Skip to content
DFT Solutions

Solutions catalogue

All 43 solutions are described in the same format, and any of them can be implemented on its own, starting with a pilot.

Digital twins

Modelling changes before they go live

Retail, marketing

Customer twin

A forecast of how customer segments respond to a rate or tariff change.

Task
The bank measures how customers respond to a new rate, tariff or product after launch.
Functions
Models segment behaviour and forecasts demand, churn and profitability for several campaign variants.
Data
CRM, product portfolio, transactions, mobile banking activity.
Pilot format
One scenario, for example how loan demand responds to a new rate, back-tested on past tariff changes.
Lending, operations

Credit pipeline twin

A map of credit pipeline bottlenecks and a forecast of decision time.

Task
Application volumes grow faster than underwriting headcount, and the bank cannot see at which stages time is lost.
Functions
Models how an application moves through systems and teams and shows delays, repeated checks and overload. The model is used to test a new order of checks, automation of a stage or a doubling of volume.
Data
Credit pipeline and BPM logs, staffing schedule.
Pilot format
One product, consumer loans or SME. The result is a bottleneck map and a forecast of decision time after the changes.
Anti-fraud, AML

Transaction flow twin

Testing of new anti-fraud and AML rules on synthetic transactions.

Task
A new anti-fraud or AML rule cannot be tested safely on real payments.
Functions
Runs the rule or model on synthetic and anonymised transactions that include atypical operations, networks of linked accounts and peak volumes.
Data
3–6 months of transaction flow, history of closed cases.
Pilot format
Comparison of the new rule with the current one on hit rate and false-positive share by transaction type.

Release testing in a copy of production

CTO, IT operations

Environment twin

An isolated copy of production deployed for a specific release.

Task
Test environments differ from production in versions, data and integrations, so some defects only appear with customers.
Functions
Deploys an isolated copy of production from an infrastructure-as-code template. The environment contains services, anonymised or synthetic databases and emulators of the NBU SEP payment system, card processing, credit bureaus and Diia, and exists for one release.
Data
Architecture description, configurations, database dumps for anonymisation.
Pilot format
One application with its dependencies.
QA, development

Tester agents

Automated execution of customer journeys in the mobile and web apps.

Task
Manual regression takes weeks, and automated tests need fixing after interface changes.
Functions
Generate test cases from requirements and code changes and run deposit, transfer and loan application journeys through the interface, continuing to work after it changes. They test APIs and describe defects with reproduction steps.
Data
Requirements and tickets in Jira, repository, existing test cases.
Pilot format
Regression for one release run in parallel with the QA team.
CTO, business continuity

Load and stress testing

Load with synthetic traffic up to the breaking point, and chaos tests.

Task
The bank does not know whether the system will handle payday, Black Friday or mass payouts after an update.
Functions
Generates traffic from profiles taken from the customer and transaction flow twins, ramps load to the breaking point and simulates component failures. The results feed business continuity plans and DORA operational resilience requirements.
Data
Peak load profiles, monitoring metrics.
Pilot format
One critical service, payments or authorisation.

AI agents for business units

Automated investigations in AML and anti-fraud

AML

Alert investigation agent

Evidence and a draft conclusion for an alert.

Task
Most alerts are false positives, yet closing each one requires gathering data by hand from several systems.
Functions
Gathers the customer’s and counterparties’ history, previous cases and open sources for an alert, orders the queue by risk and drafts a conclusion that references the source of each fact.
Data
AML system, Core banking, 5–10 thousand closed alerts.
Pilot format
A retrospective run on closed alerts.
Anti-fraud, AML

Linked-account network analysis

A graph of linked accounts for detecting money mules and transit schemes.

