WORK / CASE STUDY
Azure multi-agent architecture for fund data processing
We designed an Azure-based multi-agent system that integrates with the Xero API to streamline financial reporting across multiple businesses and Xero accounts within a fund — replacing manual consolidation with orchestrated AI agents that gather, validate, and report financial data.
The challenge
A private equity fund was consolidating financials across a portfolio of businesses, each with its own Xero organisation. Month-end meant downloading exports, reconciling inconsistencies by hand, and rebuilding the same reports in spreadsheets — slow, fragile, and impossible to audit cleanly when something looked wrong.
Private equity reporting looks simple from the outside: pull numbers from each company, add them up, produce the pack. In practice, every portfolio business had its own Xero organisation, its own chart of accounts quirks, and its own timing for when books were ready. The fund’s finance team spent the end of every month stitching exports together, chasing missing journals, and re-checking totals that never quite matched the last version of the spreadsheet.
The brief was not to replace judgement with a chatbot. It was to replace the mechanical gathering, normalisation, and cross-checks with something reliable on Azure — so the team could spend their time on exceptions and decisions, not on copy-paste.
What had to change
Three constraints shaped the design. First, each business stayed in its own Xero tenancy — there was no appetite to merge ledgers or force a single chart of accounts overnight. Second, numbers that feed investor reporting had to be traceable: if a line looked wrong, someone needed to see which organisation, which account, and which extract produced it. Third, the system had to tolerate incomplete data. Some companies close early; others lag. A brittle all-or-nothing job that fails when one subsidiary is late would be worse than the spreadsheets.
That ruled out a single prompt that “just reads Xero and writes a report.” We needed orchestrated steps, structured outputs, and clear handoffs between automation and humans.
An Azure multi-agent architecture
We built the solution on Azure with a deterministic orchestration layer and specialised agents for the work that benefits from judgement. Orchestration owns sequencing: authenticate to each Xero organisation, pull the required periods, wait or skip when a company is not ready, and assemble a run record that auditors can follow. Agents own the messy middle — mapping inconsistent account labels, flagging anomalies across entities, and drafting narrative commentary for the consolidated pack.
Integration with the Xero API was treated as a first-class system boundary. Credentials, rate limits, and per-organisation scopes live in managed Azure identity and configuration — not in agent prompts. Agents receive only the structured extracts they need for the current step, and they return structured results the pipeline can validate before anything reaches a report. Where a mapping or variance cannot be resolved confidently, the workflow stops for human review rather than inventing a number.
Gather, validate, report
In a typical month-end run, the pipeline first gathers ledger data from each Xero account in the fund. Validation agents then check completeness, period alignment, and known mapping rules, raising exceptions for anything that looks off against historical patterns or intercompany expectations. Only after those checks pass — or after a reviewer clears an exception — does the reporting stage produce consolidated views and draft commentary for the fund pack.
The important pattern matches how we design other agentic systems: agents interpret and propose; deterministic code and human approval own anything that must be correct every time. Retries re-read saved extracts and decision records, so a failed step does not ask a model to reconstruct what happened last week.
What it unlocked
The fund moved from a fragile, person-dependent consolidation ritual to a repeatable Azure process that scales as portfolio companies are added. New Xero organisations plug into the same gather–validate–report loop. Finance still owns the calls that matter — especially where mappings are ambiguous or a variance needs explanation — but they no longer rebuild the machinery of consolidation every month.
That is the outcome that matters for this kind of work: not a demo of AI reading a balance sheet, but a production reporting path the fund can trust when the numbers have to be right.
Outcomes
- Month-end consolidation reduced from days of manual spreadsheet work to an orchestrated Azure pipeline with human review only where it matters
- Consistent, validated financial extracts across every Xero organisation in the fund — with clear lineage from source ledger to final report
- A multi-agent architecture the fund can extend to new portfolio companies without rewriting the reporting stack