Agentic business banking

Banks face two pressures at the same time. Direct-to-company fintechs and ERP platforms are taking the daily business banking workflow away. General-purpose AI assistants are starting to take the channel that used to belong to the bank. Both pressures call for the same response: agentic capability inside the bank’s own perimeter, in 2026.

The first pressure comes from direct-to-company fintechs and ERP-platform players that have moved fast on the daily business banking workflow. Payments, foreign exchange, cash visibility, collections. Small and mid-sized companies have been quietly switching parts of that workflow away from their bank for several years now. Each workflow that leaves is a daily relationship the bank no longer owns.

The second pressure is newer. Small and mid-sized companies are starting to use general-purpose AI assistants to plan, decide, and act on their financial life. Comparing accounts, checking rates, drafting payment instructions, asking for guidance. Independent research published in early 2026 puts monthly use of generative AI for financial tasks at roughly a quarter of consumers, and the same patterns are showing up in the small-business segment. If a third-party AI assistant becomes the company owner’s first interface for money decisions, the bank moves one click further from the relationship.

Both pressures call for the same response. Banks need agentic capability inside their own perimeter in 2026.

Agentic Business Banking is TreasurUp’s answer. A composable banking platform that combines proven modules and services, an Intelligence Engine grounded in domain AI, and approval-gated agents that banks deploy under their own brand.

The model rests on three layers, all operating today in some form.

The first layer is the Composable Banking Platform. From atomic services to full solution suites and complete business banking channels. Web, mobile, embedded banking, APIs. Hosted by TreasurUp, deployed in the bank’s cloud, or run in a hybrid shape. Banks enter at any level and expand from there.

The second layer is the Intelligence Engine. Domain AI as a shared platform capability. Rule and optimization engines, smart insights, AI and machine learning forecasting, and a natural-language query layer that lets company users and bank staff ask plain-language questions and get sourced, calculated answers back.

The third layer is Agentic Business Banking. Agents that support the main areas of business banking for company clients and internal bank users. Workflows for cross-border payments, liquidity management and optimization, FX risk management, collections and receivables, cash flow forecasting, and front-office sales analytics. All under approval-gated execution.

The architecture is flexible by default. Banks choose deployment shape: managed SaaS hosted by TreasurUp, single-tenant in the bank’s cloud, or hybrid. Bring-your-own-LLM applies across all three. Governance, audit trails, and human approval gates are part of the platform spec.

This whitepaper sets out the platform vision, the three layers, the 12-month deliveries, the architecture, the governance model, and how banks engage.

The shift: from banking software to banking intelligence

TreasurUp’s market is shifting. For most of the last two decades, banking technology meant transaction execution. Move money. Settle it. Reconcile it. Show it. The unit of delivery was a click that became a transaction. The next decade is going to be different. The unit of delivery is moving from clicks to outcomes, and the layer that produces those outcomes is intelligence, not execution.

Banks have always been the trust layer between companies and the financial system. Company users want more than execution venues. They want judgment, context, and a partner that understands their business model and related finance. AI agents change the unit of delivery here. The question for every bank is whether those agents will sit inside the bank’s brand or inside someone else’s.

Agentic Business Banking is the answer. A platform model where banks deploy modular daily business banking capabilities and approval-gated AI agents inside their own business banking channels. Branded as theirs. Governed by their policies. Running on infrastructure they control.

Three principles run through this vision.

Agents augment. They do not execute autonomously on material decisions. Approval-gated execution is the design point, not a phase to be outgrown.

Banks own the agent. Branding, data, model choice, deployment topology, all configurable per bank.

Governance is a feature. The EU AI Act, DORA, model risk, audit trails, and explainability are part of the platform spec, not afterthoughts.

Why now: three shifts make 2026 the year to act

Three shifts have converged. Each one on its own would justify movement. Together they close the window for cautious waiting.

Banks face hyper-competition, quite suddenly

For the last few decades, banks competed but could still rely on dominant products and the size of their balance sheets. That posture no longer holds. Direct-to-company fintechs, big-tech entrants, brokers, and white-label banking providers are creating serious churn at most incumbent banks. The small and mid-sized business segment is the most exposed.