Task
Money mules, transit schemes and fraud networks cannot be detected by analysing transactions one at a time.
Functions
Builds a graph of links between accounts, devices and counterparties, and a behavioural model scores the customer’s sequence of actions. New rules are tested on the transaction flow twin.
Data
Transactions, mobile banking session logs, history of fraud cases.
Pilot format
A retrospective back-test on an anonymised 3–6 month dataset against the current rules.
Market data
  • After moving to its own behavioural models, Revolut increased fraud detection recall by 64.7%. NVIDIA ↗
  • According to Swift, a model trained on data from 13 banks detects fraud twice as effectively as a single-bank model. Swift ↗

Analytics and reporting on DWH data

Finance, risk, management board

“Ask your data” agent

An answer to a plain-language question from DWH data, with the SQL query.

Task
Answering a manager’s non-standard question takes the BI team several days.
Functions
Takes a plain-language question, for example “how did the SME portfolio in the western regions change this quarter?”, builds SQL on the data marts and returns a table, a chart and a comment. Access rights are inherited from the DWH.
Data
Data warehouse (DWH), the bank’s metrics glossary.
Pilot format
One subject area, the loan portfolio, with 100 benchmark questions.
Finance, reporting

Regulatory and management reporting agent

A draft reporting pack for the board and the NBU with variances explained.

Task
Monthly board packs and NBU reports are assembled by hand from dozens of files.
Functions
Drafts the pack with commentary on variances, such as growth in non-performing loans, reconciles figures across sources and flags discrepancies.
Data
DWH, general ledger, report templates.
Pilot format
One management report run in parallel with the manual process for two cycles.
Credit risk

Credit analyst agent

A draft credit memo for SME and corporate borrowers.

Task
Analysing financial statements and preparing a credit memo for SME and corporate borrowers takes several days.
Functions
Extracts figures from the borrower’s statements, calculates ratios, gathers data from registers and court records and drafts the memo. The agent assigns no score, and the credit decision is made by an authorised employee.
Data
Client financial statements, memo template, credit policy.
Pilot format
30–50 closed credit files, on which the agent’s draft is compared with the actual memo and preparation time is measured.

Search, processing and review of bank documents

All departments

Policy and NBU regulation assistant

A single search across internal policies and NBU regulations.

Task
Branch and compliance staff search hundreds of policies for answers by hand.
Functions
Answers questions on policies, procedures and regulations with a quotation of the clause, tracks changes in regulations and shows which internal documents need updating.
Data
Policy corpus, knowledge base, NBU regulations.
Pilot format
An isolated assistant for 30–50 employees.
Back office

Incoming document processing agent

Recognition of incoming documents and transfer of fields to the target system.

Task
Staff key credit files, powers of attorney, statements, requests from public authorities and scans into systems by hand.
Functions
Recognises the document type, extracts fields, checks the pack is complete and passes structured data to the target system. An employee reviews fields recognised with low confidence.
Data
Archive of scans and electronic documents, field dictionary.
Pilot format
2–3 document types.
Legal, compliance

Contract and internal policy review agent

Flagged risky contract terms and suggested wording for the lawyer.

Task
Approving contracts and new internal documents goes through several rounds of comments.
Functions
Compares the document with standard forms, bank policies and NBU requirements, flags risky terms and suggests wording, the final text of which the lawyer approves.
Data
Standard forms, archive of approved contracts, internal policies.
Pilot format
One contract type, procurement or collateral, measuring time to approval and the share of comments found by the agent.

Operational metrics, processes and service quality

HR, heads of business lines

Department and branch KPI agent

Plan versus actual by branch and a weekly review for the manager.

Task
KPIs are collected by hand from several systems with a delay of weeks, which makes branches hard to compare.
Functions
Calculates plan versus actual for sales, service and processing times, explains variances and prepares a weekly review for the manager. Employees see their own figures in the same form as their manager.
Data
CRM, case management system, sales plans, org structure directory.
Pilot format
One branch network or contact centre.
Operations, contact centre

Process mining and workload planning

The actual process flow from logs and a staffing forecast.

