Running entirely on your own servers and devices. No data shared. No surveillance. No dependency. No one else's infrastructure.
In August 2026, Caltech's Anima Anandkumar and Benedikt Jenik came out of stealth with Accelerated Understanding — an AI built on neural operators instead of transformers, trained to predict how physical systems evolve rather than what word comes next. It doesn't chat. It simulates weather, materials and chips, and it processed 5 trillion data points in a single prompt during testing.
Their bet: once a model can simulate a system accurately enough, you no longer need to run the experiment to know the outcome. The same logic applies to a balance sheet. A bank doesn't need another chatbot to describe its liquidity risk — it needs a system that can run the stress test.
| OSAI (open-source, iBANK) | Closed single-vendor AI | |
|---|---|---|
| Model layer | Choice of 8+ open-weight models (DeepSeek, Qwen, Kimi, GLM, MiniMax, Ling, gpt-oss, MiMo) — swap per workload | One vendor's model, one roadmap, one pricing table |
| Hosting | Self-hosted or private-cloud — data stays inside the bank's regulatory perimeter | Third-party API, data leaves the estate by default |
| Agent design | 50 narrow, auditable finance agents, each with a defined task and tool scope | One general assistant asked to do everything |
| Integration | MCP-native — agents call core banking, treasury and market data through scoped servers | Bespoke API glue per integration, rebuilt every time |
| Cost model | Compute you control, licence-free model weights | Per-seat or per-token fees that scale with usage |
Banks don't get to treat AI adoption as a pure technology decision — every model choice is also a data-residency, outsourcing and model-risk decision, and open weights change the shape of all three.
Sending customer transaction data, credit files or board materials to a third-party model API means that data leaves the bank's regulatory perimeter, even briefly. Under UK outsourcing rules (SS2/21), EU DORA, and equivalent regimes elsewhere, that triggers material-outsourcing obligations — due diligence on the vendor, concentration-risk assessments, exit planning, and in many cases regulator notification, before a single query is ever run in production.
None of that is fatal — plenty of banks clear it — but it's slow, it's recurring, and it puts a third party's roadmap and pricing changes inside the bank's own risk register. A vendor's decision to deprecate a model, change a data-retention policy, or raise per-token pricing becomes the bank's problem to manage, on the vendor's timeline.
Self-hosting an open-weight model inside the bank's own infrastructure — on-premises or in a private cloud tenancy the bank controls — keeps inference inside the existing regulatory perimeter. There's no new data-sharing agreement to negotiate, no new sub-processor to disclose, and no dependency on a vendor's continued willingness to serve the account.
OSAI sits between iBANK's core banking data and the people who need answers from it — ingesting, orchestrating, reasoning, and reporting, with security and observability running underneath every layer.
OSAI isn't tied to one hosting pattern. The 50 agents and the underlying model layer run the same way in each case — what changes is where the compute physically sits, chosen to match the bank's existing infrastructure and regulatory posture.
Every OSAI agent reaches iBANK's systems through the Model Context Protocol — a standard way to discover tools, request data, and get structured results back, instead of a one-off integration per agent.
Accounts, balances, transaction history, ledger entries via iUL
Cash positions, FX, debt schedules, capital allocation via iTreasury
Payment flows, reconciliation, merchant data via iPAY / iPAL
Controls, audit trails, regulatory filings, policy documents
1. A user or system asks OSAI a question — "flag any account with debt covenant breach risk this quarter."
2. The orchestrator plans which agents and servers are needed — Debt Schedule Agent, Risk Assessment Agent.
3. Each agent discovers the right MCP server and its available tools.
4. Structured calls go out — args in, JSON results back.
5. The Compliance and Treasury servers execute against real systems.
6. Results return as grounded context, and the model synthesises the answer — with the source data attached, not paraphrased from memory.
Every tool call is scoped, logged and attributable. An agent can only reach the servers it's been granted, and every server enforces its own auth and rate limits. That means audit teams can trace exactly which data an agent touched to reach a conclusion — the same bar a human analyst would be held to.
Narrow, auditable agents make governance tractable in a way a single general-purpose assistant never can — each agent's scope is small enough that a reviewer can actually reason about what it's allowed to touch and what it's allowed to decide.
Agents draft, flag and recommend; anything that moves money, changes a filing, or alters a customer record routes to a named approver before it executes
Model selection, validation and monitoring map onto existing model-risk frameworks (SR 11-7-style three-lines-of-defence, EU AI Act high-risk obligations)
Every answer carries the source data it was grounded in via MCP, so a reviewer can trace a conclusion back to the ledger entry or filing it came from
An agent that proposes a transaction is never the same agent scoped to approve it — the split is enforced at the MCP server permission level, not by policy alone
A general-purpose assistant asked to "help with finance" has an almost unbounded scope, which makes it genuinely hard to write a control framework around — you end up governing the tool rather than the task. Fifty agents, each with one job and one set of MCP servers it's allowed to call, turn that into fifty much smaller, much more specific control statements: what this agent can read, what it can propose, and what it is never permitted to do on its own.
