Blog
The real cost of manual treasury workflows in the age of AI agents
Manual treasury workflows cost more than labor hours. They create a compounding operations tax: delayed cash visibility, repetitive reconciliation, manual approvals, trapped liquidity, avoidable borrowing, error remediation, fraud exposure, audit burden, and slower decisions. They also carry a cost that's easy to miss until an AI initiative stalls: fragmented, manually patched workflows are exactly the kind of data environment an AI agent can't safely operate in.
That last point matters more than it sounds. An AI agent is only as good as the data infrastructure underneath it. Clean, connected data produces fast, accurate agent output. Fragmented, inconsistent data produces the same confident-looking output, except it's wrong in ways that aren't visible until something breaks. So the real question behind "how much does manual treasury work cost?" isn't just a labor question anymore. It's whether your treasury operation is structured in a way that lets an AI agent help you at all.
Key takeaways
What does manual treasury work actually cost a finance team? More than the hours it visibly takes. It's a compounding tax across seven areas: labor, delayed visibility, trapped liquidity, error and rework, fraud and control exposure, scalability, and AI readiness. Most organizations can put a rough number on the first one and never quantify the other six, which is usually where the larger cost sits.
Why can't AI just fix manual treasury workflows on its own? Because an AI agent is a multiplier of the data infrastructure it's given, not a replacement for it. Fragmented bank feeds, inconsistent entity mapping, and manually adjusted spreadsheets don't get fixed by adding an agent on top. They get automated at higher speed, which surfaces the same gaps faster and with less warning.
Where should a treasury team start if it wants to fix this? Not with a vendor purchase. Start by building a data-quality baseline for the single manual workflow costing the most today, usually reconciliation, cash positioning, or forecasting, and use that baseline to decide which use case is actually ready for an AI agent versus which one needs data work first.
See how Kyriba TAI turns treasury AI readiness into governed action: Manual treasury workflows do more than slow teams down; they limit what AI agents can safely do. Kyriba TAI is agentic AI built for treasury and finance teams that need connected data, human control, and audit-ready automation in the same workflow. With TAI, teams can query treasury data, analyze variances, generate reports, run repeatable Skills, and prepare recommendations with transparent reasoning, human-in-the-loop approvals, scoped permissions, and fully auditable action trails. Learn more here.
The manual treasury operations tax
Think of every manual workaround in treasury, the extra spreadsheet, the second login, the email approval chain, as a small, recurring tax on the business. No single instance looks expensive, but added up across a year, across every entity and every bank relationship, these efforts have a cost. And manual work is still common: TD's 2025 treasury survey found nearly 80% of treasury departments still use manual processes, with manual or fragmented financial systems cited by 36% of respondents as a top challenge, on par with macroeconomic uncertainty and market volatility. It's also not just an operations problem anymore. The 2025 EACT Treasury Survey ranked cash-flow forecasting as the single top treasury priority, with treasury technology infrastructure review or replacement also among the top priorities. Together, these findings put treasury modernization squarely on the board's agenda rather than a line item in a systems upgrade request.
The Manual Treasury Operations Tax is the total cost created by fragmented treasury workflows, across seven categories:
Labor cost: analyst time spent gathering and validating data instead of analyzing it
Visibility cost: decisions made on delayed or stale cash positions
Liquidity cost: idle cash, unnecessary borrowing, and missed yield
Error and rework cost: reconciliation breaks, duplicate payments, and the investigation that follows
Fraud and control cost: weak approval trails and slower anomaly detection
Scalability cost: every new entity, bank, or currency adding headcount instead of capacity
AI-readiness cost: the compounding cost of not being able to deploy AI agents safely because the underlying data and workflows aren't ready for them
The first six are the traditional case for treasury modernization. The seventh is the one that's changed the calculus in the last two years: manual workflows don't just cost you directly, they also block the next layer of value an AI agent could otherwise unlock.
Cost category 1: wasted analyst time
Treasury professionals often spend more time assembling information than analyzing it: logging into separate bank portals, exporting and formatting spreadsheets, reconciling statements by hand, chasing approvals, rebuilding the same report every week, and explaining variances after the fact instead of catching them in real time. Spreadsheets are still the default tool for this work.
The time cost of that dependency adds up: a survey of more than 100 major U.S. and European brands, reported by Finextra, found treasury teams wasting an average of 4,812 hours a year using spreadsheets to manage cash, payments, and accounting operations.
Treat that figure directionally given its vendor-survey origin, but it's consistent with what shows up inside most treasury operations that haven't consolidated their tools.
