AI & Finance

The opportunity is real.
So is the hype.

AI is here to augment human decision-making — not replace it.

"Technology can automate the technical work — but it cannot replace the judgment, the relationships, and the ability to help teams navigate change with confidence."

Section One
What AI actually does
well in finance.

Concrete capabilities. No buzzwords. These are the tasks where AI delivers real, measurable value — today, in production, at scale.

Reconciliations

Matching thousands of transactions across systems in minutes rather than days. AI catches discrepancies humans miss — not because it is smarter, but because it never gets tired and never skips a line.

Monthly financial review 95% ready on first draft. Discussion shifts from what happened to why it happened — which is where CFO value actually lives.
Anomaly Detection

Flagging unusual patterns in spend, revenue, or cash flow before they become problems. AI monitors continuously — something no finance team, however large, can do manually across all accounts simultaneously.

Fraud detection, budget overruns, duplicate payments — caught in real time rather than discovered in the next month-end close.
Forecasting Models

Processing far more variables than any spreadsheet model — historical patterns, seasonality, external signals — and updating in real time as new data arrives. Faster, more consistent, and free of the cognitive biases that affect human forecasters.

Rolling forecasts that update automatically rather than requiring a week of manual work each quarter.
Document Processing

Extracting data from contracts, invoices, purchase orders, and regulatory filings at speed and scale. What takes a team member hours takes AI seconds — with consistent accuracy and full audit trail.

Statutory statements for every legal entity, produced automatically. A human checks the output. Claude does the work.
Compliance Monitoring

Continuous monitoring against regulatory requirements, internal policies, and approval thresholds — with automatic flagging rather than periodic manual review. Validated by Deloitte and regulators in live deployments.

An AI compliance platform that passed regulatory scrutiny — not because it replaced judgment, but because it made judgment faster and more consistent.
Reporting Automation

Weekly and monthly reports that used to take hours now take minutes. The finance team stops assembling data and starts interpreting it. Time shifts from production to analysis — which is where CFO value actually lives.

Weekly revenue reports: from hours to 30 minutes. The bottleneck was time. Time is no longer the bottleneck.
Section Two
What AI does
not replace.

This is the section most AI vendors skip. It is also the most important one for any business leader deciding how to deploy these tools.

Judgment

AI can tell you what the numbers say. It cannot tell you what to do about them. The decision to exit a market, restructure a division, or hold a position through a difficult quarter requires context, experience, and accountability that no model has. Yet.

"Data informs. Judgment decides. Courage acts. All three are required — and only one of them is human."
Relationships

The bank that extends credit in a crisis does so because of a relationship built over years. The board that trusts the CFO's recommendation does so because of credibility earned through difficult conversations. AI builds neither.

"The most senior person on the team is the heaviest AI user. That is the signal — not because AI replaces seniority, but because senior judgment, accelerated, is the competitive advantage."
Accountability

Every number presented to a board or investor carries someone's signature — literally or implicitly. That signature means something. When the numbers are wrong, someone answers for it. AI does not answer for anything. The CFO does.

"The CFO who walks into the boardroom and defends a number is not being replaced by AI. They are being equipped by it."
Ethics & Culture

AI reflects the data it is trained on and the governance it operates within. If the culture tolerates cutting corners, AI will help cut them faster. The ethical framework — the decision about what is right, not just what is legal — remains entirely human.

"Ethics is not a compliance checkbox. It is the culture that ensures the numbers you are looking at reflect reality."
Section Three
Where most businesses
get it wrong.

These are the failure modes that repeat — across industries, across geographies, across company sizes. None of them are the AI's fault.

01
Deploying AI before cleaning the data

AI is only as good as the data it processes. Garbage in, garbage out — at machine speed. A business with inconsistent chart of accounts, incomplete master data, or manual workarounds in the ERP will not be fixed by AI. It will have its problems amplified and accelerated.

Data preparation typically accounts for 60–70% of the real project cost. Most budgets allocate 10%.
02
Automating bad processes

AI is extraordinarily good at doing what you tell it to do — including doing the wrong thing very efficiently. A broken approval process automated at scale is still a broken approval process. The question before deployment is always: should this process exist at all?

Automating a flawed reconciliation process does not fix the reconciliation. It just hides the errors faster.
03
Confusing speed with accuracy

AI produces outputs fast. Fast is not the same as correct. Finance teams that remove human review from AI outputs because "it's faster" are trading accuracy for efficiency — a trade that eventually costs far more than the time saved. The human in the loop is not a bottleneck. It is the control.

A single AI error in a regulatory filing can cost more than an entire year's AI investment.
04
No governance framework

Who owns the AI outputs? Who reviews them? What happens when the model is wrong? Who decides which tasks AI can handle autonomously versus which require human sign-off? Most businesses deploy AI without answering any of these questions — and discover the gaps at the worst possible moment.

