Agentic AI in Finance Timeline 2023–2026: How Wall Street Is Moving From Chatbots to AI Agents

A chatbot takes a question and returns an answer. An agent takes a goal, makes a plan, calls tools, takes actions and returns a result — and can maintain state across the steps. A multi-agent system divides that work across specialised agents, for example a research agent, a risk agent and a compliance agent coordinated by an orchestrator, usually with a human approval step before anything is actioned.

The important correction: agentic does not automatically mean multi-agent. A single agent using several tools is still an agentic system, and many bank deployments are exactly that. Multi-agent architecture is an option, not a requirement — and it introduces its own failure modes (agent disagreement, error propagation, authority confusion, feedback loops) rather than being automatically safer.

Banks generally do not want an unconstrained AI model with unlimited access to money, markets, client accounts or regulated systems. Instead, agentic AI is connected to specific tools, specific data, specific permissions, specific workflows and human approval gates. The critical word is bounded.

An agent may be allowed to read documents, query databases, calculate values, draft forms and reconcile records. It is not necessarily allowed to move funds, send regulated communications, approve client onboarding, change risk limits or execute trades. The more authority an agent gets, the more important its permission boundary becomes.

Human-in-the-loop vs human-on-the-loop

Human-in-the-loop: the agent must pause and get approval before a critical action. Human-on-the-loop: the agent can operate inside pre-approved boundaries while a human monitors and can intervene. Full autonomy: minimal or no human approval. For regulated finance, many high-risk applications should stay bounded and approval-gated — and neither in-the-loop nor on-the-loop is mandatory for every agent; the right setting depends on the action’s blast radius.

At the February 2026 report, Goldman was described as “in the early stages” of developing these agents and expecting to launch them “soon” — not as having deployed autonomous Claude agents across the bank. Goldman executives said they were surprised at how capable Claude was beyond coding, in accounting and compliance work that combines parsing large volumes of documents with applying rules and judgement. The bank also reported a roughly 30% reduction in the time to onboard new institutional clients using Claude’s reasoning. 🔵

Why code is easier than money

Anthropic released Claude Opus 4.6 on 5 February 2026, improving agentic task performance, coding, long-horizon work and document and spreadsheet workflows, and some reporting connects that model to Goldman’s agent development. But Goldman’s confirmation refers broadly to Claude-based agents. Unless a first-party Goldman source names the exact production model, the bank’s agent strategy is better described model-agnostically. Model versions change quickly; the architecture is what lasts: approved data → model → tools → permissions → checks → human gate → audit trail.

When a source says AI is used in trading, it can mean any of: research, analytics, data processing, surveillance, execution support, pricing, workflow automation or algorithm development. It does not automatically mean an LLM independently selects a security, chooses a position, sets size, executes and manages risk. There are at least six distinct meanings, and collapsing them into “AI trading” is the single most common error in coverage:

1. AI helps trading research — summarising filings, news, positioning.
2. AI helps algorithm development — writing and testing execution code.
3. AI adjusts existing models — tuning parameters a human owns.
4. AI suggests orders — a human decides.
5. AI manages a portfolio under constraints — inside hard limits and suitability rules.
6. AI autonomously creates and executes strategies — the frontier, not documented as mainstream.

Algorithmic trading is not agentic AI

Algorithmic trading has existed for decades: a predefined rule or model receives market data and executes. An agentic system can potentially plan, reason, select tools, adapt its workflow and make sequences of decisions. Not every algorithm is an AI agent.

High-frequency trading is not an LLM agent

High-frequency trading prioritises latency, deterministic execution, market microstructure and specialised infrastructure. LLM agents add reasoning and tool orchestration — but also latency, non-determinism and model risk. It is technically misleading to say “LLM agents execute trades at zero latency.” The more accurate concern: machine-speed decisions can amplify risk if more autonomous AI is eventually connected directly to financial decision-making.

On flash crashes: regulators are not claiming an AI-driven “Flash Crash 2.0” is expected tomorrow. The historical analogy is used to frame a research question about correlated automated behaviour — not to describe a documented agent-driven crash. And using AI does not erase existing regulation: rules on supervision, records, market conduct, best execution, risk controls and communications can still apply. Regulation here is technology-neutral.

Traditional software usually follows defined logic. Generative AI can produce unexpected output. Agentic AI adds another dimension: unexpected action. So the key governance question is not only “was the answer correct?” but “was the agent allowed to do that?“

For every action, a bank may need to reconstruct which model acted, what data it saw, which tools it called, what permissions it had, what rule it applied, what intermediate output it generated, whether another agent reviewed it, whether a human approved it, and what happened afterward. That makes auditability, permissioning, identity, monitoring and action logs into core financial infrastructure.

Liability is not a simple choice

Responsibility for an agent’s mistake is not a clean pick between the bank, the model provider and a human supervisor. It can depend on the activity, jurisdiction, contract, regulatory obligation, delegation, supervision and product design. In many regulated contexts, an institution cannot simply outsource its responsibility to an AI vendor.

Traceability, not full explainability

Banks realistically need traceability — observable inputs, outputs, actions, tool calls, approvals and policy decisions — rather than a promise of full explanation of a model’s internal reasoning. Every action needs a receipt: input → plan → tool call → data accessed → policy decision → approval → action → result.

Multi-agent systems do not automatically add safety. A “risk agent” checking a “research agent” can add defence — but if both share the same model, data and prompt assumptions, they can fail together. Real safety diversity includes deterministic rules, traditional risk engines, independent data, hard position limits and human review, not just asking another language model.

Agents do not automatically reduce operating cost. An agentic workflow can require many model calls, long contexts, multiple agents, tool execution, monitoring and human review. Goldman Sachs Research has said it expects agentic AI to materially increase token consumption, because agents think, check, retry, call tools, delegate and process more context. One chatbot handles one task; a multi-agent setup adds an orchestrator, a researcher, a reviewer, a risk check and retries. The right scorecard measures cycle-time reduction, manual touches, error rate, exception rate, human review time, compute cost, regulatory incidents and customer outcome — not “jobs eliminated” as the only metric.

The honest answer is not yes or no: tasks are likely to change before whole occupations disappear. Work most exposed includes document review, data gathering, routine coding, reconciliation, pitch materials and basic analysis. Human-heavy work — relationships, judgement, negotiation, risk ownership and regulatory accountability — remains central.

For decades, a junior banker learned by doing the slow work: reading the filings, checking the spreadsheet, updating the presentation, finding the number that did not reconcile. AI agents are very good candidates for exactly that kind of work. The productivity case is obvious. The training problem is not. If software completes the first five steps, the next generation of bankers may arrive at step six without having learned why the first five mattered.

Consider an analyst who asks an AI: “Prepare tomorrow’s client meeting.” A chatbot writes a briefing note. An agent could open the CRM, retrieve the client’s holdings, check recent earnings, review previous meeting notes, update approved analytics, find upcoming maturities, prepare the presentation and schedule a reminder. The human asked one question; the machine performed many tasks. That is the agentic shift.

Now change the instruction to “Manage the client’s portfolio.” The same convenience becomes a governance problem. What can the agent buy? How much? Using which data? Under which suitability rules? Who approved the strategy? Who stops it? That is why agentic finance will likely advance one permission boundary at a time.