In brief
AI in finance is simultaneously overhyped in the short term and underestimated in the medium term. That is why many teams are making the same mistake from opposite directions: one group is waiting for perfect tools, data and proof; the other is running too many disconnected pilots and calling that progress.
- Today’s value is concentrated in repetitive, high-volume workflows: transaction classification, invoice processing, reconciliation support, collections prioritisation and anomaly detection.
- Forecasting is not being automated end-to-end. It is being improved at the margins through better baselines, faster scenarios and earlier signal detection.
- The real differentiator over the next three to five years will be learning velocity: how quickly a company tests, kills, scales and embeds AI into real workflows.
- The biggest risk is not moving too fast. It is learning too slowly while competitors build the operating capability you postpone.
This perspective is written for CFOs, CFO-1s, finance transformation leaders and PE operating partners who want a realistic answer to one question: how do we build advantage without pretending that the technology, the data or the organisation are already perfect?
The goal is to get better faster than competitors — and compound that learning into operating advantage.
1. The reality gap: what the market promises vs what companies can actually absorb
The current AI narrative in finance is split between exaggerated confidence and quiet confusion. Vendors promise autonomous finance, self-driving forecasting and copilots that replace analysts. Inside companies, the lived reality looks very different: inconsistent master data, local process variation, ERP constraints, weak workflow integration and a persistent trust gap around outputs.
Autonomous finance, one-click forecasts, copilots replacing analysts, instant ROI.
Pilot overload, partial automation, manual overrides, low trust and unclear ownership.
This matters because AI in finance is not a pure technology problem. It is an execution problem. If the use case is not anchored in a process, linked to a KPI, owned by finance and integrated into the existing workflow, the organisation may “use AI” without materially changing cost, cash or decision quality.
The two bad strategies
- Waiting for perfection. Teams delay because the data is not clean enough, the ERP is not modern enough or the use cases are not proven enough.
- Unstructured experimentation. Teams launch many pilots, often tool-led, but never force the discipline to kill or scale them.
Both strategies feel rational. Both create lag. The first creates strategic drift; the second creates organisational noise.
2. The core idea: AI advantage is really learning velocity
The most useful way to think about AI in finance is not as a transformation project and not as a software buying decision. It is a learning system. Some companies will use the next few years to build data discipline, sharpen use-case selection, train people to challenge outputs and integrate AI into recurring workflows. Others will still be comparing tools.
| Level | What it looks like | What it leads to |
|---|---|---|
| 1. Passive | No real experimentation; waiting for maturity | No learning, no capability build |
| 2. Pilot chaos | Many experiments, weak governance, no scaling logic | Noise, vendor dependence, low trust |
| 3. Disciplined experimentation | Few focused use cases, clear KPIs, finance ownership | Selective value creation and faster learning |
| 4. Workflow scaling | Successful use cases embedded across entities and teams | Meaningful cost, cash and control gains |
| 5. Compounding advantage | AI capability becomes part of how finance runs the business | Structural edge in decision speed and operating leverage |
The mistake is to think maturity starts when the tools are perfect. In practice, maturity starts when the organisation learns how to test, challenge and embed imperfect tools better than others.
3. Where AI actually works today
Real value is concentrated in finance processes with three characteristics: repeatability, volume and clear exception logic. That is why the most credible use cases today are not the glamorous ones. They sit in AP, AR, close, reconciliation, treasury, controls and selected parts of FP&A.
FP&A: augmentation, not autopilot
In FP&A, AI helps most when it improves the quality and speed of the starting point. It can propose a baseline forecast, accelerate variance analysis, surface patterns earlier and refresh scenarios faster. What it does not do reliably is absorb strategic context, pricing decisions, restructuring actions, customer concentration risk or one-off market shifts without human judgement.
Value: Faster first draft, better pattern detection, less spreadsheet handling.
Value: More scenarios, faster trade-off discussions with the business.
Value: Less manual digging, faster management commentary.
