A Reuters round-up published on 13 July 2026 has given the debate about AI agents in banking a new reference point: 51 per cent of surveyed banks are piloting such systems. The figure comes from a KPMG survey conducted in June and is attributed through Peter Torrente, KPMG's US sector leader for banking. It is the hook, though not the interesting part.
The interesting part appears when the three projects Reuters lists are set side by side. Morgan Stanley, Goldman Sachs and Citi are building agents for three entirely different zones: the client conversation, the plumbing behind the trade, and advice in wealth management. Each bank draws the line on autonomy in a different place, and those decisions say more than any adoption rate.
What: three parallel efforts to move AI agents out of the back office and towards client contact
Who: Morgan Stanley (wealth management), Goldman Sachs with Anthropic (settlement and compliance), Citi (wealth advice)
When: Goldman Sachs public since February 2026, Citi Sky announced 22 April 2026, Morgan Stanley testing from summer 2026
Sourcing: Reuters round-up of 13 July 2026, corroborated here by three independent syndications
Open point: the methodology behind the 51 per cent is not publicly available
Morgan Stanley: always available, never in charge
On the account available to Reuters, Morgan Stanley is testing digital assistants from this summer that can engage with clients at any hour. Koren Maranca, the firm's head of artificial intelligence for wealth management, leads the work. The assistants are described as able to “analyze investments, suggest strategies and assist in building portfolios.”
A clause alongside it is rarely quoted: “agents will not have autonomy to make decisions on portfolios”. The firm is therefore building an agent that is reachable around the clock, analyses and proposes, while the authority to act on the portfolio stays expressly out of reach. Availability and decision rights are separated here. That is an architectural choice, and it determines which controls become necessary at all.
A second effort should be kept apart from this one, since the two are easily conflated. Morgan Stanley is giving selected corporate clients of its stock plan administration business access to agents for administrative workflows, with a wider release in 2027. That covers roughly 3,400 corporate clients and has no bearing on the wealth management assistant.
Goldman Sachs: agents behind the trade, not in it
Goldman Sachs has spent around six months working with embedded Anthropic engineers on autonomous agents built on the Claude model, according to a CNBC report of 6 February 2026. Marco Argenti, the bank's chief information officer, leads the work on its side. Experience with the coding assistant Devin preceded the collaboration with Anthropic.
The application areas are transaction reconciliation, trade accounting, client vetting and onboarding, and compliance work such as parsing documents and applying rule sets. This is the settlement of trades already done, and not the execution of trading decisions. Turning that into "agents for trading" asserts autonomous activity in the markets, which appears in none of the sources.
The direction of travel is nonetheless clear. David Solomon, chairman and chief executive officer, has set out the aim of limiting headcount growth through the generative AI rebuild. The agent replaces no traders here. It replaces the work behind them.
Citi: the agent as a visible team member
Citi has gone furthest towards the client, and has done so visibly. On 22 April 2026 the bank introduced Citi Sky, presented as an AI-powered member of the Citi Wealth team and built on Google Cloud and Google DeepMind technologies. The technical foundation is the Gemini Enterprise Agent Platform, together with real-time avatar technology and Gemini's live audio and video models.
The rollout begins in summer 2026, phased across the United States for Citigold clients, initially in English and Spanish. Its functions include alerts on maturing certificates of deposit (CDs), market views from the wealth division's chief investment office, and conversational support on financial questions.
Andy Sieg, head of wealth at Citi, describes the intent as follows: “This is the shift from interface to intelligence, from transactions to outcomes.” On what actually prompts clients to get in touch, he adds: “At the center is a universal question: ‘Am I financially okay?'” Thomas Kurian, chief executive officer of Google Cloud, positions the work as a template: “With Gemini Enterprise as the backbone of Citi Sky, combined with frontier models from Google DeepMind, Citi Wealth is establishing a new blueprint for agentic AI.”
The risk shift that rarely gets named
Between an agent in the back office and an agent in the client conversation lies more than a distance on the org chart. In the back office a human normally checks the output before it takes effect: the entry is approved, the alert is assessed, the report is read. An agent in a client conversation leaves that zone. Its output is the effect, because the client hears or reads it in the same moment.
Seen that way, the Morgan Stanley restriction reads less as caution and more as a clean answer to a hard question: which control can still bite in a real-time interaction. Once no human vets each output in advance, control has to move into the frame, into the question of what the agent may dispose of and what it may not.
Other banks in the Reuters round-up confirm the range. Robin Vince, chief executive officer of Bank of New York Mellon, describes digital workers as team members with their own log-in identities and nicknames. Richard James, head of AI product at UBS, reports agents that flag necessary client actions to advisers, who, as a result, could spend around 70 per cent of their time in client conversations. That figure comes from the Reuters account and could not be checked against a primary source.
What survives of the 51 per cent
The figure is real, syndicated independently several times and attributed to a named spokesperson. Neither the sample size nor the exact question wording could be established, and two publicly available KPMG surveys from the same period do not carry the number. For a metric meant to underpin investment decisions, that is thin.
It remains useful as a direction of travel and not as a measurement. Bhavi Mehta, global lead for advanced analytics in financial services at Bain & Company, puts the position plainly: “Banks are increasingly using agentic AI and figuring out more ways to use it because it has a lot of potential.” Anyone measuring their own lead or lag against an adoption rate whose denominator nobody knows is benchmarking against an estimate.
What this means in practice
Four starting points follow for leaders in distribution, operations and digital.
Now: Before any agent project, set down in writing what the system may dispose of and what it expressly may not. Morgan Stanley separates availability from decision rights; that separation travels well and determines which controls are needed afterwards. Draw the line only after the pilot and you are drawing it against expectations that have already formed.
In planning: An agent whose output a human checks before it takes effect and an agent that speaks directly to a client belong to different risk classes. Moving from one to the other is no mere scaling step. It calls for its own approvals, its own logging, and its own answer to what happens when the agent gets an answer wrong.
Now: The 51 per cent works as a mood reading and fails as a benchmark. Lift it into internal reporting and you are measuring yourself against a figure with no publicly traceable methodology. Your own operational metrics serve better: number of use cases in production, share of use cases with documented oversight, handling time before and after.
Ongoing: None of the three projects is fully in production. Goldman Sachs names no launch date, Morgan Stanley is testing, Citi is rolling out in phases. Infer competitive pressure from such announcements and you are comparing your own live operation with someone else's press release. Maturity belongs in every competitive assessment.
Christian Schablitzki
Strategy & Management Consultant · Agentic AI expert for financial institutions
More than 20 years in investment banking and derivatives trading, followed by over 10 years advising financial institutions. Currently Partner at Infosys Consulting in Germany. Certified in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.
LinkedIn profile →Keep reading – every fortnight in your inbox.
Capital markets insights, regulatory updates and AI trends. Concise, well-founded, free of charge.
GDPR-compliant. Unsubscribe at any time.