AI in finance: build a control tower, not a crystal ball

Where would an earlier signal improve a financial decision?

Finance AI prepares and prioritises evidence; accountable people still interpret, approve and sign off.

Finance leaders do not need another system that promises certainty. They need earlier signals, cleaner evidence and more time to act. Most finance teams already have the raw material: transactions, invoices, budgets, bank files, sales orders, inventory movements and operational plans. The problem is that the information arrives at different times, in different formats and with different definitions. By the time it has been reconciled and explained, the decision window may have narrowed.

Artificial intelligence can help, but only if “AI” is separated into useful jobs. Predictive models can estimate what may happen next. Anomaly detection can identify behaviour that differs from an established pattern. Document AI can turn statements and invoices into structured records. Generative AI can explain results and help people ask questions in ordinary language. None of these capabilities is a substitute for accounting policy, internal controls or the judgment of a finance professional.

The right ambition is a finance control tower: a connected way to see what changed, why it may have changed, which items need attention and what evidence supports the proposed response. A control tower does not fly the aircraft for the pilot. It brings the right signals together so that the pilot can make a better decision.

Start by distinguishing four kinds of work

The first category is record preparation . Finance data often enters through invoices, expense receipts, spreadsheets and bank files. It must be extracted, standardised and mapped before analysis. Google Cloud Document AI can extract structured information from documents, while Gemini-assisted data preparation in BigQuery can suggest transformations, standardisation and schema mapping. These are ways to prepare the evidence, not to approve a transaction.

The second category is prediction . A forecast uses historical patterns, and sometimes external drivers, to estimate future values. BigQuery ML supports time-series forecasting with ARIMA_PLUS and ARIMA_PLUS_XREG models. Google Cloud also documents a built-in TimesFM foundation model option. Forecast output can include prediction intervals, which are essential because a range is more honest than a single number.

The third category is detection . An anomaly model asks whether an observation is unusually far from an expected pattern. It can help identify a sudden margin change, an unusual payment amount, a cost centre behaving differently from its history or a return rate that breaks its normal range. BigQuery provides anomaly-detection functions for time-series and other forms of data. An anomaly is a prompt to investigate, not a verdict that someone committed fraud or made an error.

The fourth category is explanation and decision support . Gemini in Looker provides Conversational Analytics, allowing authorised users to ask questions in natural language and receive charts or tables grounded in the Looker semantic model. The semantic layer matters because business terms such as revenue, gross margin and churn should have governed definitions. Google’s documentation also warns that generative output can sound plausible while being wrong and recommends validating it before use.

When these four capabilities are combined carefully, finance can move from retrospective reporting towards continuous attention. The combination is powerful precisely because each component has a limited job.

Use case one: forecasting that exposes assumptions

Many forecasts are laborious before they are analytical. Teams collect templates from business units, fix broken formulas, align categories, paste actuals and debate which version is current. AI cannot repair weak planning governance, but it can reduce mechanical work and add a consistent statistical baseline.

Imagine a multi-site services company forecasting weekly cash receipts and labour cost. A practical design begins with a trusted historical table: invoice dates, expected and actual payment dates, customer segment, contract terms, workforce hours, holidays and known operational events. A time-series model generates a baseline forecast and an interval showing a reasonable range around that estimate. Finance then overlays known information that history cannot infer, such as a contract ending, a one-off settlement or a planned site closure.

BigQuery’s ARIMA_PLUS pipeline can infer data frequency, handle irregular intervals and duplicates, interpolate missing values, detect spikes or dips, account for level changes, holiday effects and multiple seasonal patterns, and automatically evaluate candidate models. That automation is helpful, but it does not make the result self-explanatory. The finance team should still ask whether the training period represents the current business, whether an unusual event should be retained or adjusted, and whether the forecast’s range is wide enough for the decision at hand.

A useful operating rhythm has three layers. The statistical baseline is produced consistently. Business owners submit documented adjustments, each with an owner and reason. Finance compares actuals with both the baseline and the adjusted view. Over time, that comparison reveals where human knowledge adds value and where optimism or stale assumptions repeatedly enter the plan.

Do not measure forecast success with one company-wide average. Track error by horizon, business unit and material driver. A one-week liquidity forecast should be judged differently from a twelve-month revenue outlook. Monitor directional accuracy, interval coverage and bias as well as absolute error. The question is not only “How close were we?” but also “Did we consistently overstate or understate, and did the range communicate the risk?”

