Nightworks AI

Your finance function is run by agents. Here is the evidence that survives your audit.

Close the books with a record behind every number and the controls memo; the auditor packet is planned and not yet available.

What the record has to carry

Every agent is measured nightly against its sources, every irreversible action waits on a person, and the record it leaves is one an owner can sign.

finance module · month-end closeExample data

Intercompany balance mismatch

$120.00

Which side of the LLC to Inc intercompany balance is wrong?

Field	Stated	Verdict
INC side balance at cutoff	1000.00
LLC side balance at cutoff	-880.00
Difference	120.00	Differs

If nobody acts, the pair stays held and the close cannot be signed today.

Finance · conflicting evidence · act today · period 2026-09

stopped · a person decides

Details

Example close. The two sides do not net to zero, so the item waits for a person to decide. Nothing has been posted.

Pairsyn-llc-ar-syn-inc-ap

Stateheld

Books touchednone

02 · The problem

You still own the close.

Connecting an agent to your books does not give it your accounting policy, your judgment, or permission to act. An agent with no controls around it fails in specific ways.

The specific failures

A model asked to reconcile will produce a reconciliation whether or not the data supports one, because producing a plausible answer is what generating text optimises for.

The same task run twice can take two paths to two answers.

It has never read your accounting policy.

Its reasoning is not evidence you can hand an auditor.

OpenAI documents persistent hallucinations; Anthropic has observed models disobeying direct instructions in simulated corporate stress tests. These are reasons to build controls around agents, not predictions that every agent will fail.

OpenAI research, September 2025 · Anthropic research, June 2025

Numbers you can stand behind

A confident answer is not enough to sign off on a balance.

At the close, you need the source records, the reconciliation, and a clear view of anything still unresolved.

Decisions that stay yours

Preparing an entry and posting it are different decisions.

You need to see what will change, why it is needed, and who is approving it before it reaches the books.

An audit package. A board-ready story.

The close is not finished when the journal entries are posted.

Your audit package needs the supporting records and approvals. Your board reporting needs a clear account of performance, material changes, and open risks. Both should start from the same numbers you signed off on.

03 · The system

Every agent is graded by something that does not work for it.

Night Auditor, Agent Grader, Authority Ladder, and Operator Override: independent checks, clear permission, and control in your hands.

Night Auditor

You find out that night, not at quarter end.

An independent auditor reconciles every agent's work against source records while you sleep.

Agent Grader

The grade comes from the record, not the agent's self-report.

Every agent is graded by something that does not work for it.

Authority Ladder

Nothing irreversible happens on an agent's own authority.

Preparation is separate from permission, and permission is a recorded human decision with a limit attached.

Operator Override

Pause or stop any agent or workflow managed through Nightworks AI.

Stop for a test, investigate an unexpected result, or intervene when work moves outside your intent. The control belongs to you, whether you need to halt one activity or all connected work. Stopping is a control, not a request for the agent to reconsider.

04 · The methodology

Four questions a close record has to answer.

They are set out on the How it Works page, because an auditor cannot take a model's reasoning.

How it Works

05 · Proof

Proven in production before it was sold.

This platform runs live operating companies, with real money and real month-ends. Nightworks AI itself runs on the platform as tenant zero.

What that means

That is the standard for your trial, too: your source records, your approval boundaries, and your exceptions. A polished demo cannot tell you whether an agent can handle a missing invoice, a disputed balance, or a close that cannot wait.

The value is in what your team can use afterward: the reconciliation behind a number, the explanation behind a correcting entry, and the supporting material for an audit package or board report. Judge the work by what you can inspect and sign, not by how confidently an agent describes it.

Read the Paw Brands case study.

06 · Versus

How the alternatives handle your close.

Doing it manually has real costs. Pasting a bank statement into a chat window has different ones. Building your own system has others. Nightworks AI has to beat all three.

Manual

Doing it manually today

Familiar tools, experienced people, and direct control. Also another month of exports, spreadsheet checks, and chasing approvals.

The tradeoffs

A manual process can be the right choice when volume is low or judgment dominates. As work grows, the burden is still yours: handoffs, coverage when someone is out, reconciling versions, and assembling the audit trail after the fact.

Chat window

Pasting a statement into a chat window

Someone on your finance team has probably already pasted a bank statement into a chat window. Finance teams try the chat window first, and it usually happens before anyone decides to adopt it.

The tradeoffs

It answers with confidence whether or not the statement supports the answer.

Ask it twice and you can get two different reconciliations, and neither is a record you can hand an auditor.

Build it yourself

Building your own system

Maximum flexibility and ownership, with your team responsible for everything after a first success.

The tradeoffs

An internal system can fit your business precisely. It also needs someone to maintain integrations, keep context current, test model changes, enforce permissions, investigate failures, and provide support during the close. That commitment starts when the system is live, not when a first demo succeeds.

Nightworks AI

Why Nightworks AI

You get a close you can sign: the evidence is attached, and every exception waits for a person.

The tradeoffs

Every exception reaches you as a typed item with its evidence, and posting waits on your recorded decision.

An unresolved exception stays held and comes back to a person, not closed by an agent saying the work is done.

The record is built as the work happens: what was requested, the evidence, who approved it, and what actually changed.

07 · Trust

The evidence has to survive the person who signs.

No agent spends, files, or contacts anyone without a recorded human decision; the permission to do otherwise does not exist.

What control means in practice

A useful record connects what was requested, the evidence behind it, who approved it, and what actually changed.

08 · How engagement works

Start with one recurring workflow.

This is priced against standing work, not against five days at month-end: the people who reconcile, chase exceptions, assemble the packet, and sign. Put one agent to work on a real process, then decide from the results. The 30-day trial starts as soon as deployment is complete.

Deploy one agent

One agent implemented and deployed in your business within 30 days.

Prove the workflow

From the day it is deployed, it runs for 30 days in your environment under your own recorded limits.

Keep it or walk away

At the end of those 30 days you pay or you separate. You do not pay for the build. The decision is about the agent and workflow you trialed: the work completed, the exceptions raised, and the value to your team.

Talk through your workflow.