When the agent is wrong, whose name is on the mistake?
bot blamenoun — the accountability gap that opens when an AI agent acts and no human owns the outcome. The firm blames the bot. The vendor blames the prompt. The vendor's vendor blames the model. The client is still out the money.
Coined July 2026, as financial institutions began moving AI agents from suggesting decisions to executing them.
Why this is a financial services problem first
Agents are no longer drafting; they are doing. In wealth and asset management they are rebalancing portfolios, preparing suitability rationales, screening onboarding, and touching money movement. Every one of those actions used to have a person attached to it, and that attachment was the control environment.
Regulation has not moved. Fiduciary duty, suitability, best execution, and fair-lending obligations were never transferable to software, and no examiner will accept "the model decided" as a response. The obligation stays with the firm; only the decision left.
So the exposure is asymmetric. The productivity gain from an agent is incremental and daily. The failure is rare, public, and expensive, and when it lands, the first question from the client, the regulator, and the press is the same one: who owned this? Most deployments today cannot answer it in one name.
The Bot Blame Framework
Five clauses every agent deployment should have a name against, before go-live.
Named ownership
Every agent has exactly one accountable human owner, with the authority and the means to stop it. A committee is not a name. If the owner cannot halt the agent inside five minutes, the ownership is decorative.
Decision boundaries
Written limits on what the agent may decide alone, expressed in the units that matter: dollars, client count, and reversibility. Reversible actions can be delegated broadly; irreversible ones narrowly or not at all.
Evidence by default
Every consequential action leaves a replayable record: what the agent saw, what it decided, what it did, and which human reviewed what. If the audit trail is reconstructed after the incident, it is not an audit trail.
Fault mapped before launch
Liability allocated across firm, vendor, and model provider in the contract, not in the post-mortem. If the agreement is silent on agent error, the firm has silently accepted all of it.
Rehearsed failure
The bad day is practiced before it happens: kill switch, client communication, remediation, and regulator notification, run as a drill with the named owner in the chair. Firms rehearse fire evacuations; an agent incident deserves the same.
Blame is a design decision. Make it before the agent does.
About
Tarun Gulati has spent three decades leading commercial programs across North American banking, wealth and asset management, with over $900M in contracted value across the institutions where these agent decisions are now being made. He writes here about accountability in agentic AI.
Conversation welcome: [email protected]