PRISM Methodology
PRISM 01Finance

Finance AI, built in the right order.

The finance org we usually meet runs several ERPs that don’t talk to each other, a consolidation tool that trusts none of them, and a close held together by manual effort. You don’t fix that by pointing a model at it. You fix what the data means, then the workflows — and last, the agents.

Where we stand

Five positions most vendors won’t take.

The opinions underneath every finance engagement we run.

01

Copilots make analysts 10% faster at work that shouldn’t exist.

02

A fast close is a governance metric, not an efficiency one.

03

Put AI where the ERP ends — not on top of it.

04

Auditability is the moat. Accuracy is just table stakes.

05

Working capital is the highest-ROI AI project in finance — and nobody funds it.

The gap in the market

Everyone has half the answer.

Finance AI keeps failing in translation — between the people who understand the close and the people who can ship software.

The consultancies

Speak finance. Can’t ship.

They know the close inside out — and deliver decks, pilots, and point solutions that don’t survive contact with production, the next audit, or the next acquisition.

The engineering firms

Ship software. Don’t speak finance.

They build beautifully — and lose the room the first time a CFO describes a carve-out mid-TSA with three reconciliation platforms to consolidate. You can’t automate a workflow you can’t parse.

How LightCI is built

Every engagement pairs a finance operator with a forward-deployed engineer.

Big 4-trained finance and tax people who have sat in the seat, alongside engineers who ship production systems. The operator translates record-to-report pain into requirements; the engineer builds against them. Same room, every working session.

Two modes

Business as usual — and the roll-up window.

The same method serves two very different moments in a portfolio company’s life.

01

Continuous operations

The monthly rhythm of the finance org — the close, reporting, collections, payables. Agents live inside recurring workflows and compound month over month.

02

The roll-up window

Post-close integration — new entities, redundant systems, deadlines measured in weeks. The work is manual and time-sensitive, so every week compressed pulls value creation forward. And the spend is a one-time cost, at exactly the moment the layer is cheapest to install.

Data readiness

Finance data is tribal knowledge. We encode it.

Plugging an agent straight into a stack of ERPs works one time in ten. So we start where the complexity lives: what the data means. That knowledge sits in your team’s heads — we encode it into a semantic layer, so agents reason over curated, governed data. Raw exports never touch a model.

SOURCE SYSTEMSBUILT WITH YOUR SMEsAGENTSERP · multiple instancesConsolidation systemSpreadsheetsBilling & payrollBank feedsContracts & PDFsThe semantic layerENTITY & COA MAPPINGFX & POLICY LOGICDEFINITIONS FROM YOUR TEAMPERMISSIONS & AUDIT TRAILCLOSE AGENTSCOLLECTIONS AGENTSAP AGENTSBOARD-PACK AGENTS
01

Sit with the knowledge workers

Working sessions with the controllers, analysts, and AP clerks who own each number. What a field actually represents, which system wins a conflict, why the exception exists — captured, not assumed.

02

Curate in layers

Mappings, definitions, and business logic built up in layers AI can reason over — chart-of-accounts translation, entity structure, FX treatment — instead of raw table dumps and hope.

03

One layer, every agent

Every agent reads and writes through the same layer: consistent definitions, scoped permissions, a full audit trail. Add an agent, and it inherits the understanding.

The roll-up dividend

The layer pays for itself before any AI touches it — reporting gets cleaner on day one, and the next acquisition inherits the layer instead of rebuilding it.

The operating model

Agents that do the work — not copilots on top of it.

Software agents run the workflow end-to-end — matching, reconciling, drafting, posting — with the controller staying in control. Some live in the daily and monthly rhythm of finance; others exist for the exceptional windows: audits, carve-outs, integrations.

Record to report

Close orchestration · consolidation · flux narrative

Order to cash

Collections · DSO · dispute triage

Procure to pay

Invoice capture · 3-way match · touchless AP

FP&A & board

Variance · scenarios · the board pack

Treasury

Cash forecasting · sweeps

Tax & compliance

Multi-entity filings

Procurement

Spend & vendor intelligence

Audit & controls

Exception monitoring

Architecture first, agents last

Where an existing system already solves the problem, we don’t build an agent. We map where current applications win, where the semantic layer must exist, and only then where agents earn their place — model- and platform-agnostic by design. Not everything needs a model, either: the most rules-heavy workflows pair deterministic automation with an LLM only where judgment lives.

Worked example

Watch an invoice code itself.

The classifier proposes the full coding with a confidence score — above threshold, it posts untouched.