Task
The bank cannot see at which back-office stages time is lost or how many staff are needed at peak.
Functions
Reconstructs the actual process flow from logs, shows loops and idle time and forecasts workload and staffing needs for the week or month. It is used together with the credit pipeline twin.
Data
BPM logs, case management systems, shift rosters.
Pilot format
One back-office process.
Contact centre

Service quality agent

Review of all contact-centre interactions against the quality checklist.

Task
Quality control reviews a sample of 1–3% of calls and chats.
Functions
Reviews all interactions against the bank’s checklist for accuracy of information, script compliance, complaints and contact reasons, and produces a team summary and training examples.
Data
Call recordings, chats, quality checklist.
Pilot format
Interactions on one line, comparing the agent’s scores with supervisors’ scores.

Task, time and budget accounting for IT projects

PMO, IT team leads

Time tracking agent

A draft timesheet compiled from worklogs, the calendar and commits.

Task
Timesheets are filled in at month end, so some hours are not booked to a project or cost code.
Functions
Collects Jira worklogs, calendar meetings, commits and service desk tickets and allocates hours by project and CAPEX / OPEX cost code. The timesheet, with gaps flagged, is written after the employee confirms it.
Data
Jira / YouTrack, Exchange / Outlook calendar, Git, service desk, cost code directory.
Pilot format
One team or department in parallel with the current process.
Finance (controlling), PMO

Project cost and budget agent

Actual project costs, a cost-at-completion forecast and invoice reconciliation.

Task
The actual cost of a project becomes known a month late, and contractor invoices are checked by sampling.
Functions
Calculates costs from confirmed timesheets and rates, compares them with the budget and forecasts cost at completion. Hours on contractor invoices are reconciled with logged tasks, and discrepancies are flagged for review.
Data
Timesheets (the time tracking agent or the existing system), contractor rates and contracts, invoices, budget, schedule.
Pilot format
3–5 projects in parallel with a manual check of invoices.
PMO, CIO

Project status and risk agent

A weekly RAG rating for a project with the reasoning and a list of risks.

Task
Status reports are written by hand, so the risk of a missed deadline becomes visible late.
Functions
Gathers progress, overdue tasks, blockers and dependencies from Jira, Confluence and meeting notes, proposes a RAG status with the reasoning and drafts the steering committee pack. The PM approves the pack.
Data
Jira, Confluence, meeting notes, schedule, risk register.
Pilot format
5–10 projects over several weekly reporting cycles.
CIO, PMO, HR

Portfolio and resource twin

A team utilisation forecast and change scenarios for the project portfolio.

Task
New projects start without an assessment of available staff, and one slipped deadline affects dependent projects.
Functions
Models the portfolio on the same principle as the credit pipeline twin, from teams, roles, utilisation from timesheets, backlog estimates and dependencies. Forecasts staff shortfalls for the quarter and calculates the effect of a new project, a slipped deadline or a contractor leaving.
Data
Timesheets, task estimates, org structure and roles, leave calendar, portfolio plan.
Pilot format
One IT department.
PMO, team leads, service desk

Task triage and assignment agent

A breakdown of a request into sub-tasks and a suggested assignee with the reasoning.

Task
New requests are allocated in meetings to whoever is available, so team workload is uneven.
Functions
Finds duplicates and breaks a request or epic into sub-tasks with acceptance criteria. Suggests an assignee based on component, history of similar tasks, workload from time tracking and the portfolio twin, and leave, and the team lead makes the assignment.
Data
Work items, components and history from Jira / YouTrack, skills matrix, leave calendar, timesheets.
Pilot format
One team or service desk queue in suggestion mode.
CIO, PMO, heads of business lines

Delivery analytics agent

Delivery metrics from Jira in answer to plain-language questions.

