In London in 1832, clerks from thirty-one competing banks gathered each afternoon on Lombard Avenue to settle accounts. Every financial institution had already mastered the fundamentals — monitoring debits, crediting accounts, settling balances. The unsolved drawback was architectural: How do you allow each day transactions between opponents who essentially don’t belief one another?
What they invented collectively was a “belief structure” that allowed opponents to transact reliably, on daily basis, with out a government imposing the principles. The ensuing London Clearing Home (LCH) labored as a result of each participant had a registered id, a fame that traveled with them, and a shared understanding of what habits would get them expelled from the community.
We’re at that very same second with AI brokers.
The technical protocols for agent-to-agent communication are being constructed. The belief structure has not been, and up to date analysis from our workforce at Salesforce AI Analysis makes the excessive stakes clear. In a examine of three.5 million feedback throughout 78,000 AI brokers interacting on an agent-only social platform, our researchers discovered that with out coordination mechanisms, even succesful brokers speak previous one another. Sixty-five % of agent feedback added nothing to the posts they replied to; spam and off-topic noise dominated. The examine calls this “interplay theater,” exercise that appears like dialogue however carries virtually no alternate.
That discovering has direct implications for enterprise leaders. As autonomous brokers start to barter on behalf of organizations throughout industries and borders, governance, authorized, and moral guardrails have to be designed earlier than the transactions change into consequential.
Under are the 5 issues each enterprise chief needs to be desirous about proper now.
1. Requirements increase guidelines
Guidelines are deterministic. “Don’t disclose buyer knowledge” is a rule, and an AI agent can comply with it. Enterprise negotiation runs on requirements: contextual, interpretive, value-laden judgments about when to carry agency, when to concede, and when strategic positioning crosses into manipulation.
A long time of AI growth centered on rule-following. We are actually asking these programs to navigate requirements, and that requires a essentially completely different structure, together with governance frameworks that consider how an agent determined, not simply what it determined.
2. Id and fame should journey with the agent
Nameless participation within the Clearing Home was unattainable, and the identical precept applies right here. Cross-organizational negotiation requires verifiable id and fame that accumulate throughout 1000’s of interactions — a baseline requirement for any entity performing autonomously on behalf of a principal. Recreation concept and business legislation converge on why: good-faith habits between self-interested events will depend on the expectation that right now’s conduct shapes tomorrow’s alternatives, and on the evidentiary document that makes enforcement attainable when it doesn’t. Strip away that persistence and each interplay begins from zero, with no solution to reward consistency or maintain unhealthy actors accountable.
As a result of fame is cumulative and non-transferable, an agent’s credibility in the end derives from its principal — which supplies organizations with established belief, regulatory monitor information, and institutional accountability an actual head begin over technically succesful newcomers. In 2025, Salesforce AI Analysis pioneered the Agent Playing cards idea: normal metadata speaking an agent’s capabilities, limitations, compliance posture, and authority to commit on behalf of its principal. Consider it as a résumé for brokers. Google adopted the idea in its A2A specification. The subsequent frontier is fame infrastructure — a manner for brokers to construct, show, and lose standing over time. As with individuals, the brokers that earn the precise to maintain negotiating would be the ones that show constant at scale.
3. Boundaries scale higher than scripts
The intuition when governing AI is to specify each state of affairs prematurely. That method labored for rule-based automation and breaks down for probabilistic brokers working in complicated enterprise environments. A greater mannequin comes from how we govern human professionals. Surgeons function inside requirements of care, skilled ethics, and oversight mechanisms reasonably than scripts for each attainable affected person presentation. Agent governance should work the identical manner: extensive latitude inside outlined boundaries, with analysis frameworks that assess judgment high quality. The problem multiplies when brokers from completely different organizations, educated on completely different knowledge, should navigate shared requirements collectively.
4. Accountability have to be structured, traceable, and human
When an agent commits to a worth, accepts a contract time period, or escalates a dispute, there may be no ambiguity about who’s accountable. New organizational roles will emerge to fulfill this requirement: AI operations officers and agent managers with authority over deployments and duty when one thing goes improper. Audit trails have to be refined sufficient to outlive authorized scrutiny, capturing how selections have been made, what info was thought-about, and what options have been evaluated. Rising requirements like Agent Playing cards are early examples of what auditability infrastructure appears like in apply. Designing for auditability from the beginning is way simpler than retrofitting it after deployment.
5. Calibrated escalation is the road between automation and legal responsibility
Realizing when to cease will be the most crucial functionality of all. Brokers that escalate too usually defeat the aim of automation. Brokers that by no means escalate change into a legal responsibility. The belief structure should exactly calibrate the brink to consequence. Routine selections ought to move autonomously. Excessive-stakes selections involving regulatory compliance, main contractual commitments, or vital monetary publicity ought to robotically floor to human judgment. These are strategic design decisions: ought to the agent escalate mid-negotiation, solely at ultimate approval, earlier than strolling away/going “pens down” or by way of periodic audits after the very fact? The reply will depend on the area, the chance, and the connection at stake.
The takeaway for leaders
- Id that accumulates fame
- Structured accountability that traces again to a human
- Calibrated escalation that is aware of when to cease
These 5 ideas kind the inspiration of the belief structure that agent-to-agent commerce would require.
The London Bankers’ Clearing Home emerged as a result of thirty-one competing establishments acknowledged that the collective advantage of a trusted system outweighed the aggressive price of constructing one. We’re at that very same second with agent-to-agent communication between corporations right now. The frameworks don’t but exist, the governance requirements haven’t been written, and the authorized structure for machines that negotiate within the shadow of humanity stays largely uncharted territory. The organizations that have interaction with these questions early will form the requirements everybody else complies with later.
The time to interact is now.
This pondering emerges from an unlikely partnership: a chief scientist and a chief authorized officer who found that the toughest issues in agentic AI reside on the intersection of their disciplines. We’re grateful to Adam Earle, Patrick Stokes, Portia Bamiduro, Jacob Lehrbaum, Rob Wolf, Itai Asseo, and Karen Semone for his or her insights and contributions to this work.
This piece adapts pondering from our Fortune op-ed, “The nineteenth century banking drawback that AI hasn’t solved but” (March 2026), distilled right here into 5 sensible takeaways for enterprise leaders governing AI brokers at scale.