Regulators are increasingly asking these questions. The answer "we don't have a framework yet" is no longer acceptable.
05
No understanding of what it costs

Every query costs something. Every document processed costs something. Every API call, every model run, every automated report has a price — and at scale, those prices compound. Most finance teams deploying AI have no cost model for the AI itself. The irony is significant.

One organisation accidentally spent $500M in a month on AI tokens. The revenue model, it turns out, was essentially uncapped accidental billing.
Section Four
How to approach it
the right way.

A framework built from real deployments — including an AI compliance platform validated by Deloitte and regulators, and 30% back-office cost reduction delivered through AI-enabled finance transformation.

1
Clean the data first

Before any AI deployment, audit your data quality. Standardise the chart of accounts. Close the gaps in master data. Eliminate manual workarounds. This is unglamorous work. It is also the work that determines whether your AI deployment succeeds or fails.

2
Start high-volume, low-risk

Begin with tasks that are high in volume, low in risk, and easy to verify — reconciliations, document extraction, standard reporting. Build confidence, build the team's comfort level, and build the governance framework before moving to higher-stakes applications.

3
Build governance before you need it

Define ownership, review processes, escalation paths, and error protocols before deployment — not after the first incident. Include your compliance and legal teams. If you are in a regulated industry, engage your regulator early. Surprises in regulated environments are expensive.

4
Model the cost — like any investment

Build a business case. Model the token costs, the integration costs, the maintenance costs, and the human oversight costs. Compare them to the labour and time they replace. If the ROI is not there, do not deploy. AI is a capital investment decision — treat it like one.

The Cost Nobody Talks About
Running AI is not free.
Most budgets assume it is.
Token Costs

Every query, every document processed, every API call has a price. At low volume, the costs are negligible. At scale — thousands of documents per day, continuous monitoring, automated reporting — they compound fast. A finance team running high-volume AI workflows needs a token budget, just like any other operating cost line.

"A 42-second sketch went viral: CEO finds out tokens cost money. It was funny. It was also the most accurate financial education on AI spend I had seen that week."

The Hidden Costs

The token bill is the visible cost. The hidden costs are usually larger:

  • Data cleaning and preparation (typically 60–70% of total project cost)
  • ERP and systems integration
  • Ongoing model maintenance and prompt optimisation
  • Human oversight — you still need people reviewing AI outputs
  • Security, compliance, and audit infrastructure
  • Retraining and change management for the finance team
Cost Risk by Use Case
Invoice & document processingLOW
Standard reconciliationsLOW
Automated reportingMEDIUM
Continuous compliance monitoringMEDIUM
Real-time anomaly detectionMEDIUM
Large-scale forecasting modelsHIGH
Unstructured document analysis at scaleHIGH
Agentic AI with no human-in-loopHIGH
What Is Tokenmaxxing?

A token is the basic unit of AI processing — roughly three-quarters of a word. Every time you send text to an AI model, it is broken into tokens. Every token costs something. Tokenmaxxing is the discipline of getting maximum value from every token you spend — choosing the right model for the right task, batching efficiently, and optimising how you frame requests. Don't use a $0.015-per-token model for a task a $0.0002-per-token model handles equally well. The CFO who applies the same cost discipline to AI spend as to any other operating cost will consistently outperform the one who doesn't.

But What If My Data Stays in the ERP?

Tokens are consumed whenever you use an AI model — regardless of where your data comes from. The deployment model changes where costs live, not whether they exist:

Public AI (Claude, ChatGPT via API)
Data goes to the provider's servers. Per-token billing. Lowest upfront cost, highest ongoing variable cost. Data privacy requires a business agreement.
Private / Self-Hosted AI
Data never leaves your infrastructure. No per-token bill — but you pay for compute, servers, GPU time, and energy. Full data control. Higher upfront cost. What large banks and regulated institutions typically choose.
Embedded AI in ERP (SAP Joule, Oracle AI)
AI built into your existing system. Data stays within the ERP. Token costs are bundled into your licence fee — invisible but not absent. Simpler governance, less flexible.

The tokenmaxxing discipline applies to all three. In public AI it is literal token efficiency. In private and embedded models it translates to compute efficiency and licence value. The principle is the same: maximum value from every unit of AI processing, whatever form that takes.

"AI is a powerful tool.
Like all tools, it is only as good as
the judgment behind it."

The businesses that will win with AI in finance are not the ones that deploy the most AI. They are the ones that deploy it most deliberately — with clean data, clear governance, honest cost models, and human judgment firmly in the loop.

Talk to Aethon About AI in Your Finance Function
More on AI, finance, and leadership —
every Tuesday on LinkedIn.
Tokenmaxxing, sovereign AI investment, SpaceX IPO analysis — CFO thinking, not tech hype.
Follow on LinkedIn