Task
Team throughput and release dates are calculated through a manual export from Jira to Excel.
Functions
Calculates task cycle time, throughput and waiting time between stages. Answers questions such as “why is project X behind?” with links to the work items and forecasts the completion date as a range from the team’s historical throughput.
Data
Status history in Jira / YouTrack, timesheets, project schedule, team register.
Pilot format
One IT department’s portfolio over 6–12 months; back-test of the date forecast on completed projects.
Market data

Automated reconciliation, period close and budget control

Accounting, procurement

Procure-to-pay and invoice agent

Three-way matching of invoice, purchase order and certificate with a discrepancy queue.

Task
Invoices, purchase orders and certificates are matched by hand, so duplicates and amount discrepancies are found by sampling.
Functions
Recognises invoices and certificates and matches price, quantity and terms against the purchase order and contract, and for state-owned banks also against the Prozorro contract. Discrepancies enter a queue with an explanation, and the accountant posts matched documents.
Data
ERP procurement and accounts payable modules, contract register, invoices and certificates, supplier master data.
Pilot format
One invoice stream, for example from IT suppliers, in parallel with manual processing.
Accounting, finance

Period-close agent

A month-end close checklist with reconciliations and draft accruals.

Task
The month-end close consists of dozens of manual reconciliations whose status is not recorded in a system, so errors are found after closing.
Functions
Runs the close checklist with owners and reconciles sub-ledgers with the general ledger and intercompany balances. Draft accruals include the calculation and sources, and the accountant approves the posting.
Data
ERP general ledger and sub-ledgers, close procedure, prior-period reconciliations.
Pilot format
Two closes in parallel with the current process.
Finance (controlling), heads of departments

Budget control agent

A spend request checked against the remaining cost-centre budget.

Task
Overspend becomes visible in the monthly report, and plan-vs-actual variances are explained by hand.
Functions
Checks a spend request against the remaining cost-centre budget and existing commitments and alerts the approver before approval. Explains plan-vs-actual variances by line, and for IT projects takes actual costs from the project cost and budget agent.
Data
Budgets, commitments and actuals in the ERP, purchase requests, cost-centre register.
Pilot format
3–5 cost centres.
Procurement, facilities, every department

ERP assistant for employees

A draft purchase request from a text request, and its approval status.

Task
Employees send requests and status questions by email to accounting and procurement because the ERP interface is complex.
Functions
Takes a plain-language request such as “we need 5 laptops for the new team” and drafts the purchase request with cost codes and a catalogue supplier. Access rights come from the ERP, and the employee submits the request.
Data
Requests, procurement catalogue and statuses in the ERP, cost-code registers, role-based access model.
Pilot format
One request type, for example IT equipment or office supplies, in one department.
Market data
  • Oracle added 22 agentic applications to Fusion for finance, HR and supply chain, which run inside the transactional systems. TechRadar ↗

Meeting preparation and CRM data maintenance

Corporate banking, SME, private banking

Relationship manager assistant

A client briefing before the meeting and structured notes and tasks after it.

Task
The relationship manager gathers data from several systems before a meeting, and the outcome often does not reach the CRM.
Functions
Before the meeting, compiles a briefing with products, balances, turnover, open requests and register changes. After it, structures the notes, creates tasks and opportunities and drafts a follow-up email, and the relationship manager confirms the entry.
Data
CRM, products, balances and turnover from Core banking, request history, calendar, state registers.
Pilot format
10–20 SME or corporate relationship managers.
CRM owner, operations, compliance

CRM data quality agent

A correction queue for CRM duplicates and mismatches with Core banking.

Task
The CRM holds duplicates, empty and stale fields and stalled deals, and its data diverges from the registers.
Functions
Finds duplicates, empty and stale fields, stalled deals and mismatches with Core banking and the Unified State Register, EDR. Each correction with the source of the value goes to the owner’s queue, and an employee changes the CRM records.
Data
CRM, Core banking, the EDR and other state registers, the bank’s data quality rules.
Pilot format
One client segment.
Retail, SME, marketing

Next best action agent

A suggestion to the manager based on events in the client’s transactions.