The threat is workflow-shaped. It rarely arrives as a wholesale switch of bank accounts. It arrives one workflow at a time. The company starts sending international payments through a non-bank provider because it is cheaper and faster. It runs FX conversion through a multi-currency wallet because the bank’s spreads are wide and the user experience is dated. It pulls cash flow visibility from its accounting software because the bank’s portal does not surface the same picture. Each of those workflows used to live with the bank. Most of them no longer do. The relationship erodes one piece at a time.

Independent research published in late 2025 quantifies the picture. Satisfaction with bank merchant services sits at 15 percent among small merchants and 22 percent among mid-sized merchants. Around 40 percent of small and mid-sized merchants are considering a shift to specialist payment technology providers. The opening is closing fast, but it has not closed yet: 66 percent of merchants still prefer their traditional bank for financial services overall.

Figure 1. Bank merchant satisfaction by segment.  Chart redrawn in TreasurUp branding from data published in Capgemini Research Institute, World Payments Report 2026, September 2025.

The battle for the primary channel is moving into AI

In meetings with banks worldwide over the last six months, the question has shifted. A year ago, banking leaders asked whether AI assistants were a real channel risk for their small and mid-sized company clients. Today they ask how fast it is moving and what to do about it. Independent research helps frame the answer.

Small and mid-sized company owners are starting to use general-purpose AI assistants for financial tasks. Research published in 2025 and 2026 found that almost a quarter of consumers use generative AI tools for financial tasks at least monthly, with the top use cases being understanding products, comparing options, and getting financial guidance. The same behavioural patterns appear among small-business owners. Owners of small and mid-sized companies behave much more like retail customers than like the centralized finance functions of large enterprises when it comes to financial decisions.

The implication for banks is direct. If the small-business owner gets used to asking a third-party AI assistant questions about their company finances, that assistant becomes the channel through which the daily relationship runs. The bank app, the relationship manager, the call centre, all move one step further from the owner’s attention. The relationship has not been switched. It has been mediated. In the medium term, an external AI assistant could move from suggesting actions to initiating them on the company’s behalf. Surveys put the willingness to use a third-party AI financial agent at around 57 percent of customers if their own bank does not offer one. At that point the bank is a back-end product provider rather than a relationship owner.

Three responses are possible for any individual bank. Wait and see how the AI channel develops, adapt to being a product provider behind someone else’s AI interface, or compete by offering the company’s AI experience inside the bank’s own brand. Banks taking the third path need agentic capability inside their own perimeter, and they need it before the habit of using a third-party assistant becomes settled.

Figure 2. Three strategic postures for banks responding to the agentic shift.  Framework adapted from McKinsey & Company, How gen AI agents threaten retail banks’ customer relationships, April 2026. “How TreasurUp helps” row added by TreasurUp.

Regulatory clarity is finally arriving

Vague compliance risk is no longer a justification to wait. The EU AI Act’s risk-based framework, DORA’s operational resilience requirements, and updated model risk guidance from supervisors give banks a defined path to deploy AI in regulated workflows. The questions have shifted from “is this allowed” to “what controls and documentation does the supervisor expect.” For the agentic use cases described in this paper, the answer is documented and deployable.

The cost of waiting is asymmetric. Independent analysis suggests early adopters of agentic AI in banking could open a gap of around four percentage points of return on tangible equity relative to slow movers, capturing years of productivity benefit before those advantages get competed away. Banks that ship credible agentic business banking experiences in 2026 set the bar their competitors will spend 2027 catching up to.

The three-layer platform

Figure 3. TreasurUp’s three-layer platform. Source: TreasurUp, 2026.

Layer 1: The Composable Banking Platform

The composable foundation. From atomic services to full modules to complete channel outsourcing. All the components a bank needs to build and run a competitive business banking channel: web, mobile, ERP and TMS, and APIs.

Solution suites range from liquidity management, payments, collections and receivables, foreign exchange, and API developer portal components, through to customer onboarding, communications, entitlements, and workflow.

Banks enter at any level. They can buy a full suite, pick specific modules, or consume individual services via API. That flexibility shortens sales cycles, reduces perceived risk, and creates land-and-expand paths.

Layer 2: The Intelligence Engine

The differentiator. Domain AI lives here. Not as a separate product but as a shared capability that makes every module smarter.