Every tool call, every piece of retrieved context, and every output an agent produces is logged with the agent's identity, the servers it touched, and the model version that generated the response. That's the same evidentiary standard an audit team would expect from a human analyst's working papers — which is what makes it possible to sign off on agent output rather than simply trusting it.
Rather than one general assistant, OSAI runs fifty single-purpose agents grouped across the finance function — each with a defined scope, so its output can be reviewed the way you'd review any analyst's work.
Running entirely on private infrastructure. 82 specialized autonomous agents interconnected via Model Context Protocol (MCP) across core operational verticals.
Predicts SKU-level demand using historical sales, seasonality, and market trends.
Calculates safety stock, reorder points, and economic order quantities (EOQ).
Flags potential out-of-stock scenarios before they impact sales.
Identifies slow-moving inventory and models discount strategies to clear shelf space.
Adjusts prices in real-time based on competitor tracking, velocity, and stock levels.
Analyses product performance to recommend additions or deletions to the product catalog.
Tracks supplier lead times, fill rates, and defect percentages.
Optimizes behavioral triggers and personalized offers to recover lost digital sales.
Allocates ad spend across on-site and off-site retail advertising channels.
Forecasts long-term value tiers for cohort-based marketing targeting.
Analyzes return patterns to mitigate fraud and optimize restocking costs.
Audits EHR notes against ICD-11 and CPT codes to prevent billing mismatches.
Compiles patient history and insurance criteria to expedite approval requests.
Diagnoses rejected insurance claims, fixes errors, and drafts appeal documentation.
Analyzes appointment history to optimize scheduling intervals and overbooking rates.
Screens electronic health records against complex trial inclusion criteria.
Predicts ER and ward patient volume to build optimal nursing schedules.
Models chemical interactions to narrow down viable drug candidates.
Aggregates and structures data outcomes during clinical trial execution.
Monitors temperature-sensitive shipments (cold chain) and checks serialization compliance.
Pulls credit bureau data, banking history, and cash flows to issue loan decisions.
Conducts deep corporate entity lookups and macro risk mapping for mid-market business loans.
Verifies asset documentation, income statements, and property appraisals automatically.
Monitors asset drift against target allocations and executes tax-loss harvesting.
Assesses client risk tolerance through behavioral inputs to recommend investment paths.
Parses earnings transcripts, news feeds, and alternative data for micro-signals.
Simulates historical loss vectors to dynamically price auto, home, or life insurance policies.
Triages digital insurance claims by processing accident photos and descriptions.
Flags anomalies in claim filing behaviors, networks, and repair shop invoicing.
Tracks IoT sensor data on factory machinery to schedule fixes before failures occur.
Dynamically routes raw materials and steps across assembly lines to maximize throughput.
Analyzes visual streams from the production line to spot micro-defects in real-time.
Computes multi-stop delivery routes factoring in traffic, fuel, and driver hours.
Monitors spot rates across sea, air, and rail to secure lowest-cost logistics contracts.
Directs slotting strategy based on product weight, volume, and velocity of movement.
Tracks geopolitical, weather, and financial updates affecting tier-1 and tier-2 suppliers.
Identifies where users get stuck in a trial and triggers contextual product tours.
Correlates specific feature usage paths with high long-term retention behaviors.
Flags free/trial accounts that hit usage thresholds indicating readiness for sales outreach.
Monitors drops in login frequency, API usage, or support ticket escalation spikes.
Recommends optimal timing and pricing tiers for upselling accounts nearing contract end.
Synthesizes support metrics, NPS, and usage data into an account health metric.
Forecasts solar and wind generation capability using micro-weather modeling.
Automates demand-response triggers to redirect electricity during peak consumption times.
Aggregates distributed energy resources (like home batteries) to trade power back to the grid.
Tracks Scope 1, 2, and 3 emissions across enterprise travel, bills, and shipping.
Vets and compiles validated carbon credit investments for corporate compliance.
Ensures corporate filings match CSRD, SEC, and other shifting climate reporting mandates.
Evaluates inbound contract requests and routes them to the correct internal legal expert.
Compares inbound NDAs, MSAs, and SOWs against the corporate standard playbook.
Scans executed contracts to map renewal dates, payment terms, and delivery milestones.
Runs automated PEP, sanctions list, and adverse media screenings during onboarding.
Identifies PII data exposures across company systems to ensure GDPR and CCPA safety.
Reviews internal messaging and trading windows for market manipulation indicators.
Parses resumes against complex job requirements without introducing demographic bias.
Scores project submissions or coding challenges, providing objective feedback metrics.
Handles interview scheduling, follow-ups, and common FAQs for applicants.
Matches internal salaries against regional, industry, and macroeconomic talent datasets.
