← Back to blog

November 18, 2026

BNY and Microsoft used AI agents to negotiate the terms of a U.S. Treasury bill

BNY and Microsoft used AI agents to negotiate the terms of a U.S. Treasury bill in a pilot announced Tuesday. The agents exchanged offers within preset limits, while traders reviewed and approved the proposed terms and the trade was executed through existing systems. It is a concrete example of AI working across company boundaries with people retaining the final decision, although the companies have not published comparative performance results. BNY announcement, republished by MarketScreener

Executive Snapshot

  • OpenAI introduced ChatGPT for Sales, bringing account research, meeting preparation and follow-up into connected workflows.

  • Microsoft and Salesforce introduced a public-preview Copilot integration with more than 38 prebuilt sales skills, while Accenture is piloting the integrations.

  • OpenAI released GPT-6.2 Astra and Sol; Microsoft added its own coding and decision models at Ignite.

  • BNY’s Treasury pilot shows how agent-to-agent negotiation can work with defined limits and human approval.

  • BCG argues that redesigning wholesale credit around connected work could materially shorten decisions, an estimate that needs validation in each bank.

Top Developments

New models put professional work and cost under the microscope

OpenAI released GPT-6.2 Astra and Sol for professional outputs, coding and computer use. Both are available to eligible Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the API; workspace administration can affect access. OpenAI’s performance comparisons are its own evaluations. The practical test for a sales team is whether a model produces a more accurate, usable account brief at an acceptable total cost, including review and correction.

At Ignite, Microsoft made Code 2 and Code 2 Flash available in GitHub Copilot and Foundry, and announced general availability of Microsoft-Decision-1.1 for tasks such as classification and routing. The growing menu makes task-level evaluation more useful than choosing one model for every job.

Agent visibility becomes a shared enterprise problem

Microsoft launched an AI Transparency Initiative with more than 40 organizations, including Mastercard and Volvo Cars, to advance approaches to transparency, security and governance across providers. Its Agent 365 Business Analytics is in Frontier preview and is intended to show usage and spending. This is a product and standards effort, not proof that cross-vendor auditability is solved. For a sales agent, the useful record includes sources consulted, permissions used, proposed changes, approvals and errors.

AI for Sales and Field Execution

The account brief moves closer to the next action

ChatGPT for Sales combines OpenAI’s sales plugin with more than 30 new or expanded partner plugins for work including account selection, call preparation, proposals and follow-up. For a field representative, the opportunity is a single preparation flow that connects account history, open commitments and the next meeting. Start with approved sources and require review before customer messages or CRM changes. Measure verified facts, preparation time and follow-through, not just briefs generated.

The YouEx.ai ChatGPT app brings CRM work into the same conversation where sellers are already researching accounts and preparing for meetings. After connecting their YouEx.ai account, sellers can ask questions about their pipeline, look up or update records, and create follow-up actions in plain language. For field teams, that means an opportunity to keep customer context and next steps current without breaking the flow of the day to navigate a separate system.

Salesforce context arrives inside Copilot

Microsoft and Salesforce’s Copilotforce announcement includes Salesforce in Copilot in public preview, with more than 38 prebuilt sales skills. Accenture is a pilot customer validating the integrations in its sales organization; no pilot outcome figures were published in the announcement. For manufacturing or MedTech sellers, test whether the right account, meeting and service context arrives with source visibility and existing access controls before expanding the workflow.

The 38 skills fall into these groups:

  • Account intelligence and research: Account Context, Research Prospect, Account Plan, Account Tiering, Stakeholder Map, Expansion Whitespace, Executive Movement.

  • Deals and opportunities: Deal Review, Deal Advance Gap, Deal Signals, Deal Slip Scenario, Close Plan, Handle Objection.

  • Pipeline and forecasting: Pipeline Review, Team Pipeline, Forecast Narrative, Salesforce Hygiene Check, Weekly Wrap, Win-Loss Review.

  • Customer success: Customer Health, Renewal Radar, Customer Voice.

  • Leads and prospecting: Lead Triage, Assign Target to SDR, Draft Outreach.

  • Meetings and calls: Call Prep, Call Follow-Up, Summarize Conversation Transcript, Schedule Meeting, Calendar Events, Event Invite List.

  • Daily work: Daily Briefing, Inbox Sweep, Log Activity, Query Activity History, Rep Context, Find Internal Answer.

  • Personalization: Voice Profile.

Wednesday Focus: Financial Services

A trade proposal with a human signature

In the BNY and Microsoft pilot, BNY’s protocol verified participants, applied predefined controls and reconciled the agreed terms against a central ledger. The agents could not execute the trade themselves. That architecture suggests a useful pattern for commercial pricing or complex renewals: define negotiable boundaries, preserve a shared record, and route the proposed agreement to an accountable person. One pilot does not establish scale, cost savings or suitability for other transactions. Source

Credit decisions depend on the handoffs around the model

In a new BCG analysis, the firm estimates that end-to-end agentic redesign could reduce time to wholesale credit decisions by 40% to 60% compared with traditional automation. This is a consulting estimate, not a measured result from a named bank in the cited summary. BCG calls for traceable facts, deterministic calculations, guardrails and human judgment. The parallel in field sales is straightforward: a faster summary helps, but an up-to-date account file and a clear exception owner help the entire customer process move.

What to Watch

  • Whether sales plugins can reliably cite the right customer record, respect permissions and produce accepted updates in live deployments.

  • Whether BNY and Microsoft publish repeatability, exception rates, negotiation times and evidence from additional counterparties.

  • Whether emerging transparency standards yield portable, inspectable records across vendors and agents.

  • How GPT-6.2 and Microsoft’s models perform on representative customer work when quality, latency and human review are measured together.

Source List

BNY and Microsoft Treasury pilot · OpenAI GPT-6.2 · ChatGPT for Sales · Microsoft and Salesforce · Microsoft Ignite models · AI Transparency Initiative · Agent 365 · BCG wholesale credit