Manual Activity | Hidden Cost | Business Impact |
Downloading bank data | Analyst hours spent collecting data | Delayed cash visibility |
Spreadsheet consolidation | Manual validation and formatting | Higher error risk |
Approval chasing | Time spent following up with stakeholders | Payment delays |
Manual reconciliation | Investigation and rework | Slower close and reporting |
Recurring treasury reports | Repetitive preparation time | Less time for strategic analysis |
This is the category an AI agent is built to absorb directly. In a governed reasoning loop, an agent retrieves balances, forecasts, and payment data across every entity simultaneously, in one pass, with every call logged for later review. The work that used to take an analyst an entity at a time now happens in parallel, and the analyst's time shifts from data-gathering to reviewing and approving what the agent surfaced.
Cost category 2: delayed cash visibility
Treasury can't optimize what it can't see. When positions take hours or days to consolidate across banks and entities, teams default to conservative decisions: holding more cash than necessary as a buffer, delaying funding calls, or making forecasting assumptions on data that's already stale by the time it's used.
The cost of delayed visibility isn't just reporting lag. It's the cost of making funding, investment, and liquidity decisions with yesterday's information. This is also the single most-cited AI opportunity in treasury for a reason: EuroFinance research found cash forecasting is the leading cited AI use case at 38%, ahead of process automation at 32%, information summarization at 16%, and data quality and decision-making at 14%, and the same research notes that AI's potential here depends directly on data quality and system integration.
Academic research backs the underlying logic rather than just the popularity: Salas-Molina et al. found predictive accuracy is highly correlated with cost savings in daily cash-management models, which means fixing visibility is a measurable financial lever, not just an operational nicety.
Cost category 3: trapped liquidity and avoidable borrowing
Disconnected cash visibility has a direct financial cost. Idle cash sits in low-yield accounts. Balances get trapped across entities, banks, or regions that don't talk to each other. Treasury borrows short-term while cash sits available somewhere else in the organization. Short-term investment windows get missed because nobody consolidated the picture in time.
Fragmentation Problem | Potential Financial Cost |
Cash held across disconnected accounts | Excess idle cash |
Poor forecast accuracy | Larger liquidity buffers |
Delayed visibility into shortfalls | Emergency borrowing |
Slow approvals | Missed early-payment discounts |
Manual intercompany funding | Inefficient capital allocation |
An agent working a liquidity positioning workflow computes headroom per entity after minimum balances and commitments, identifies the most efficient transfer routes accounting for cutoff times and FX effects, and surfaces yield opportunities on surplus balances. Using a four-mode autonomy model as an example, this type of workflow might begin with human approval for each transfer and progress toward supervised execution within policy only after recommendations have proven consistently accurate.
Cost category 4: errors, exceptions, and rework
Manual processes don't just take time; they create more time later. Every reconciliation break, duplicate payment, missed cutoff, or unexplained variance opens a second workflow: investigation, correction, documentation, and review. Spreadsheet formula errors and inconsistent entity or account mapping compound the problem quietly, since they often surface only when something downstream doesn't reconcile.
This is where a governed AI agent can show clear, measurable impact, because reconciliation is high-volume, rule-based, and well suited to automation once the matching rules and controls have been validated. HelloFresh, for example, uses Kyriba TAI to identify hard-to-reconcile payments in seconds. Before using TAI, this work could take “ages,” according to HelloFresh’s Director of Group Treasury. The capability helps the lean treasury team spend less time creating and searching for data and more time analyzing it.
Cost category 5: fraud, controls, and audit exposure
Manual workflows create risk, not just inefficiency. Email-based approvals are hard to control and easy to route around. Manual beneficiary updates increase fraud exposure. Fragmented data makes anomaly detection harder because nothing is looking across the full picture at once. Audit trails end up scattered across inboxes and spreadsheet versions, and segregation-of-duties evidence can be difficult to reconstruct after the fact.
This isn't a hypothetical risk: AFP's 2025 Payments Fraud and Control Survey found 79% of organizations experienced attempted or actual payments fraud in 2024, and its 2026 survey found 76% of U.S. organizations experienced it in 2025, while only 17% used AI to help combat it. Checks remain a particular exposure inside that picture: AFP data summarized by the Federal Reserve found 63% of organizations experienced attempted or actual check fraud back in 2024, while 91% still reported using checks and more than 75% had no immediate plans to stop.
