Human-in-the-loop reporting: designing finance workflows that use AI responsibly

How to move faster at month-end without sacrificing accuracy, auditability, or investor trust.

AI can make finance reporting dramatically faster: variance analysis in seconds, first-draft commentary in minutes, portfolio dashboards that update automatically.

But finance has a non-negotiable constraint that most "AI automation" conversations skip:

Speed does not matter if the numbers are not trusted.

In investor and lender reporting-especially across multiple SPVs-your output must be:

• consistent month-to-month
• reconcilable to the source system
• explainable under scrutiny
• approved by accountable humans

That is why the best reporting workflows are human-in-the-loop: AI accelerates analysis and drafting, while finance owns reconciliation, judgement, and sign-off.

This post lays out a practical blueprint for building responsible AI into your reporting process-without creating a new source of risk.

---

What "human-in-the-loop" means in finance reporting

Human-in-the-loop (HITL) reporting is a design pattern:

AI proposes. Humans approve. Systems log.

AI is used for what it is good at (pattern detection, summarisation, drafting), while humans retain ownership of what only humans can do reliably (materiality judgement, accounting policy decisions, stakeholder-facing sign-off).

This is different from:

• "Full automation" (fast, but risky when outputs are wrong or unexplainable)
• "No AI" (safe, but slow and inconsistent when teams are stretched)

HITL is how you get the upside without breaking trust.

---

Why responsible AI matters more in SPV portfolios

In a single entity, mistakes are easier to spot. In multi-entity SPV portfolios, small inconsistencies compound:

• different charts of accounts
• coding differences between properties
• capex vs opex drift
• intercompany movements
• changes in mapping from one month to the next

Now layer on AI, and you introduce new risks:

• "confident" narratives that do not match the ledger
• incorrect variance attribution
• missed exclusions (one-offs, reclasses, prior-period adjustments)
• inconsistent KPI definitions (NOI, capex buckets, finance costs)

So the goal is not "AI everywhere." The goal is AI where it is safe and useful, with controls that keep the reporting system defensible.

---

The responsible split: what AI should do vs what humans must own

AI is great for (high leverage, low risk when reviewed)

1) Variance detection and triage

• surface the biggest MoM/QoQ movements
• flag outliers by SPV, account, vendor, category
• identify concentration risk ("80% of movement came from 2 SPVs")

2) Drafting "what changed" commentary

• first-pass narrative that ties movements to drivers
• structured templates ("headline -> drivers -> so what -> actions")

3) Suggesting classifications and mappings

• propose where new accounts should map in a portfolio COA
• flag likely miscoding (capex in repairs, interest in bank fees)

4) Scenario re-runs and sensitivity summaries

• "what if occupancy -5pp?"
• "what if rates +150bps?"
• "what if capex slips 2 months?"

5) Pack assembly

• compile charts/tables
• generate consistent section headings and narrative structure

The key: AI generates recommendations and drafts-never the final truth.

Humans must own (the accountability layer)

1) Close and reconciliation

• bank recs
• debt rollforward checks
• key balance sheet reconciliations
• period lock discipline

2) Accounting policy and classification decisions

• capex vs opex rules
• gross vs net treatment
• one-offs vs run-rate categorisation

3) Materiality judgement

• what matters to investors vs what is noise
• what needs disclosure vs what does not

4) Final sign-off on anything investor-facing

• investor letters
• board packs
• covenant reporting
• audited/regulatory outputs

In short: AI can speed up the work, but humans own the consequences.

---

The 7 design principles for human-in-the-loop reporting workflows

1) One source of truth

AI should not "invent" numbers. Your workflow must be anchored to:

• actuals from the accounting system
• defined consolidation rules
• consistent KPI definitions

If you cannot trace a statement to data, it does not belong in the pack.

2) Traceability and drill-down

Every number and claim in commentary should be explainable as:
Portfolio line -> SPV -> account -> transaction

When investors ask "why?", your system should answer quickly-without spreadsheet archaeology.

