An opportunity map for agentic AI across Services, Markets and Banking — grounded in Citi's disclosed FY2025 financials and real industry benchmarks.
Citi delivered record FY2025 revenue and positive operating leverage across all five businesses — while committing to keep bending the cost curve. The wholesale franchise, with its scale and process intensity, is where AI compounds fastest.
Treasury & Trade Solutions (cross-border / international payments, USD clearing) + Securities Services. FY2025 revenue $21.3B · opex $10.8B · cross-border value $416B.
Fixed Income + Equities (incl. Prime). High-volume, low-latency. FY2025 revenue $22.0B · opex $14.1B.
Investment Banking (M&A, ECM, DCM) + Corporate Banking. FY2025 revenue $8.2B · opex $4.5B.
Shared spine: onboarding/KYC, reference & client data, trade & payment processing, reconciliations, settlement, controls, and regulatory reporting — where most addressable expense (and AI opportunity) concentrates.
Each estimate is built bottom-up from a disclosed segment cost pool × an addressable share × a conservative AI savings rate — and anchored to a documented "quantum-leap" benchmark. Click any card to expand.
Pick an opportunity and adjust the addressable cost pool and AI savings rate. The financial impact and the portfolio total recompute live. Defaults reflect the conservative ranges in the analysis.
From copilot to autonomous multi-step agent — grounded in Citi data and bank-grade governance. Pick a scenario and run the agent.
Bubble size = estimated annual run-rate impact. Hover any bubble for detail. Sequencing weighs ROI certainty, cost-pool size, and data/regulatory dependencies.
The inside view — which you and the team have — decides the real sequence. This is an outside-in frame to react against: rough order of magnitude throughout, and much of it likely already in flight.
Illustrative by design — the real sequence, numbers, and scorecard belong to the team. The intent: show where hands-on outside help could contribute from week one.
A serious review will — and should — attack this brief. These are the challenges it expects, answered plainly; several are conceded outright, and each concession maps to something the weeks 1–4 baseline settles with internal data. Click any card.
A fair challenge — and the reason the headline is stated as gross capacity-plus-expense potential rather than savings. The committed program is a net figure over a defined window; the estimate here sits before overlaps between rows, investment cost, and redeploy-or-release decisions. The honest claim: AI is one lever capable of carrying a large share of the committed number — not a new number on top of it.
Conceded — and the distinction is the whole game. Capacity becomes P&L only through explicit decisions: released positions, avoided backfills, reduced vendor and BPO spend, or redeployment into revenue work (which surfaces as operating leverage rather than expense). The proposed scorecard carries 'hours redeployed vs. released' as a first-class metric for exactly this reason.
Triangulated from disclosed segment expense, industry cost-structure studies, and peer analogs — rough order of magnitude by design. The sliders exist so every pool can be replaced by internal data the team already holds. Any single number should be expected to move materially; ranking and direction are the robust part.
They do. The portfolio view is presented as a landscape, not a sum — gross, overlapping potential rather than an additive total. A bottom-up build on internal data nets the overlaps in weeks — which is the weeks 1–4 baseline exercise.
All true. Benchmarks are used as bounds on what has been achieved somewhere, not forecasts for Citi — the modeled savings rates sit deliberately below them. The failures are design inputs too: Klarna's over-automation of judgment work is precisely why human-in-the-loop tiers gate every regulated or judgment-dense step here.
Agreed — and the sequencing says so. The first wave is SDLC plus agent-assisted investigation that drafts repairs and rationales for human sign-off; no autonomous action in regulated flows. Autonomy earns its way in through low-risk, reversible steps, against an evidence bar Model Risk and Compliance define. The reference-data work the orders require is the same foundation these use cases need — complementary programs, not competing ones.
Cost depends on build-vs-buy, platform reuse, and restructuring choices only internal planning can make. Directionally, programs of this shape run net-negative for 12–24 months before crossover, and the self-funding logic in the 90-day frame — early lighthouse savings financing the next wave — is how the curve stays shallow. A credible investment case is a weeks 9–12 output, not a pre-meeting guess.
They failed on liability, jurisdiction, and data ownership. What's proposed is internal consolidation — one client record, one policy engine, a shared service inside the bank's own perimeter — which is exactly the design those failures argue for.