Manual Control Gap | Risk Created | Evidence Needed |
Email approvals | Weak approval traceability | Approver, timestamp, decision rationale |
Manual beneficiary setup | Fraud or misdirected payments | Verification record, approver history |
Spreadsheet edits | Version-control issues | Change history |
Disconnected bank portals | Incomplete payment visibility | Unified activity log |
Manual exception tracking | Missed or delayed resolution | Exception queue and closure notes |
At a supervised level of autonomy, a payment fraud and anomaly detection agent could monitor the payment queue continuously against beneficiary whitelists, historical patterns, and policy thresholds, routing only flagged items to a human review queue instead of asking a team to review everything. The control gets stronger, not weaker, because the evidence chain behind every flagged and cleared payment, the data used, the reasoning, the reviewer's decision, and the final action, is logged automatically rather than reconstructed from memory during an audit.
Cost category 6: scalability bottlenecks
Manual treasury operations scale linearly with complexity, which is another way of saying they don't scale at all. More bank accounts mean more reconciliation. More entities mean more intercompany funding complexity. More currencies mean more FX exposure to track. More payment volume means more exceptions. More regions mean more banking formats and cutoff times to manage by hand.
Manual treasury workflows scale by adding people, not capacity. Every additional entity, bank relationship, or currency adds a proportional amount of manual work unless the underlying workflow is standardized and automated.
Intercompany settlement and netting is the clearest illustration: a rules-based process (apply netting rules, calculate net positions, generate settlement instructions) that a manual team has to coordinate across every ERP instance, currency, and jurisdiction involved. An agent applying documented netting rules can execute this consistently at scale, with the potential for greater autonomy as rule compliance and performance are demonstrated, which is precisely the workflow where manual effort grows fastest as an organization adds entities.
Cost category 7: poor AI readiness
This is the cost category most organizations don't put a number on, and it's the one that determines whether every other fix on this list is even possible. Despite the attention AI is getting in treasury, maturity remains low: PwC's 2025 Global Treasury Survey, based on 350 treasurers globally, found only 26% rated their AI capabilities as moderately or very mature, even as predictive analytics, anomaly detection, and RPA are increasingly embedded in forecasting, cash visibility, reconciliation, and exposure data gathering elsewhere in the function.
AI cannot reliably improve treasury decisions if the underlying workflow is fragmented, inconsistent, or stale. An agent is a multiplier of your data infrastructure: clean, well-connected data produces fast, accurate outputs, while fragmented data from multiple entities and banking platforms produces confident-looking outputs that are wrong in ways that aren't immediately visible. Layering automation on top of a broken workflow doesn't fix the workflow. It just makes the errors move faster. IBM's finance AI research, conducted with APQC, points to why the gap persists: organizations achieving stronger AI returns aren't simply deploying the most use cases; they're redesigning workflows, standardizing data, and governing AI as an operational capability before they scale it.
The most reliable predictor of a difficult agentic finance pilot isn't governance immaturity. It's data quality problems that weren't assessed before the pilot began: bank feeds that refresh inconsistently, ERP data that lags by hours, entities with no direct banking connectivity, forecast inputs that get manually adjusted before use. All of that is normal in a real treasury environment, and all of it affects an agent's output in ways that are hard to separate from the agent's own performance once a pilot is already running.
Manual, fragmented workflows are, in effect, an AI-readiness debt. Fixing them isn't preliminary to an AI agent program. It's the first phase of it.
The seven costs, at a glance
Cost Area | What Manual Work Looks Like | What It Costs | What an AI Agent Changes |
Labor | Data gathering, reporting, approval chasing | Analyst hours and overtime | Parallel, multi-entity data retrieval and drafting |
Visibility | Delayed consolidation of cash positions | Conservative decisions on stale data | Same-day position and forecast, human-validated |
Liquidity | Cash trapped across entities and banks | Idle cash, excess buffers, borrowing | Headroom analysis and transfer recommendations |
Errors and rework | Manual matching and break investigation | Rework, slower close, exception backlog | High-volume auto-matching with an exception queue |
Fraud and controls | Scattered evidence and email trails | Audit burden, compliance risk | Continuous monitoring with a logged evidence chain |
Scalability | More volume requires more headcount | Higher operating cost per entity added | Rules-based execution that scales without headcount |
AI readiness | Fragmented data, inconsistent workflows | Delayed automation ROI, failed pilots | The data-quality baseline that makes every use case above possible |
How to estimate the cost of manual treasury workflows
A simple framework for putting a number on it:
Annual Manual Treasury Cost = Labor Cost + Error/Rework Cost + Liquidity Cost + Risk/Control Cost + Opportunity Cost
1. Labor cost. Number of treasury staff involved, hours per week each spends on manual reporting, reconciliation, and approvals, fully loaded hourly cost, and how often the manual process repeats. If three treasury professionals each spend 10 hours a week on manual data collection, reconciliation, and reporting, that's 30 hours a week of capacity that could otherwise go to forecasting, liquidity analysis, or risk management.