3) Deterministic calculations, probabilistic commentary

Keep calculations deterministic (models, consolidations, rules).
Use AI on the interpretation and drafting layer, where humans can review.

4) Structured templates beat free-form writing

Give AI a fixed structure that matches how finance communicates:

• Headline
• Top 3 drivers (with - impact)
• Context
• Cash/covenants
• Actions + owners + timing

This prevents vague, overconfident language and keeps commentary comparable month-to-month.

5) Confidence thresholds + escalation rules

Define what triggers mandatory human review, e.g.:

• any unmapped accounts
• large reclasses
• big MoM variance beyond threshold
• covenant headroom below buffer
• cash runway below X months

AI can highlight these-but humans decide and approve.

6) Audit logs and versioning

A responsible workflow records:

• what data was used
• what changed since last period
• what the AI generated
• what the reviewer edited/approved
• when it was published and to whom

This is what makes AI use "finance-grade."

7) Least-privilege access and data minimisation

Only give AI access to what it needs:

• restrict entity scope
• restrict sensitive fields
• avoid unnecessary PII
• document retention rules

(And always align security practices with your internal governance and legal/compliance requirements.)

---

A practical month-end workflow: "AI copilot" with human control

Here is a simple blueprint that works well in real finance teams.

Step 1: Close and reconcile (human-owned)

• SPV-level close checklists completed
• period locked
• key reconciliations signed off

Output: "Actuals are stable."

Step 2: Consolidate and standardise (system-owned)

• multi-entity consolidation runs
• portfolio chart of accounts mappings apply
• unmapped accounts report produced

Output: "Portfolio numbers reconcile."

Step 3: AI generates analysis drafts (AI-owned, not final)

• top drivers of NOI movement
• cash runway movements and key risks
• outlier SPVs/assets/accounts
• first-draft "what changed this month" commentary

Output: "A structured draft + supporting evidence."

Step 4: Reviewer edits and approves (human-in-the-loop)

• validate each claim against drill-down
• add operational context AI cannot know (tenant event, contractor delay, insurance renewal)
• rewrite anything that is ambiguous or overly certain
• confirm what is material

Output: "Approved narrative."

Step 5: Publish investor/board pack (human sign-off)

• final pack generated
• approvals recorded
• distribution controlled

Output: "Investor-grade reporting."

---

The most important guardrail: "AI must show its working"

If you want commentary that builds confidence, require this rule:

Every narrative claim must reference a driver and a number.

Bad (unreviewable):

• "Costs increased due to one-offs."

Good (reviewable):

• "NOI down -82k MoM: -51k from lower occupancy in SPV 3 (refurb downtime), -23k from insurance renewal timing in SPV 7, -11k from higher interest after rate reset."

That structure makes review fast and makes the story defensible.

---

Where this fits into our platform approach

We are building a reporting layer for real estate SPVs that makes human-in-the-loop workflows practical at scale:

• one-stop visibility across multiple Xero or QuickBooks entities (SPVs)
• standardised COA mappings so SPVs roll up cleanly
• FP&A (budgeting, forecasting, cash planning)
• scenario planning (rates, occupancy, refurb programmes) with clear cash impact
• an "AI CFO / advisor" layer that drafts commentary like "what changed this month" and highlights risks-with human review and consistent logic on top

The point is not replacing finance judgement. It is removing the repetitive work so finance can focus on decisions and stakeholder confidence.

---

A quick checklist you can copy into your process

Before you ship any AI-assisted pack, ask:

• Do portfolio totals reconcile to SPVs?
• Are all accounts mapped (no silent "other")?
• Are KPI definitions unchanged from last period (or explicitly disclosed)?
• Does commentary quantify the top drivers (with - impacts)?
• Can every statement be drilled to SPV/account/transaction?
• Has a named person approved the final narrative?
• Is there an audit trail of changes and approvals?

If you can tick those off, you are using AI responsibly.