Invoice classifier — AP inbox
Incoming invoiceRECEIVED

Meridian Surgical Supply

INV-20419 · Net 30

$148,250.00

Arthroscopy towers with endoscopic camera systems — qty 4, delivered to Clinics East.

Proposed coding
GL account1520 — Surgical Equipment
Cost centreCC-114 · Clinics East
Asset classMedical equipment
CapitalizationCapitalize · 7-year useful life
Confidence0.88
0.85

Above the 0.85 auto-approve threshold. Posted to the ledger — logged, reversible, attributable.

By month four, 80–90% of invoices clear without a human touching them — every decision logged.

Closed-loop learning

Every correction trains next month’s model.

01Classify02Human reviewsexceptions03Correctionscaptured04Model retrainsmonthlyCONFIDENCE THRESHOLDTIGHTENS EVERY CYCLE

What the audit finds

Where a federated finance team’s hours actually go.

Measured with the finance team before anything gets built — and remarkably consistent across multi-entity orgs.

Invoice coding & post-coding review28%
Reconciliations & accruals22%
Reporting roll-forward18%
Payroll & labour journals15%
Ticket triage12%
Genuine analysis5%

The punchline

The first three categories are two-thirds of finance hours — and almost none of it is analysis. The team you hired to think spends its month moving numbers between systems.

The commercial shape

An audit, then fixed price per use case.

No seats, no open-ended retainers. Every statement of work names a metric — and payment is tied to hitting it.

01~3 weeks

The audit

Two tracks: where the data lives — ERPs, consolidation systems, the spreadsheets around them — and where the hours go, ranked with the finance team. The access we need is to people, not systems. Output: a ranked backlog — recurring, high-frequency workflows first, each with a quantified prize.

02Per use case

The build

Each statement of work names the metric, the timeline, and the total cost — fixed price, with payment tied to hitting the metric. The first build stands up the semantic layer that everything after it inherits.

03Managed → handed over

The run

Guardrails and security prep with IT before any agent gets authority. A managed run keeps the agents maintained while we train the internal team — and hands over when they’re ready to own it.

Governance in finance

The audit log is non-negotiable.

In finance, auditability is the price of admission.

The audit log

Every decision, override, and version logged — a record, not a recollection.

Confidence is first-class

Every output carries a score; every score carries a threshold.

No new system of record

Your ERP stays your ERP — agents post back through controls your team already trusts.

The CFO’s dashboard

One weekly view. Five numbers that don’t lie.

The numbers the fee is accountable to — not adoption theater.

Hours vs baselineBy workstream — the number the fee is priced against.
Auto-code rateTrending toward 80–90% by month four.
Exception-queue agingHow long flagged items sit before a human clears them.
Days-to-close trendThe governance metric, month over month.
Spend per workflowCost per invoice coded, per commentary drafted.

One rule

Every number on it has a measurement methodology the auditors have seen.

Case studies

The methodology, applied four times.

Anonymized programs across four industries — the shape of the result holds.

The Close · Record-to-Report

Sponsor-owned outpatient healthcare platform

~€450M revenue · 180 sites · 40+ legal entities · 30+ bolt-ons

Close orchestration across every entity — automated eliminations, tolerance flags, a drafted flux narrative each morning.

14→5
business days to close
Day 6
board-ready numbers
~3 FTEs
redeployed to analysis
Weekly
reforecasts, up from monthly

Order-to-Cash

Industrial distribution roll-up

~$1.2B revenue · 14 business units

A collections agent that scores risk, classifies disputes, and hands every collector a ranked worklist each morning.

62→49
days sales outstanding
~$26M
working capital released
accounts per collector
−40%
aged >90-day balance

Procure-to-Pay

Multi-site consumer services platform

~$600M revenue · 250+ locations · 6–8 acquisitions a year

Touchless AP end to end. New acquisitions plug into the pipeline instead of adding headcount.

78%
of invoices fully touchless
11→3
days invoice-to-posted
Flat
AP headcount across 7 acquisitions
$1.2M/yr
early-pay discounts captured

FP&A & the Board

Sponsor-backed B2B software company

~$120M ARR · first institutional CFO

A reporting agent that writes the variance narrative in house style — and answers scenario questions live in the boardroom.

5d→½d
board-pack prep time
2 FTEs
freed from roll-forward work
Live
scenario answers in the boardroom

Ready to move

Start with the audit.

Three weeks with your finance team. We come back with where the hours go, what the data looks like, and a ranked backlog — each item with a metric, a timeline, and a fixed price.

Talk to LightCI