R2R and SSC: where the real ROI usually starts
If a CFO wants immediate evidence that AI can create finance value, this is still the most credible place to start. High-volume, repetitive tasks lend themselves to partial automation far more than judgement-heavy planning processes. Transaction classification, reconciliations, invoice capture, posting support and exception routing all sit here.
- Transaction classification: high automation potential in mature environments with consistent chart-of-accounts logic.
- Reconciliation support: matching logic and prioritised exception handling can materially reduce close effort.
- Invoice processing: touchless processing rises meaningfully when vendor and PO discipline are strong.
- Close task orchestration: AI is less about replacing close owners and more about surfacing exceptions earlier and improving sequencing.
O2C and working capital: targeted gains, not magic
Collections prioritisation, payment prediction and dispute triage can create genuine value, particularly where working capital is a strategic lever. But the gains are uneven. If billing quality is weak, customer data is fragmented and dispute ownership is unclear, AI helps less than expected.
Controls, audit and compliance: high-value, low-glamour
This is the area many businesses overlook. AI is strong at scanning broad populations for exceptions, unusual patterns and signals that a control may have failed. That does not remove the need for control owners, but it can materially improve coverage and shorten response times.
4. Where AI still disappoints
The weak spots are remarkably consistent. They are the areas where organisations most want certainty but the technology still depends on context, integration and judgement to produce reliable output.
Useful narrative, weak operating reality. The business context still dominates the model.
It removes manual work; it does not replace challenge, context or leadership interaction.
Often produce interesting outputs, but weak adoption because they are not embedded or auditable.
Create a shadow layer that fragments data, accountability and controls.
This is why disciplined skepticism matters. If a claim cannot be tied to workflow adoption and measurable business impact, it is still narrative rather than proof.
5. The CFO playbook: how to move now without pretending everything is ready
The best approach is neither caution disguised as prudence nor experimentation disguised as strategy. The right approach is disciplined experimentation: a small number of focused moves, each tied to value and embedded quickly enough to create learning.
Step 1: define three value pools
- Cost: where can manual finance effort be reduced meaningfully?
- Cash: where can working capital, collections or forecast timeliness improve?
- Decision quality: where can management see, test and react faster?
Step 2: choose only a few use cases
Most organisations should not start with a portfolio of fifteen ideas. They should start with three to five use cases with visible economics. That makes it easier to force focus, governance and post-pilot decisions.
Step 3: run 8–12 week experiments
Each experiment should answer a commercial question, not just a technical one. Did processing time fall? Did exception quality improve? Did forecast cycle time shorten? Did collections productivity rise? If not, stop.
Step 4: kill or scale aggressively
This is where most businesses underperform. They neither stop weak pilots nor scale the strong ones fast enough. The discipline to kill is what protects speed; the discipline to scale is what converts learning into advantage.
Step 5: build capability in parallel
Do not wait for a separate “people plan” later. Finance teams need to learn how to interpret outputs, challenge weak logic and use AI inside their normal operating rhythm. The future edge is not that everyone becomes a data scientist. It is that finance becomes better at challenging and using model outputs in business context.
6. The Monday morning test
If a CFO, CFO-1 or PE operating partner cannot answer the questions below, the AI agenda is probably still too abstract.
- Which three use cases should deliver measurable impact in the next 12 months?
- What KPI defines success for each use case?
- Who in finance owns each use case?
- What gets killed in 90 days if it does not work?
- What operating habit changes if one of these use cases succeeds?
Good AI agendas are not broad. They are specific enough to create decisions, funding choices, operating changes and clear stopping rules.
7. Closing thought
AI in finance will become more powerful over time. That part is likely. What is far less certain is who will actually benefit. The winners will not be the companies with the most polished language or the most vendor demos. They will be the ones that learn fast, organise around a few valuable use cases and build the discipline to embed what works.
In other words: the future advantage will not come from perfection. It will come from learning velocity.