2. Error and rework cost. Monthly reconciliation breaks, average resolution time per break, duplicate or failed payment rate, exception volume, and internal audit remediation hours.
3. Liquidity cost. Average idle cash held as a buffer, short-term borrowing caused by forecast misses, missed yield or investment return, and missed early-payment discounts.
4. Risk and control cost. Time spent on manual approval workflows, audit evidence collection, payment fraud exposure, beneficiary-change controls, and compliance documentation.
5. Opportunity cost. Strategic projects delayed, scenario analysis not performed, forecast improvements not pursued, and decisions treasury couldn't make fast enough to matter.
Before running this calculation for an AI agent business case specifically, it's worth adding one more input most organizations skip: a documented data-quality baseline for the process in question. Which entities have direct bank connectivity and which require manual workarounds, how fresh each data source actually is against what the use case needs, and where the known gaps are. That baseline typically takes one to two weeks to build for a use case of moderate complexity, and it's the single most valuable input into whether an agent will perform well or simply automate the same gaps faster.
Manual vs. agentic AI: what changes
Treasury Workflow | Manual State | With an AI Agent |
Cash positioning | Bank portal downloads and spreadsheets | Consolidated, same-day cash view with human validation |
Forecasting | Static spreadsheet assumptions | Continuously updated, variance-flagged forecast with human validation |
Liquidity transfers | Manual headroom calculation and approval routing | Agent recommendations with human approval |
Reconciliation | Manual matching and break investigation | Automated matching with exception-based human review |
Payment fraud detection | Manual review of flagged items | Continuous monitoring with flagged exceptions routed to human review |
Intercompany settlement | Manual coordination across ERPs and currencies | Rules-based execution within defined controls |
Audit evidence | Scattered files and email threads | Logged evidence chain: data, reasoning, recommendation, approval, action |
The autonomy levels follow directly from how much is at stake per decision and how stable the process is: high-frequency, well-defined, lower-stakes work earns more autonomy over time, while lower-frequency, higher-consequence decisions keep a human owner regardless of how well the agent performs elsewhere.
Common questions
Common questions and answers:
What are the hidden costs of manual treasury workflows?
Beyond labor hours: delayed cash visibility, trapped liquidity and unnecessary borrowing, reconciliation breaks and rework, weaker fraud controls, heavier audit burden, scalability limits that force headcount growth, and reduced readiness to deploy AI agents safely.
How do manual treasury processes affect cash visibility?
When positions take hours or days to consolidate across banks and entities, decisions get made on stale data. That typically pushes teams toward conservative choices, like holding excess cash as a buffer, rather than optimizing in real time.
Why do manual treasury workflows increase risk?
Email-based approvals are harder to trace, manual beneficiary changes are a common fraud vector, and fragmented data makes anomaly detection slower because no single system is watching the full picture at once.
How can finance teams calculate the cost of manual treasury work?
Add labor cost, error and rework cost, liquidity cost, risk and control cost, and opportunity cost. Each has specific, countable inputs, like hours per week on manual tasks, monthly reconciliation breaks, and average idle cash held as a buffer.
Why do fragmented treasury workflows make AI harder to use?
An AI agent is a multiplier of the data infrastructure underneath it. Fragmented, inconsistent data produces confident-looking agent output that's wrong in ways that aren't visible until something breaks downstream. Connecting and standardizing the data comes first.
When should a company deploy an AI agent for treasury workflows?
Once the underlying data quality has been assessed and baselined for the specific use case. High-volume, rule-based, well-understood processes like reconciliation and fraud detection are typically the safest starting points; lower-frequency, high-stakes decisions keep a human owner regardless of how mature the automation becomes elsewhere.
What's the difference between automating treasury workflows and using an AI agent?
Traditional automation follows a fixed script and breaks visibly when conditions change. An AI agent reasons across live data, adapts within an authorized scope, and still requires the same governance discipline, defined authority, logged evidence, and human approval for material actions, just applied to a system that can handle more volume and complexity than a fixed script can.
Related Reading
Related resources
