Daily positions
Quantity, local/base value, cost, accrued income, currency
Data Analysis Agent for Finance
An agent that finds what moved the numbers.
Connect positions, transactions, benchmarks, plans, actuals, and notes so the agent can trace performance, variance, and exposure questions to the records and definitions behind them.
The recurring problem
The answer is scattered across positions, plans, and notes.
Important performance and exposure questions depend on reconciliations across positions, transactions, benchmarks, finance systems, plans, and notes.
The representative scenario follows a quarter-end portfolio review: explain relative underperformance, test the attribution, then isolate the holdings behind a threshold exception. It uses synthetic positions, transactions, benchmarks, market facts, and notes. It is not investment advice or a production client result.
Contribution
Attribution
Benchmark
Forecast vintage
Materiality
Exposure
Concentration
Base currency
What the agent connects
Six sources the agent has to reconcile.
Illustrative dataset
Northstar Portfolio Office
A fictional multi-strategy investment and operating-finance team preparing a quarterly portfolio review.
- Date range
- 1 January 2025–30 June 2026
- Disclosure
- Fully synthetic portfolio, transaction, benchmark, note, and exception records. Optional SEC and FRED reference data would remain separately attributed.
Transactions
Trade/settle dates, security, quantity, price, fees, transaction type
Benchmark & market
Weight, return, price, FX, valuation timestamp
Security & company facts
Identifiers, sector, geography, asset class, filing facts
Investment notes
Issuer, thesis tag, author, date, review status
Limits & exceptions
Threshold, observed value, status, owner
Security identifiers connect positions, transactions, benchmark members, prices, company facts, and notes. Portfolio and account hierarchies define approved roll-ups.
Return, contribution, benchmark-relative performance, attribution method, exposure, concentration, base currency, cash-flow treatment, and materiality.
A security identifier changes after a corporate action and one benchmark constituent weight arrives a day late. The analysis must expose the mapping and valuation-date effect.
Watch the agent investigate
What dragged the portfolio?
Representative Interface · Illustrative Analysis. Every value uses synthetic records; this shows representative agent behavior, not a client result.
Portfolio review lead
Which holdings contributed most to underperformance this quarter, and was the difference allocation or asset performance?
The agent’s interpretation
Compare the Growth Portfolio with its approved benchmark for Q2 2026, use time-weighted performance in USD, include external cash flows, and apply the documented sector-allocation / security-selection method.
positions_daily
transactions
benchmark_daily
market_prices_fx
security_master
investment_notes
- Reconcile positions and cash flows
- Align security and benchmark identifiers
- Calculate contribution and relative effects
- Trace the largest effects to holdings, trades, and notes
Example Output
Three holdings explain most of the relative shortfall.
Illustrative analysis shows selection in Industrials as the largest negative effect, partly offset by allocation to Health Care. One result changes when the late benchmark weight is aligned to the correct valuation date.
Holding
Effect
Driver
Review
Aster Systems
−43 bp
Selection
Thesis update
Northline
−31 bp
Allocation
Weight timing
Cedar Health
+18 bp
Selection
No exception
Evidence & definitions
6 sources · USD base · Q2 2026 · benchmark definition v3 · synthetic records
Follow-up question
Recalculate after correcting the benchmark-weight date and show any concentration breaches.
The agent narrows the answer
The relative shortfall narrows by 7 bp. One holding remains 0.6 percentage points above the review threshold and is added to the human review list.
Attribution table · concentration exceptions · quarterly review reportReport / table output available in an implementation
Human decisionThe result informs the investment-review agenda; it does not recommend or execute a trade.
Where else the agent works
Three finance jobs for the agent.
01 Agent job Portfolio and investment performance A portfolio misses its benchmark, but the management pack does not explain which positions, sector weights, asset returns, or currency effects produced the difference.
People asking
Portfolio managers, investment analysts, CIOs, operating partners, and performance-reporting leads.
Real questions
- Which holdings contributed most to underperformance this quarter?
- Was the gap driven by allocation, asset performance, or currency?
- Which concentration or thesis assumptions changed materially?
Sources the agent uses
Positions and transactions · Security master and prices · Benchmark weights and returns · FX and investment notes
Work the agent performs
Reconcile opening and closing positions and external cash flows; calculate period returns and weighted contributions; compare approved portfolio and benchmark weights/returns; separate attribution effects under the selected methodology; trace notable movements to trades and notes.
Output it prepares
A contribution bridge, holding-level attribution table, concentration exceptions, and a review pack with the calculation definitions attached.
Human decision or workflow
Prepare the investment review and identify positions or assumptions requiring human discussion—not an automated trade instruction.
What must be validated
Benchmark, corporate-action, FX, fee, cash-flow, and attribution conventions must be agreed and validated against the official performance system.
02 Agent job Financial planning and management reporting Actuals miss forecast while the explanation is distributed across the general ledger, planning model, departmental files, and operating drivers.
People asking
CFOs, FP&A leaders, controllers, department leaders, and board-reporting owners.
Real questions
- Which departments explain the largest unfavorable variance to forecast?
- Which operating drivers changed?
- Is the variance timing, volume, rate, mix, or classification?