Task
Falling turnover, large incoming payments or a maturing deposit or loan do not stand out in the general flow of transactions.
Functions
Tracks transaction events and proposes a next step from the approved product range to the manager with the reason. A mass campaign is first tested on the customer twin. The agent does not set terms or rates.
Data
Transactions, product portfolio, campaign and feedback history, product rules.
Pilot format
A back-test on historical data and shadow mode on one segment.
Contact centre, branches

Call and request summary agent

The summary, topic and next step of a conversation in the client record.

Task
The operator writes the summary and topic by hand, so the CRM history is incomplete and some complaints are not registered.
Functions
Transcribes calls and chats, writes the summary, agreements and next step to the client record, classifies requests and routes complaints. The operator edits the draft before saving, and the service quality agent assesses the conversation.
Data
Call recordings, chats, CRM, topic and complaint taxonomy.
Pilot format
One contact-centre line.
Market data
  • Bank relationship managers call themselves “data entry clerks” because of their CRM work, and their turnover at many banks is 15–35%. McKinsey ↗
  • Salesforce released an agent for bankers that structures meeting notes and creates tasks in the CRM. Salesforce ↗

Control and audit

Monitoring, validation and audit of bank models

CRO, model validation

Model registry and drift monitoring

An automatically generated model validation report.

Task
The regulator expects roughly quarterly validation of model accuracy, robustness and fairness.
Functions
Keeps a registry of models and AI systems with their risk class, tracks data and performance drift, runs regression tests on benchmark data and produces the validation report.
Data
Model inputs and outputs, realised defaults and confirmed fraud cases.
Pilot format
Connecting the 1–2 most critical models, usually retail scoring.
CRO, credit risk

Challenger models and shadow scoring

An alternative model in shadow mode on live applications.

Task
The bank does not know whether the current model remains optimal, and replacing a model directly in production carries risk.
Functions
Builds an alternative model and runs it in shadow mode on live applications while the current system makes the decision. The models are compared on Gini and AUC using realised defaults.
Data
Application and default history, current scorecard.
Pilot format
One retail product in shadow mode.
Internal audit

Audit trail for AI decisions

Reproduction of model decisions at the auditor’s request.

Task
During an inspection, the bank must reproduce why a model made a particular decision.
Functions
Logs the model version, input data, output and decision factors for each request and produces an explanation for the customer and a pack for the auditor.
Data
Stream of model decisions.
Pilot format
Reproducing 20 random decisions at internal audit’s request.

Platform inside the perimeter

A shared platform for launching and controlling agents

CIO/CTO

Agent orchestration platform

Shared MCP connectors, memory and task routing between agents.

Task
Each chatbot has its own integrations and memory, and one agent cannot hand a task to another.
Functions
Contains a registry of agents and tools, shared memory, vector search and connectors based on the Model Context Protocol standard. The platform routes tasks between agents, for example the alert investigation agent calls the document processing agent.
Data
List of systems to connect: Core banking, CRM, DWH, the AML system, Jira.
Pilot format
Two agents from the catalogue on a shared platform.
CRO, compliance

Action approval loop and agent policies

An approval queue for agent actions with consequences for a customer or the bank.

Task
An agent with write access to the bank’s systems needs limits: what it does on its own and what it only proposes to an employee.
Functions
Sets the permitted tools, limits and read-only mode for each agent at platform level. Actions with consequences enter an approval queue with the supporting evidence, and the employee’s decision is recorded in the audit trail.
Data
The bank’s role model, delegation of authority matrix.
Pilot format
Policies for one agent with the right to act, tested against scenarios that try to bypass the restrictions.
CIO, AI governance

Agent evaluation and observability

Quality tests and cost tracking for agents.