Use case two: anomaly detection as a continuously updated attention list

Traditional controls often rely on thresholds: review every payment above a fixed amount, every discount above a percentage or every variance outside a budget band. These rules are understandable and should remain where policy requires them. They can also produce large queues and miss unusual behaviour below the threshold.

Anomaly detection adds context. A $20,000 payment may be normal for one supplier and extraordinary for another. A 5% expense increase may be harmless in a growing unit but concerning in a stable one. BigQuery’s anomaly-detection tools can analyse time series with ARIMA_PLUS, ARIMA_PLUS_XREG or TimesFM, and can use approaches such as k-means, principal component analysis and autoencoders for other patterns. Output can include a probability and upper and lower bounds.

In a finance setting, candidate uses include unusual payment timing, duplicate-like amounts, abrupt changes in average selling price, expense claims outside a peer pattern, unexpected movements between accounts, inventory shrinkage and return rates that depart from their baseline. The model should enrich an existing control environment, not silently replace it.

The review process must explain why an item appeared. Show the value, expected range, relevant comparison and source transactions. Let reviewers label the outcome: true issue, legitimate event, data-quality problem or model false positive. Those labels improve the control and help distinguish model quality from upstream-data failure.

Precision and recall should be considered together. A highly sensitive detector may find more true issues while overwhelming the team with false alerts. A very strict detector may create a manageable queue while missing meaningful events. Set thresholds according to risk and reviewer capacity, then test them on historical periods containing known incidents and normal seasonal changes.

Most importantly, do not describe an anomalous record as fraudulent. It is different from the pattern the model expected. Investigation, supporting evidence and authorised judgment determine what it means.

Use case three: reconciliation that routes differences instead of hiding them

Reconciliation is a strong AI candidate because it combines repetitive comparison with important exceptions. The goal is not to ask a language model whether the books balance. The goal is to prepare, propose and explain matches while retaining deterministic control over posting and sign-off.

Take bank-to-ledger reconciliation. Statements may arrive in a standard feed, CSV or PDF. Internal transactions have their own identifiers and timing. A practical workflow first converts both sides into a standard structure. Document AI can extract values when the source is a document. Gemini-assisted BigQuery data preparation can suggest cleaning and mapping steps, which staff can preview, edit and schedule. The system then applies exact matches: reference, amount, date and account. Next it applies approved tolerances and known one-to-many patterns. More flexible matching can propose candidates based on description, timing and value, but should attach a score and explanation.

The output is not a forced balance. It is three queues: confirmed matches, proposed matches requiring review and unresolved items requiring investigation. A reviewer accepts or rejects proposals, adds a reason and escalates old or material differences. Posting remains subject to the organisation’s normal segregation of duties.

The same pattern can support intercompany balances, accounts-receivable cash application, supplier-statement reconciliation and subledger-to-general-ledger checks. It can also help compare contract terms, purchase orders, receipts and invoices. In every case, a source record and audit trail matter more than a fluent narrative.

Measure the percentage matched exactly, percentage matched with approved tolerances, reviewer acceptance rate for proposed matches, unreconciled value and age, time to close, reopened items and adjustments posted after sign-off. A rising auto-match rate is not success if the number of later corrections also rises.

Use case four: decision support grounded in governed definitions

Senior finance staff spend considerable time translating a question into a data request. “Why did margin fall in the north?” may require an analyst to locate tables, reconcile definitions, write a query, build a chart and explain the result. Conversational analysis can shorten that path, especially for follow-up questions.

Looker’s Conversational Analytics uses Gemini to interpret natural-language questions and responds from data through the Looker semantic model. Because the semantic model holds governed definitions, it can understand that the company’s “net revenue” excludes certain items or that “active customer” follows a defined rule. Users can receive charts or tables and inspect how the answer was generated.

This supports a better conversation, not an unquestioned answer. A CFO might begin with “Show gross margin by region versus plan,” then ask “Which three product groups explain most of the variance?” and “Is the change driven by price, volume or mix?” The AI helps navigate; the governed model and underlying records remain the source of truth.

Decision support can also combine structured analysis with grounded narrative. A monthly performance pack might present the approved numbers, highlight statistically unusual changes and draft a commentary linked to the relevant measures. The controller reviews the explanation, adds business context and approves the final communication. The AI should never invent a causal story merely because two lines moved together.