Rule and optimization engines handle payments, collections and receivables, ERP workflow, and liquidity management, plus onboarding and bank and client workflow support. They combine deterministic rules with AI interpretation of client and bank input.

Smart insights and analytics turn data into recommendations, not dashboards. AI and machine learning cash flow forecasting predicts liquidity with multi-scenario modelling. Natural-language query layers let company clients and bank staff ask questions in plain language and get sourced, calculated answers back.

The positioning is domain AI for business banking. Not generic AI bolted onto banking software. The moat is the combination of large language model reasoning with proprietary domain logic: risk and liquidity management, daily payments and collections, onboarding, and workflow depth that comes from a decade of building this product with banking-background practitioners.

Layer 3: Agentic Business Banking

The realization. AI agents that do not just provide insights. They prepare actions, with human oversight on every material step.

Company-side agents identify payments, receivables, liquidity optimization, and FX risk needs. They surface anomalies. Bank-side agents validate transactions, monitor risk limits, and brief relationship managers before a client calls.

Approval gates are architecture. Agents propose. The company user or the bank user disposes.

TreasurUp’s AI solutions

The agent roster covers the workflows where banks are most exposed to fintech and AI-assistant competition.

Figure 4. TreasurUp’s agent roster. Company-side agents serve the small or mid-sized company. Bank-side agents serve relationship managers and the middle office. Approval-gated execution sits between them. Source: TreasurUp, 2026.

Cross-border payments agent. Prepares payment instructions, identifies routing options, and flags exceptions. The company user approves. Execution flows through the bank’s existing channels.

Collections and receivables agent. Matches incoming flows, identifies reconciliation gaps, and proposes follow-up actions on overdue items. Bank staff and company users see the same workflow, with role-appropriate visibility.

Liquidity management and optimization agent. Watches positions across accounts, currencies, and entities. Identifies options for sweeps, transfers, or investment actions inside the company’s policy framework. The company user keeps the decision.

FX risk management agent. Identifies exposures from confirmed and forecasted cash flows, prepares hedge proposals against company-defined policy, and presents the reasoning the company user needs to approve or adjust.

Cashflow forecaster. Multi-scenario forecasting that ingests ERP data, bank data, and market context. The forecast is sourced and explainable, not a black-box number.

Copilot for Business Banking. A natural-language interface for company clients. Plain-language questions, sourced and calculated answers, with the underlying data accessible in one click. “What is my expected cash position next Friday?” “Which suppliers are unpaid past 60 days?” “Should I sweep my surplus euro balance into the overnight account?” The company owner asks in language they already use. The agent answers from the company’s own data.

Client reporting and sales analytics agent. A bank front-office capability. Prepares pre-call briefings for relationship managers: position changes, upcoming maturities, missed opportunities, anomaly flags. The kind of context that used to live in someone’s head, or nowhere at all.

Trade anomaly detector. A middle-office agent that monitors transaction patterns and surfaces deviations to the right reviewer. Reduces manual exception handling without removing human judgment from the decision.

Sales intelligence radar. Aggregates client signals across the bank: behaviour changes, product usage shifts, upcoming events. The output is a prioritized list of relationship-manager actions, not another dashboard.

For company clients, the shape of the day changes. Payment, FX, and liquidity decisions stay with the company. The data work that leads up to those decisions mostly does not.

For relationship managers, the bank-side agents make a measurable difference. Independent research finds that banks deploying agentic AI in the frontline report between three and 15 percent higher revenue per relationship manager and 20 to 40 percent lower cost to serve. In many banks, relationship managers spend only 25 to 30 percent of their time in client dialogue. Agentic AI returns 10 to 12 hours per week to each banker, with most of that time going back into the conversations that drive the relationship. That is the productivity gain banks can put in front of their company clients in 2026.

Figure 5. A typical relationship manager week, before and after agentic AI.  Chart redrawn in TreasurUp branding from data published in McKinsey & Company, Agentic AI is here. Is your bank’s frontline team ready?, December 2025.

Deployment

The architectural commitment is flexibility. TreasurUp meets banks where they are: as a fully managed SaaS hosted by TreasurUp, deployed inside the bank’s own cloud tenancy, or in hybrid shapes that split the stack. The platform, Intelligence Engine, and agents are the same across all shapes. What varies is where they run and which model powers them.

Four deployment choices, bank’s call.