Task
The bank does not measure the quality of an agent’s answers or the cost of a request.
Functions
Traces agent steps, runs benchmark tests before a model or prompt change, and calculates accuracy and hallucination metrics and cost per request and per department. Results are recorded in the model registry.
Data
Agent logs, benchmark questions from departments.
Pilot format
One existing agent or chatbot.

Model deployment on the bank’s hardware

CIO

Private Fabric: an AI platform inside the bank’s perimeter

An AI platform on the bank’s hardware with an optional air-gapped mode.

Task
The bank’s GPU servers, if it has any, run separately, with no shared scheduler or model registry.
Functions
Deploys the control node, scheduler, model registry, console and GPU nodes on the bank’s hardware and pools scattered GPUs into one cluster. The platform can run in air-gapped mode with no outbound connections.
Data
Description of existing or planned GPU hardware, the security team’s network segmentation requirements.
Pilot format
Deployment on 1–2 of the bank’s servers; the bank’s security team verifies there are no outbound connections.
CIO/CTO

Private inference for open models

Open models through an OpenAI-compatible API, billed by GPU time.

Task
Agents need language models, per-token fees to external providers grow with volume, and the data leaves the perimeter.
Functions
Serves text and multimodal open models from Hugging Face through an OpenAI-compatible API on vLLM, so the bank’s existing applications connect without code changes. Billing is by GPU time.
Data
List of tasks and expected load in requests per day.
Pilot format
One model for the policy assistant or the alert investigation agent.
CRO, Data Science

Fine-tuning and federated learning

Models fine-tuned on the bank’s products and terminology.

Task
General-purpose models do not know the bank’s terminology and products, and data from branches or partners cannot be pooled in one place.
Functions
Fine-tunes open models on the bank’s data inside its perimeter and runs federated learning on NVIDIA FLARE across several sites without moving raw data. The weights and adapters belong to the bank.
Data
A training set for a specific task, for example closed alerts, classified documents or transactions.
Pilot format
Fine-tuning a model for one task from the catalogue; comparison with the base model on a hold-out sample.
Market data
  • PNC CEO Bill Demchak said: “We will have our own GPU compute.” Evident ↗

Custom PM and ERP application development

CIO, PMO

PM system built around the bank’s processes

An application for approving IT requests under the bank’s procedures.

Task
Change requests arrive by email, the portfolio is compiled by hand, and Jira covers delivery of the work but not its approval.
Functions
Takes a request through estimate, prioritisation and committee approval to delivery and benefit assessment. The application drafts the business case, detects duplicates, uses the time tracking, project status, task assignment and delivery analytics agents and syncs with Jira.
Data
Change approval procedures, Jira / YouTrack, AD, department and budget registers.
Pilot format
A first version for one request stream, in parallel with the current process.
Finance, procurement, facilities

ERP modules and workspaces

Modules for requests, contracts and approvals integrated with the ERP.

Task
The contract register, requests and approvals are kept in spreadsheets, and data is re-keyed into the ERP.
Functions
Modules for purchase requests, the contract register and approvals, a supplier portal and an accountant’s exceptions queue write data to the ERP through its API. The document processing, procurement and budget control agents and the ERP assistant are built into them.
Data
The bank’s ERP API, contract register, documents, role-based access model.
Pilot format
One module, for example contract approval.
CIO, internal system owners

AI features in existing applications

Search, summaries and form auto-fill in the bank’s existing systems.

Task
Staff copy data from documents into forms by hand and search for information across several screens.
Functions
Integrates document search, a summary of a client or case history, form auto-fill from a document and request classification into existing applications. Models are called through the orchestration platform under the bank’s agent policies.
Data
The internal system’s API or database, the documents users work with.
Pilot format
One feature in one application for one user group.
Market data
  • Gartner predicts that by the end of 2026, 40% of enterprise applications will feature task-specific AI agents. In 2025 the share was under 5%. Gartner ↗

Discuss a pilot for the selected solution.

Discuss a pilot