For material decisions, require the user to see the period, filters, metric definition and source. Log questions and feedback. Maintain a set of recurring questions with expected outputs and retest them when definitions, models or data pipelines change.

A 90-day path from idea to controlled pilot

During the first month, choose one decision rather than a broad transformation. Good candidates include a thirteen-week cash forecast, daily revenue anomalies, one high-volume reconciliation or conversational analysis for a defined management dashboard. Name an executive owner, process owner, data owner and control owner. Record the current cycle time, error rate, review effort and decision latency.

Map the data and definitions. Which system owns the number? How are corrections handled? Are dates, currencies, entities and account mappings consistent? Resolve the most consequential quality issues before adding AI. BigQuery data insights can help technical teams explore table metadata, patterns, relationships and potential quality issues, but AI-generated descriptions and inferred relationships should be reviewed by data stewards.

During the second month, build in a sandbox using historical periods. Keep a holdout set for evaluation. Compare forecasts with actuals, anomaly alerts with known outcomes, or proposed reconciliation matches with completed work. For conversational analysis, establish a small library of questions and verified answers. Record not just model accuracy but the operational consequences of error.

During the third month, run a parallel pilot. The AI proposes; the existing process remains authoritative. Review differences weekly. Once quality and controls meet the agreed threshold, allow a narrow production use, such as auto-clearing only exact, low-risk matches or distributing forecasts as a secondary baseline. Preserve a kill switch and a manual route.

Governance is part of the product

Finance data is sensitive, and some generative features have different compliance boundaries from the underlying data platform. Google Cloud’s documentation explicitly notes that Gemini in BigQuery does not support all the same compliance and security offerings as BigQuery. Teams should review data location, access, encryption, logging and contractual requirements before enabling it.

Apply least-privilege access, segregate model development from transaction approval, and prevent AI tools from initiating payments, changing master data or posting journals without an authorised workflow. Mask or exclude information that is not required. Keep the original evidence and every human override.

For generative explanations, require validation. Ground the output in governed data, use fixed definitions and show the query or source. For predictive models, monitor performance, bias and drift. For anomaly models, monitor queue volume and reviewer outcomes. For reconciliation, monitor correction rates after close.

These controls do not weaken the business case. They make it sustainable. Finance earns trust by making uncertainty and accountability visible.

The result: earlier evidence, not automated certainty

AI can give finance a more continuous view of the business. It can prepare records, establish a forecast baseline, identify unusual movements, propose reconciliations and make governed analysis easier to access. The value appears when people spend less time assembling facts and more time interpreting them.

The most credible finance-AI programme will not promise perfect foresight or a fully autonomous close. It will show, process by process, that signals arrive earlier, exceptions are better prioritised, evidence is easier to trace and decisions are made with a clearer view of uncertainty. That is what a control tower is for.

• BigQuery forecasting overview https://docs.cloud.google.com/bigquery/docs/forecasting-overview
• ARIMA_PLUS model and forecasting pipeline https://docs.cloud.google.com/bigquery/docs/reference/standard-sql/bigqueryml-syntax-create-time-series
• BigQuery anomaly-detection overview https://docs.cloud.google.com/bigquery/docs/anomaly-detection-overview
• AI.DETECT_ANOMALIES https://docs.cloud.google.com/bigquery/docs/reference/standard-sql/bigqueryml-syntax-ai-detect-anomalies
• ML.DETECT_ANOMALIES https://docs.cloud.google.com/bigquery/docs/reference/standard-sql/bigqueryml-syntax-detect-anomalies
• Gemini in BigQuery overview, including validation and compliance cautions https://docs.cloud.google.com/bigquery/docs/gemini-overview
• BigQuery data preparation overview https://docs.cloud.google.com/bigquery/docs/data-prep-introduction
• BigQuery data insights https://docs.cloud.google.com/bigquery/docs/data-insights
• Looker Conversational Analytics overview https://docs.cloud.google.com/looker/docs/conversational-analytics-overview
• Document AI overview https://docs.cloud.google.com/document-ai/docs/overview
• Vertex AI generative-AI evaluation service https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/evaluation-overview

This article is original editorial work for a business audience. Product facts were checked against the official sources above. Any performance figures are targets to validate against your own baseline, not vendor guarantees.

Bring us the operating need, risk, or opportunity. We will connect the people, security, AI and enterprise systems required to act.