A forward deployed pod of 3–5 tkxel engineers works inside your environment to build AI agents for KYC refresh, claims intake, underwriting support and reconciliation, with audit trails, human sign-off and model documentation your risk team can review.
Book discovery callAgents gather customer documents, screening hits and transaction history from your KYC platform and case manager, draft the alert narrative and a recommended disposition, and route it to an analyst. Nothing closes without a named reviewer, and every source the agent read is logged.
For insurers, agents read FNOL emails, adjuster notes, photos and policy wordings, populate the claim in Guidewire or your core system, check coverage against the policy schedule and flag subrogation or fraud indicators for the adjuster, who keeps authority over payment.
Agents pull financial statements, bureau data and covenant terms into a draft credit memo or underwriting summary, citing the page each figure came from. Underwriters edit and approve; the agent never sets price, limit or decision, so fair lending accountability stays with your people.
Agents match ledger, custodian and bank statement records, trace breaks through trade and cash movements across systems, and propose the journal entry or counterparty query. Proposed entries wait in a maker-checker queue, so your existing segregation of duties stays intact.
We build the agent alongside its paperwork: purpose, data lineage, prompt and model versions, test sets, known limitations and monitoring thresholds, written so your second line and model validation team can challenge it under your model risk policy.
Agents assemble evidence packs for examiners and internal audit, tie regulatory return inputs back to source systems, and keep an immutable log of every prompt, retrieval, tool call and human decision, retained to match your records schedule.
We’ll map one workflow, its controls and its reviewers in a 30-minute call.
Book a scoping callTeams build a demo, then find validation needs documentation, test evidence and monitoring nobody captured. Rebuilding for review often takes longer than the build.
Customer, transaction and claims data live in core systems with tight access rules, so outside teams wait weeks for extracts that go stale.
US model risk guidance now leaves generative and agentic AI to each institution’s existing governance, so you have to set your own bar and defend it.
A platform that cannot show why it recommended a disposition is hard to defend to an examiner, an auditor or a complaints team.
We sit with operations, compliance and model risk to pick one workflow, define what the agent may and may not decide, and agree the evidence your validators need. Third-party risk onboarding and access requests run in parallel so engineering is not blocked.
Deliverables: workflow map, decision boundary, control matrix, data access request
The pod builds in your cloud tenancy or data center, using your identity provider, logging and approved model endpoints. Every release ships with test results against historical cases your analysts have already decided.
Deliverables: working agent, evaluation set, integration code, audit log schema
Your second line reviews the documentation and challenges the test evidence while the agent runs in shadow mode or with full human review on live cases. We close findings before anyone considers reducing review depth.
Deliverables: model documentation, validation findings log, pilot results, monitoring thresholds
Your engineers take ownership through paired runbooks and on-call shadowing. Code, prompts, tests and monitoring stay in your repositories with no licence tied to tkxel.
Deliverables: runbooks, source repositories, monitoring dashboards, training sessions
We build in your repositories and environments from day one, so there is no separate rebuild between proof of concept and release.
Each pod is led by an engineer who has shipped AI systems before, drawn from tkxel’s team of around 150 AI professionals. The same people stay through delivery.
Autonomy matrices, audit trails, and human review are designed with your risk and security teams before release, backed by ISO 27001 controls.
Every engagement starts with agreed business metrics and evaluation sets. We report against them throughout, and the pod exits when your team can run the system.
AI agents in financial services mostly handle the preparation work around a human decision: gathering KYC documents, drafting AML alert narratives, reading claims files, building credit memo drafts and investigating reconciliation breaks. The agent reads from your systems, proposes an action and logs its reasoning; an analyst, adjuster or underwriter approves it. Fully autonomous decisions on credit, claims payment or suspicious activity reports are rare, and we think they should stay that way until your governance and your supervisors are comfortable with the evidence.
Not directly any more. SR 26-2, issued by the Federal Reserve, OCC and FDIC in April 2026, replaced SR 11-7 and places generative and agentic AI models outside its scope, with the agencies signalling a separate request for information on AI. That does not remove obligations under BSA/AML, fair lending, UDAAP, privacy or third-party risk rules. We build as if the agent will be validated anyway: documented purpose, test evidence, known limitations and monitoring, so your model risk team can apply its own standard. Confirm the details with your counsel.
Yes, and that is how a forward deployed pod works by default. Our engineers build in your cloud tenancy or data center, use the model endpoints your institution has approved, such as a private Azure OpenAI or Amazon Bedrock deployment or a self-hosted open-weight model, and authenticate through your identity provider. Customer and transaction data does not leave your environment. If your policy requires it, engineers work on your managed devices or virtual desktops.
No, not in the systems we build. The agent prepares the file, cites its sources and proposes an outcome; a named person decides. Here is the limitation to plan around: agents can cut the time an analyst spends per case, but in regulated decisions they rarely remove the review step altogether. If your business case depends on taking people out of a review queue entirely, it is likely to disappoint, and we would rather say so during scoping.
We log every prompt, retrieved document, model and prompt version, tool call, output and human decision, with timestamps and user identifiers, into your log store or SIEM under your retention rules. Logs are write-once, so an examiner or internal auditor can replay exactly what the agent saw and proposed on any case. The same records feed monitoring dashboards that show override rates by reviewer and case type.
For one workflow, plan on roughly 12–16 weeks from kickoff to handover, with production use beginning during a supervised pilot. The code is rarely the critical path. Third-party risk onboarding, data entitlements and validation queues usually are, so we start them in week one. Institutions with long vendor approval cycles or a validation backlog should expect the upper end of that range or longer.
Governance, Agents
1 October 2026 · 8 min read
Read article
Governance
24 September 2026 · 8 min read
Read article
Agents, Governance
17 September 2026 · 8 min read
Read article