Loading…
AI Agent Development · New York
Trade settlement ops at a mid-sized broker-dealer in Midtown still runs five analysts who reconcile DTCC breaks against the order management system by hand. T+1 made the window tighter, not the process simpler. An agent connected to CTM, the OMS, and the custodian feed clears 80% of the breaks before a human sees them. The remaining 20% land in a queue with the reasoning attached.
Insurance claims ops at Guardian, MetLife, and the carriers in Lower Manhattan have the same shape of problem. First Notice of Loss triage, eligibility checks, fraud signal screening, document requests to the insured. Each step is a 4 to 12 minute manual action repeated 600 times a day. We build agents that handle the deterministic 70% so adjusters work the exceptions.
Every project runs as a fixed scope. Discovery first, fixed quote after, then four to eight weeks of build with weekly demos.
Describe the workflow.
The pattern shows up across every New York vertical we talk to. Ops teams of 4 to 30 people spend most of their day moving structured data between systems that almost talk to each other, making the same routing or reconciliation decision over and over with slight variations.
At a bulge-bracket investment bank near Bryant Park, the middle office runs trade affirmation against CTM, reconciles fails with custodians, and produces the regulatory reports for FINRA and the SEC. The 7am to 11am window before US market open is the hottest part of the day. Headcount is expensive, training time for a new analyst is 4 to 6 months, and the work is not interesting to anyone qualified to do it.
Commercial real estate firms around Park Avenue do lease abstracting at the same intensity. A 60-page office lease has 80 to 120 data points that have to land in the lease admin system (Yardi, MRI, VTS). A senior analyst takes 3 to 5 hours per lease. An agent with a document parser and a clause classifier lands the same data in 4 minutes, with the analyst reviewing the extracted JSON instead of the raw PDF.
Insurance claims FNOL triage at MetLife, Guardian, and the regional carriers along Sixth Avenue is the third shape of the same problem. A claim comes in by phone, email, broker portal, or partner API. The adjuster checks eligibility against the policy admin system, screens for fraud signals, requests missing documents from the insured, and assigns severity. That sequence takes 25 to 40 minutes per claim and is largely deterministic.
Hotel revenue management at the Marriott and Hilton franchisees in Times Square has a fourth variation. Rate decisions across 18 to 24 distribution channels, competitor scraping every 4 hours, event calendar overlays for Broadway and the Javits Center, plus the channel manager push. Revenue managers are paid for the judgment call. They should not be paid for the data assembly.
Every one of these workflows is a candidate for an agent. Not a chatbot. An agent with tool use, structured output, an audit trail, and a human gate at the irreversible action.
Six components show up in almost every agent we ship for finance, insurance, real estate, and hospitality ops in the city.
Plan-execute or ReAct depending on the workflow. Built with LangGraph state machines so transitions are explicit, not magic. The graph is reviewable and testable, not a 200-line prompt.
CTM, DTCC, Bloomberg, FactSet, Guidewire, Duck Creek, Yardi, Opera, SiteMinder, Salesforce. Every tool call has a typed schema, retry policy, and a fallback when the upstream API is degraded.
Short-term context lives in the agent state. Long-term memory uses pgvector or Pinecone, indexed by tenant, policy number, deal ID, or whatever your row-level key is. We do not let the model invent context.
Any irreversible action (book a settlement, deny a claim, send a contract) routes to a human approver in Slack, Teams, or a custom queue. The agent waits. Approval includes the reasoning trace.
FINRA Rule 4511, NY DFS Part 500, SOC 2. Every prompt, tool call, decision, and override is written to an append-only log with model and prompt versioning so a compliance review can reproduce any past decision.
We measure task completion rate, hallucination rate, and tool-call accuracy against a labeled set you provide before the agent goes near production. Regression suite runs on every prompt change.
New York concentrates the industries where the unit economics of an agent work most clearly. The financial services labor market in the metro is roughly 470,000 people. Insurance employs another 160,000. Commercial real estate, hospitality, and legal services push the back-office workforce well past a million. A 30% reduction in manual data assembly time at any non-trivial firm in that group is a meaningful number on the P&L.
We have talked to teams at Goldman Sachs, JPMorgan, Bloomberg, and Palantir about agent patterns at the architecture level, and we have shipped systems for smaller firms in the same verticals. Squarespace and the late-stage SaaS companies in SoHo and the Flatiron run a different shape of agent (customer ops, billing dispute triage, growth analytics) but the building blocks are the same.
We work remotely with New York clients on Eastern time. Discovery calls fit between 9am and 5pm ET, demos happen weekly, and the codebase ships to a repository you own. No office visits required, and no agency-style timezone gymnastics either. The team is in North America and overlaps with your business hours by design.
Industries where we see strongest fit: investment banking middle office, insurance claims and underwriting, commercial real estate ops, hotel revenue management, and law-adjacent document processing.
Tell us what your ops team does every day that you think a software agent could handle. We'll reply within one business day with a rough scope and price range. No commitment, no sales call required.