Managed SaaS hosted by TreasurUp. TreasurUp hosts the full stack: platform, Intelligence Engine, agents, foundation models, and data integrations. Fastest to deploy and the lowest operational lift on the bank side. The right option for banks that want capability quickly without standing up new cloud operations, and for early rollouts that may move to a bank-hosted shape later.

Single-tenant in bank cloud. The full TreasurUp platform deploys to the bank’s own cloud subscription, in the bank’s region of choice. Bank data never leaves the bank’s cloud. Bank security teams own the perimeter. TreasurUp ships software; the bank operates it under shared responsibility.

Hybrid. Some banks want the platform managed by TreasurUp but with sensitive data and inference inside their boundary. Hybrid deployments split the orchestration layer (TreasurUp-hosted) from the data and model layer (bank-hosted). The path for banks that want speed without compromising data residency.

Bring-your-own-LLM. An option that applies across the three shapes above. The bank picks the foundation model: a leading commercial model, a bank-private model, or a multi-vendor mix. TreasurUp’s orchestration layer integrates with the bank’s chosen model endpoints. Banks that have negotiated their own AI vendor terms or built their own internal models keep that posture.

Figure 6. Four deployment choices for Agentic Business Banking. Source: TreasurUp, 2026.

Why this matters for procurement. Managed SaaS is the right fit for many banks that want speed, simplicity, and TreasurUp’s operational expertise. For banks with stricter data residency, model risk, or examination access requirements, bank-tenanted and hybrid deployments clear most of the questions that make AI deals in financial services difficult. Data residency, model risk, regulatory examination access, business continuity, and vendor concentration are all addressable in the bank’s existing control framework. In every deployment shape, TreasurUp is a platform vendor, not a third-party processor of bank or company data.

The composable layer underneath stays consistent across deployments. The same five suites, the same modules, the same agents, the same approval flows. What varies is where it runs and which model powers it. Not what it does.

This architecture is also why TreasurUp can ship agentic capabilities in months, not years. The platform exists. Modules are already deployed at multiple banks. Agentic features extend the existing stack rather than requiring a new one.

Governance and trust

Approval-gated execution

Every agent action with financial, regulatory, or accounting impact passes through an explicit human approval. The approval is logged with the agent’s reasoning chain, the input data, the alternatives considered, and the decision. The platform does not expose autonomous-execution APIs for material actions. Approval-gated execution is structural, not configurable.

Figure 7. Approval-gated execution. Every agent action with financial, regulatory, or accounting impact flows through a logged human approval before execution. Source: TreasurUp, 2026.

EU AI Act and DORA alignment

The platform is designed for EU AI Act and DORA compliance from the ground up. Components are classified against the EU AI Act’s risk tiers and ship with the documentation banks need for high-risk system obligations. Agent and module components are catalogued and classified by criticality, support TLPT-style resilience testing, and stream incidents to the bank’s existing SIEM via standard interfaces.

Model risk management

Every agent has a registered specification covering model, version, training data lineage where applicable, inputs, outputs, decision rights, and fallbacks. Banks run independent validation against agent outputs using sample logs, gold-standard datasets, and the monitoring tooling included in the platform. Drift, accuracy, and behaviour-change alerting run continuously, with notifications to the bank’s model risk team when an agent’s behaviour changes. Banks can disable, override, or roll back any agent without TreasurUp involvement.

Audit trails and explainability

Every agent interaction is captured end-to-end: the user query, the data retrieved, the reasoning, the proposed action, the approval, the executed action, and the post-action check. Logs are bank-owned, exportable, and retained per the bank’s policy. Agents present reasoning in plain language at the point of decision, not buried in logs. When the FX Risk Management agent surfaces a hedge proposal, the company user sees the position concerned, the policy constraints applied, the market context, the alternatives ranked, and why this proposal scored highest. Explainability is a user-experience requirement, not a compliance afterthought.

Data protection

GDPR and banking confidentiality are designed in from the start, not bolted on. In bank-tenanted and hybrid deployments, business banking data stays within the bank’s data sovereignty boundary. In managed SaaS deployments, data is processed under the contractual and technical controls that apply to a regulated banking vendor.

Human-in-the-loop is permanent

This is worth restating, because the temptation, when agents perform well, is to remove the human approval gate. The product roadmap does not do that. Where automation increases over time, it does so through better agent quality and tighter scope, not through removing humans from material decisions.