Sources the agent uses
General-ledger actuals · Budget and forecast versions · Cost-center hierarchy · Headcount, volume, price, and utilization drivers
Work the agent performs
Map actual and plan accounts, preserve forecast vintage, calculate amount and percentage variances, decompose material lines using approved drivers, and surface late journals or reclassifications.
Output it prepares
A variance bridge, department exception table, commentary draft, and management-reporting workbook.
Human decision or workflow
Assign owners, update the forecast, and determine which variances require intervention or explanation.
What must be validated
Account mapping, sign convention, forecast version, materiality, allocation, currency, and close status must remain visible.
03 Agent job Risk, exposure, and exception investigation A threshold or covenant monitor identifies an exception, but the underlying positions, counterparties, facilities, and source records still require investigation.
People asking
Risk managers, treasury, investment operations, finance directors, and compliance reviewers.
Real questions
- Where has exposure crossed an agreed threshold?
- Which positions or counterparties account for it?
- Is the change economic, transactional, or a data-quality issue?
Sources the agent uses
Positions and counterparties · Legal-entity hierarchy · Facilities and covenant schedules · Limits, collateral, cash forecasts, and exception logs
Work the agent performs
Aggregate through the approved entity hierarchy, normalize currencies, apply documented threshold and netting definitions, compare current and prior snapshots, and drill into transactions or missing classifications.
Output it prepares
An exposure matrix, threshold exceptions, underlying-record table, and assigned review queue.
Human decision or workflow
Escalate the exception, obtain missing evidence, correct data, or conduct a qualified risk/compliance review.
What must be validated
Legal netting, regulatory classification, ratings, covenant interpretation, and compliance conclusions remain with qualified people and systems of record.
From scenario to implementation
Begin with an agent that can follow the numbers across sources.
Working Data Analysis Agent
The reusable foundation already supports conversational investigation, cross-source joins, evidence references, permissions, and reviewable analytical outputs. Finance implementation starts by selecting the parts relevant to the review—not by rebuilding that analysis loop.
- Conversational questions and follow-up investigation
- Cross-source joins with visible evidence and definitions
- Charts, contribution tables, exception views, reports, and downloads
- Permissions, guidance, and configured or self-service connections
Configure the calculation contract.
The same question can produce a different answer when benchmark, valuation date, hierarchy, currency, materiality, or forecast version changes. Those choices become explicit configuration for the organization and its users.
- Position, transaction, ledger, plan, benchmark, market, and note connections
- Return, attribution, variance, exposure, threshold, and netting definitions
- Security, account, portfolio, department, counterparty, and legal-entity hierarchies
- Role-based source scope, review rules, report formats, and human decision boundaries
Keep the agent reviewable
Keep the method—and the decision—visible.
Validate positions, cash flows, prices, corporate actions, and benchmark versions against systems of record.
Make calculation methodology, valuation date, currency, fees, exclusions, and materiality visible with every output.
Scope access by approved portfolio, account, legal entity, and user; do not imply full role-level control where it has not been implemented.
Require qualified human review for investment, lending, credit, covenant, regulatory, and compliance conclusions.
A bounded client implementation
Start with one review the team already runs.
Choose the review and intended result
Select portfolio review, variance explanation, or exposure investigation; identify the accountable users; and record the current review effort, cycle time, or exception-resolution measure.
Name the recurring questions
Agree the management questions, materiality, benchmark or plan, and expected review output.
Connect approved evidence
Use representative or client-approved positions, transactions, finance data, definitions, and notes.
Define the calculation contract
Document valuation dates, cash flows, currencies, attribution or variance methods, hierarchies, and thresholds.
Set access and boundaries
Identify permitted portfolios, accounts, entities, sensitive records, and mandatory human review.
Configure the analysis and outputs
Shape the agent, contribution/variance tables, evidence view, and reporting format.
Validate known cases
Reconcile an agreed question set with official finance or performance outputs and document failure modes.
Evaluate and scope expansion
Review the working capability against the agreed questions and measures, then choose the next source, report, workflow, or application integration only where justified.
Define the result before deployment
Measure whether the review gets faster—and better supported.
Agree the current effort and review cycle before implementation, then evaluate whether the agent reduces time spent assembling evidence, accelerates recurring answers, and helps reviewers resolve material exceptions without weakening control.
The baseline, review period, and acceptable evidence are agreed before deployment; no improvement is assumed.
- Time to answer recurring portfolio or management questions
- Effort required to prepare review and reporting packs
- Reconciliation and exception-resolution cycle time
- Adoption, repeat use, and reviewed-output completion
Extend where the business requires
Carry the finding into the finance workflow.
Where analysis alone is not enough, Company A can extend the product with proprietary feeds, governed reporting applications, exception queues, approvals, pipelines, recurring monitoring, or connections to other agents and finance systems.
Possible next workflows
- Add planning and general-ledger sources
- Create governed management-reporting packs
- Connect exception review and approval workflows
- Build portfolio or finance applications around the analysis layer
- Operate recurring validation and definition reviews
What custom development can add
- Data-source integrations and pipelines
- Portfolio and management-reporting applications
- Exception and approval workflows
- Permission-aware internal interfaces
- Managed reporting and ongoing improvement
Identify a bounded first implementation