Case study / AI treasury forecasting / 2025
Flowcast AI treasury and liquidity copilot
I designed a treasury workspace that brings cash forecasts, source checks and comparable scenarios into one approval flow. After implementation, more financial plans passed their first review without substantial comments.
The product name and branding elements have been changed to preserve confidentiality. The original product solutions and the results of their implementation are shown.
01
Overview
Project info
Flowcast is a desktop treasury workspace for a 13-week GBP cash forecast across eight bank, ERP, payroll and FX sources. I connected scenario comparison, the preserved base case and data freshness checks so financial plans reached approval with their assumptions and consequences visible. My work covered product strategy, AI interaction, financial-risk UX and the design system.
02
The challenge
A financial plan needs to explain how a proposed action covers a cash shortfall, which assumptions it uses and whether its sources are current. Treasury data updates at different times. Reviewers need comparable amounts and timing before they can accept a plan without substantial comments.
- Eight changing sources
- Bank, ERP, payroll and FX data had to share one entity, currency, timing and freshness model.
- Comparable plans
- The base case and alternatives needed the same horizon and measures, with the original assumptions still visible during review.
- Approval readiness
- A plan needed current decisive sources and clear consequences before an authorised person could approve it.
03
Why AI
Treasury forecasting combines time-series calculations with business events that need interpretation. AI prepares drivers and scenarios; the interface brings their sources and assumptions into the financial-plan review.
Forecast preparation
The model estimates a 13-week cash position and presents uncertainty alongside the status of its inputs.
Driver attribution
It links forecast movement to receivables, payroll, tax and FX events with visible sources.
Scenario preparation
It prepares comparable base, stress and remedy scenarios for human review.
04
AI system model
Eight treasury sources feed a forecast that stays connected to its drivers, comparable plans and a human approval record.
- 01
Collect
Read bank, ERP, payroll, receivables, tax and FX sources.
- 02
Normalise
Align entity, currency, timing and freshness before a source enters the forecast.
- 03
Forecast
Produce a 13-week cash range, driver attribution and conditional confidence.
- 04
Compare
Keep the base case, stress case and proposed remedy on the same horizon, preserving the original plan.
- 05
Approve and record
Require role-based human confirmation and write the selected scenario, sources and timestamp to the audit trail.
05
Key decisions
-
Make the shortfall actionable
I placed projected close, reserve floor, breach horizon and minimum headroom beside the forecast. The team could use that shared context to prepare and review a financial plan.
-
Compare plans on the same basis
I brought the base case, stress case and proposed remedy into one view. Amounts, timing and sources stay comparable, while the approved base case remains the reference for review.
-
Check the evidence before approval
I made source freshness visible before a plan reached approval. Missing or stale decisive data pauses the action and opens a recovery path, so the team can address the affected assumptions before resubmitting.
A financial plan is ready for review when its sources, assumptions and consequences can be checked together.
06
The solution
Liquidity command centre
The first view brings current cash, projected close, the reserve floor and the shortfall horizon together. The review action leads from the liquidity risk to the financial plan that addresses it.
- Observation
- A plan needs a clear account of when the shortfall occurs and which reserve it affects.
- Decision
- Bring cash position, projected close, reserve floor and next action into the first view.
- Effect
- The team starts planning and review from the same liquidity position and time horizon.
Sourced forecast drivers
Each driver connects a forecast movement to its amount, source and timing. The reviewer can inspect the assumptions behind the proposed plan alongside the explanation of how the forecast was built.
- Observation
- A forecast total does not show which business events a proposed response depends on.
- Decision
- Attach source, timing and contribution to every decisive driver.
- Effect
- The assumptions behind a financial plan remain available for inspection during its first review.
Comparable scenario consequences
The base case, stress case and proposed remedy remain visible together. Closing cash and minimum headroom use the same 13-week horizon, and exploring an alternative preserves the approved base case.
- Observation
- Reviewing alternatives requires comparable measures and a stable reference.
- Decision
- Keep all three scenarios in one view with consistent amounts, timing and sources.
- Effect
- The reviewer can compare the proposed plan with its base case and inspect the consequences before accepting it.
Role-based approval
The approver reviews the selected plan, amount, supporting sources and forecast effect before recording a decision. The AI recommendation and the authority to approve it remain distinct.
- Observation
- A financial plan needs both a reviewable rationale and an authorised decision.
- Decision
- Bring the selected scenario, its evidence and the approval gate into one review state.
- Effect
- The approver can accept the plan or request a revision with the proposal and its context still visible.
Data failure and recovery
Missing or stale decisive data pauses approval. The interface identifies the affected forecast inputs and offers reconnection, a fallback source and a new forecast calculation.
- Observation
- A plan based on stale data can reach review with assumptions that no longer hold.
- Decision
- Show source freshness and block approval until decisive data has been restored.
- Effect
- Data issues can be addressed before the plan is submitted for approval, with the affected assumptions kept visible.
07
Human oversight and recovery
Failure and recovery states
Normal forecast
- When
- The required sources are current and decisive inputs are available.
- Product response
- The forecast is available for scenario comparison and plan preparation.
Low confidence
- When
- A decisive source approaches or exceeds its freshness limit.
- Product response
- The range widens, the reason is named and approval becomes conditional.
Missing source
- When
- A decisive feed is unavailable and its contribution cannot be included in the forecast.
- Product response
- Approval pauses while the interface shows the affected data and a recovery path.
Human approval
- When
- A remedy is selected and the required source checks pass.
- Product response
- An authorised approver reviews the transfer and records the decision.
Return for review
- When
- The approver finds a source, amount or assumption that needs revision.
- Product response
- The proposal returns with its baseline and evidence history intact.
Safety and review controls
Visible source timing
Decisive claims show the source, refresh time and contribution to the forecast.
Conditional confidence
Confidence and forecast range respond to the health of the evidence.
Proposal-only AI
The model prepares an action but cannot approve or execute a payment.
Role-based confirmation
The record stores the scenario, sources, amount, role and confirmation time.
Quantified safe stop
When data fails, the product blocks action and reports the cash excluded from the forecast.
08
AI evaluation
After implementation, product analysts reported an increase in financial plans passing their first review without substantial comments. This outcome reflects plan acceptance in the implemented workflow.
Evaluation methods
What the measure counts
A financial plan passes on the first review without substantial comments. The measure concerns approval readiness and the quality of the submitted plan.
Before and after implementation
The result compares first-pass acceptance before and after implementation. Scenario comparability, the preserved base case and source freshness were the focus of the design changes.
09
Results
Following implementation, the share of financial plans accepted on first review without substantial comments rose from 70% to 76%. I designed comparable scenarios, a preserved base case and visible source freshness to support the preparation and review of those plans.
10
Reflection
The central design decision was to keep the plan, its assumptions and its sources together through approval. Scenario comparison only helps when the base case remains stable and the reviewer can see whether the data is current. That connection made first-pass plan quality the relevant outcome for this work.
11
Let’s Connect
Building an AI, FinTech or Enterprise product? I can help shape complex decisions into clear workflows with evidence, controls and measurable outcomes.
Download CV