What we will explicitly not offer

Four explicit non-goals. These boundaries are not temporary. They are the design choices that make the rest of the vision deployable in regulated banking.

No autonomous trade execution. Agents propose. Humans approve. The trade ticket flows to the bank’s execution venue through standard channels. This is the strongest signal of trust posture.

No autonomous limit changes. Agents do not modify counterparty limits, product limits, or policy thresholds. These are bank governance decisions that stay with bank governance.

No agent-to-agent transactions across institutions. Cross-institution agentic flows, where a company’s agent communicates with a counterparty’s agent and a bank’s agent end-to-end, are not in scope. The interoperability question is real, but it is a 2027 question. In 2026, agents work within the bank-and-company boundary.

No autonomous credit decisions. Agent insights inform bank credit and risk teams. The decisions stay with the bank’s credit and risk frameworks.

Engagement

There is no single starting point. Banks join Agentic Business Banking from different positions and on different timelines.

For banks that already run TreasurUp modules, the agentic layer extends what is already there. It is a configuration and rollout exercise, not a new procurement. The lead modules are the FX Risk Management agent and the Cashflow Forecaster, both ready to deploy on top of existing FX and liquidity suites. Live agent deployments are targeting the second half of 2026.

For banks new to TreasurUp, the standard platform deployment plus agentic configuration moves through procurement, security review, and model risk validation in parallel from week one. End-to-end time from contract to first company-user-facing agent is in the order of six months for the lead modules.

A small number of banks shape the roadmap as design partners, focused on a specific segment of company clients (early-stage business, established small business, mid-market) or on a specific internal bank function. Two design partner slots open per quarter through 2026.

A note on what comes next

The agentic shift in business banking is not a far-out story. Small and mid-sized company owners are already using AI-native tools that sit outside the bank perimeter. Each one of those flows is a daily relationship moving away from the bank that should own it.

The architecture is composable, intelligent, and agentic, in that order. Composable because banks need to enter at the level that fits their estate, not buy a monolith. Intelligent because domain AI, grounded in banking workflow, is what makes the agents useful instead of generic. Agentic because the unit of work in business banking is moving from clicks to outcomes, and the bank that hosts those outcomes inside its brand earns the daily relationship.

If your bank is thinking about how to bring agentic capability inside its own perimeter in 2026, this is the conversation to have.

References

This whitepaper draws on independent research from leading consultancies and primary research providers. All data points referenced inline are paraphrased; readers are encouraged to consult the original sources linked below.

1. McKinsey & Company,  How gen AI agents threaten retail banks’ customer relationships, April 2026. Cites McKinsey Global Banking Annual Review 2025, October 2025.

2. Capgemini Research Institute,  World Payments Report 2026, September 2025. Survey of 2,600 merchants across 15 markets.

3. McKinsey & Company,  Global Banking Annual Review Survey, May 2025 (n = 30,021).

4. Boston Consulting Group,  How retail banks can put agentic AI to work, March 2026. See also Agentic AI will shake up banking, shrinking global profit pools, McKinsey, November 2025.

5. McKinsey & Company,  Agentic AI is here. Is your bank’s frontline team ready?, December 2025. Survey of 406 US and Canadian bankers.

Further reading consulted but not cited inline:

• McKinsey & Company, The end of inertia: Agentic AI’s disruption of retail and SME banking, August 2025.

• McKinsey & Company, Banking’s agentic AI opportunity, December 2025.

• Deloitte, Managing the new wave of risks from AI agents in banking, 2026; and 2026 Banking and Capital Markets Outlook, October 2025.

• McKinsey & Company, Digital-led with a human touch: The next era in small-business banking, August 2025.

Published by TreasurUp B.V., Utrecht, Netherlands, 2026. For inquiries about Agentic Business Banking and partnership opportunities, contact info@treasurup.com or schedule a working session at https://treasurup.com/contact/.

Figures 1, 2, and 5 are redrawn in TreasurUp branding from data published by Capgemini Research Institute and McKinsey & Company, with full attribution shown beneath each figure. Original publications are linked in the References section and remain the copyright of their respective publishers. Figures 3, 4, 6, and 7 are original TreasurUp diagrams.