within those entities"; they have evaluated effectiveness and presented their conclusions. And they have disclosed to the auditors and the audit committee all significant deficiencies in the design or operation of internal controls, and any fraud involving people with a significant role in those controls.
The phrase to sit with is "made known to such officers by others within those entities". It was drafted for a world in which the thing that surfaces a problem is a person who noticed. A model that flags an anomaly, drafts a commentary or clears a reconciliation is not "others within those entities", and it will not raise a hand in a meeting. That does not relieve the duty; it relocates the work. The officers still have to have designed controls that bring the material information to them, and where a machine is the first — sometimes the only — thing that sees it, the design has to specify how what the machine saw reaches a human who can escalate. The ethics code catalogues the threats that arise on the other side of that boundary: see technology threats in the finance function.
Rule 13a-14 supplies the sharpness. It puts the certification in the filing, requires each specified officer to sign it, and at paragraph (c) forbids signature by power of attorney or any other form of conferred authority. Nothing about that is delegable, and nothing about it is corporate. Compare the underlying control duty, which the entity owes — see an AI tool inside a control you must evaluate. The regime deliberately splits the two: a company maintains and evaluates; two named people put their names to the result. That split is why the sub-certification chain matters so much, and why a chain that never asks about AI systems leaves the signatory holding a representation nobody made to them.
The mismatch this collection is built around lands here in its most personal form. A certification is a statement made on a date about a quarter; a model behaves differently in the second half of that quarter than in the first, for reasons that may be a supplier's. What an officer can honestly certify is not that the model was right, which nobody can know, but that the controls around it were designed to bring what mattered to their attention and that they evaluated whether those controls worked. Being able to say that with a straight face requires having asked questions during the period, not at the point of signature. The competence to stay answerable for figures a model helped produce is the substance of this credential, and section 302(a)(4)(B) is where the law puts it.