Evaluates tenure, tenure trends, and engagement metrics to flag key roles at risk of quitting.
Maps existing internal skill matrices against upcoming product or service roadmaps.
Runs continuous multivariate adjustments on ad copy for specific demographics.
Tracks real-time social media conversations to alert PR teams of negative spikes.
Allocates dynamic budgets across Google, Meta, and LinkedIn based on real-time ROAS.
Analyzes incoming support emails to tag priority, urgency, and routing needs.
Resolves high-volume inquiries (e.g., track order, change email) securely using back-end APIs.
Flags frustrated customers within live-chat queues to escalate directly to senior human teams.
Estimates residential and commercial property values using geospatial and historical comps.
Evaluates rent rolls, occupancy data, and maintenance data to maximize NOI (Net Operating Income).
Scans complex commercial leases to extract rent escalations, CAM charges, and termination options.
Automates identity, background, credit, and historical eviction checks.
Decodes tenant requests to categorize emergency levels and dispatch regional contractors.
Adjusts HVAC and lighting levels based on sensor-detected occupancy patterns.
Standardizes and extracts bid data from vendor responses to present a clean comparison matrix.
Scans corporate credit card data and expense entries to catch non-compliant, out-of-contract purchasing.
Analyzes spend distribution to prevent single-source dependencies across the supply chain.
Correlates raw material price drops with supplier contract rules to trigger renegotiation windows.
Matches vendor delivery logs against contracted SLAs to claim penalties or rebates automatically.
Forecasts final university acceptance and enrollment numbers based on historical applicant behaviors.
Evaluates LMS logins, assignment drops, and test grades to flag students needing academic triage.
Scans course content to check for alignment with national accreditation standards.
Dynamically adjusts quiz difficulty tiers based on a student's previous answer accuracy.
Drafts formative grading commentary on student essays to highlight structural gaps for teachers.
Each agent moves through the same lifecycle before it ever touches production data, and stays inside it afterwards — treated less like a chatbot feature and more like a piece of financial software with its own change history.
OSAI routes each agent to whichever open-weight model fits the task — swapped, upgraded or self-hosted without rewriting the agents that sit on top.
117B-parameter MoE with adjustable reasoning depth — fits self-hosted deployments needing enterprise-grade control.
1T-parameter MoE with native multimodal support and long context — suited to agentic and planning-heavy workflows.
Ultra-long context, multilingual, strong at RAG and long-document analysis — a natural fit for compliance and filings.
Strong multilingual coding and high inference speed at efficient compute cost — good for automation-heavy agents.
Hybrid-attention MoE tuned for high-throughput inference — well suited to always-on monitoring agents.
Trillion-parameter MoE with evolutionary chain-of-thought reasoning — strong for board-pack and reporting generation.
744B parameters optimised for coding and reasoning workloads, with sparse attention for long-context efficiency.
MIT-licensed, reasoning-heavy, trained with large-scale reinforcement learning — strong at tool-calling agents.
No — open-weight refers to the model itself, not the bank's data. The model's parameters are published so it can be inspected and self-hosted; the bank's transaction data, documents and query history stay entirely inside the bank's own infrastructure and are never sent to the model's original publisher.
Because the model layer is decoupled from the agent and application layers, a flawed or deprecated model can be swapped for another open-weight model without rewriting the fifty agents sitting on top of it. That's a materially faster remediation path than waiting on a closed vendor's own patch cycle.
Not necessarily. A bank can start in a private-cloud tenancy it controls and move specific agents on-premises over time, following the hybrid pattern described above — the agents and MCP integrations don't need to change when the hosting does.
A chatbot plugin answers questions about data you show it. OSAI's agents are scoped to specific tasks with defined tool access via MCP, so they can be reviewed, tested and governed the way any other piece of financial software is — not treated as a black box bolted onto a search index.
Yes. The agent library is modular by design — a bank might start with the FP&A and Reporting & Governance categories and add Treasury or M&A agents later, without re-architecting anything already in production.
The same governance chain accountable for any analyst's work product — because every agent output is logged with its source data and the human approver who signed off on any action, responsibility sits with the bank's existing review structure, not with an opaque third party.
Universal Ledger — the account and transaction record every agent reads from
Liquidity, FX and debt data behind the Treasury and Capital Allocation agents
Payment flows and reconciliation feeding the AR/AP and Vendor Spend agents
Merchant and settlement data behind pricing and margin agents
Ticketing and marketplace revenue feeding cohort and churn agents
Close to 30 years across EY, Accenture, Oracle, Morgan Stanley, AIB and the Bank of England, where he served as Chief Architect. At Vocalink, he co-created the Oyster Card and PayM — infrastructure now used by millions every day.
He founded iBANK Incubator to put that same institutional rigour behind the ventures, not just the banks, building the next economy — concentrating his own conviction where it matters most: Quantum and AI.
OSAI is iBANK's proposal for a self-hosted, open-model, MCP-orchestrated finance agent layer.