Loading…
AI Agent Development · Chicago
A mid-tier FCM in the West Loop runs six analysts every evening to reconcile CME and CBOT futures trades against their internal blotter. The settlement file lands at 4:45pm Central. By 6:30pm the breaks have to be identified, classified, and either resolved or escalated. The work is structured, the rules are written down, and the volume is growing faster than headcount. An agent connected to CME ClearPort and the give-up system clears the routine 70% so the team focuses on real disputes.
Insurance ops at Allstate's Northbrook campus, Aon's River North office, and the regional carriers across the metro follow the same shape. First Notice of Loss triage, eligibility checks against the policy admin system, fraud signal screening via ISO ClaimSearch, and document requests. The adjuster makes the judgment call. The agent does the assembly.
Fixed-scope projects get a fixed quote scaled to the work. Discovery first, quote after, four to eight weeks of build, demos every Friday on Central time.
Tell us what your team does at end of day.
Chicago is the country's densest concentration of derivatives trading, commercial insurance, and supply-chain operations. Each of those verticals shares a back-office pattern that does not need a 12-person team to execute: read a file, join it to another file, apply a rule, write the result somewhere, route the exception to a person.
On the trading side, the settlement and reconciliation cycle for CME and CBOT futures has a fixed evening window. Ops at the FCMs, the prop shops in the Board of Trade building, and the clearing groups along Wacker spend the hours between market close and overnight clearing chasing breaks. The give-up flow between executing brokers and clearing firms creates breaks even when nothing is technically wrong. The agent reads the CTM-equivalent ClearPort file, joins it against the OMS, and proposes a resolution per break.
In insurance, the catastrophe season and the daily auto-claim volume keep adjusters at carriers like Allstate, State Farm (Bloomington-adjacent), and the regional mutuals working through a queue that does not shrink. First Notice of Loss arrives via phone, broker portal, email, and partner API. Each FNOL needs eligibility verification, fraud screening, and triage to the right adjuster pod. That sequence is 18 to 30 minutes of mostly deterministic work per claim.
On the logistics side, distributors in the Pilsen industrial belt, the 3PLs along I-55, and the e-commerce ops that grew up around the Grubhub and Groupon talent pool all run exception queues that look identical from a system-design perspective. A shipment is delayed, a courier no-shows, an SKU is missing, an address fails. The exception arrives as an EDI 214, a customer ticket in Zendesk, or a partner API webhook. Ops makes the call and updates the WMS.
The Morningstar and Relativity engineering offices have shown that the same agent patterns work for research data ingestion (fund prospectus parsing, fee-schedule extraction) and for legal e-discovery review (pre-classification of responsive documents before attorney review). Same building blocks, different domain schema.
We build the agent. Your team owns the runbook, the audit log, the prompts, and the evaluation set.
The six things that appear in almost every agent we ship for trading ops, insurance, supply chain, and legal in the metro.
LangGraph for stateful workflows where breaks-resolution or FNOL-triage steps need to be reproducible. AutoGen for multi-agent coordination when one agent gathers data and another writes the recommendation. We pick by workflow shape, not preference.
CME ClearPort, CBOT settlement files, give-up systems, FIS and Calypso OMS adapters, custodian APIs for daily position files. Every tool call has a typed schema, idempotency keys, and a retry policy that respects exchange rate limits.
Guidewire ClaimCenter, Duck Creek, ISO ClaimSearch, AM Best, NICB referrals, EDI 837 and 835. The agent normalizes inputs into a typed claim object before reasoning, so the LLM never sees malformed data.
pgvector or Weaviate for long-term memory indexed by policy number, trade ID, or shipment number. Short-term context in the agent state. Deduplication so the same FNOL or break does not get worked twice when retried.
Append-only log of every prompt, tool input, tool output, decision, and human override, with model version pinned. Satisfies CFTC, FINRA, Illinois DOI, and SOC 2 review requirements without bolt-on tooling.
Task completion rate, hallucination rate, tool-call accuracy. We use Promptfoo or Braintrust against a labeled set you provide. Regression tests run on every prompt change before merge.
Chicago has the rare combination of deep finance, deep insurance, and deep logistics in the same metro. CME Group runs the largest derivatives exchange in the world from the Wacker building. Morningstar publishes the data the buy side runs on. Relativity builds the platform half the country's litigation teams use. Outcome Health, Sprout Social, and Enova represent the SaaS layer that grew up around that infrastructure.
The ops shape of these businesses is what makes them a good fit for agent work. End-of-day reconciliation, daily claims volume, continuous exception queues, and overnight processing windows are the moments where a software agent earns its keep. A trader does not want to be the person reconciling at 9pm. A senior adjuster does not want to be the person screening for fraud at 11am. The agent handles the assembly, the human owns the judgment call.
We work remotely with Chicago clients on Central time. Discovery calls fit between 8am and 5pm CT, demos happen weekly, and the codebase ships to a repository you own. The team overlaps with your business hours by design, not by accident.
Industries where we see strongest fit: futures and derivatives trading ops, P&C insurance claims and underwriting, supply chain and 3PL exception handling, and legal-tech e-discovery pre-review.
Tell us what your trading, claims, or ops team does after market close (or before market open) that you think a software agent should handle. We'll reply within one business day with a rough scope and price range.