Dark Factory
non-engineering
Lights-out operation of non-software functions — finance close, claims, procurement, analytics production — is technically nearer than anyone in those functions believes and further than any vendor claims; the binding constraint is that no regulator or auditor has yet said what an unattended process must look like.
Clustered only. No lab work behind it. Cannot be cited.
Confidence
40%unresearchedExpiry
85duntil review · 27 Nov 2026Lead time
—not yet mainstream · opened 14 May 2026Ownership
Unownedcandidate — a named human electsWhere it is
The dark-factory idea moved from software (Dark Harness, on the Near horizon) to everything else this year, mostly in vendor decks. The evidence is a handful of partial cases: a month-end close where 80% of journal entries are agent-prepared and human-approved; a claims flow where straight-through processing on low-complexity claims passed 90%; an analytics team producing standard reports unattended. None is lights-out. Every one keeps a human on the approval step because nobody knows what an auditor will accept. In Australia CPS 230 makes the question sharper: an unattended material process is an operational-risk event waiting for a definition. The field is a candidate because the technology is not the gate; adoption is.
Why a Quantium decision hinges on it
Quantium's own analytics production is a non-engineering function, and the firm's clients run the finance, claims and procurement functions the field is about. If the adoption gate opens, the consulting question becomes 'which of your functions goes dark first and what does the control environment look like', and that is a question the firm should be able to answer with evidence rather than a vendor's slide.
Field attributes
Position
What is demonstrated, what is hype, what would have to be true.
The shape every position request answers. Signal-tier fields carry a draft; assessed and tested fields carry a validated one.
- 01Month-end close with 80% of journals agent-prepared and human-approved at a mid-size company; close shortened from 8 to 3 days.
- 02Straight-through processing on low-complexity claims above 90% at one AU insurer, with human review sampled rather than universal.
- 03Standard report production run unattended for a quarter inside the lab's own analytics pipeline, with a human on exceptions only.
- 01'Autonomous finance' decks that describe human-approved automation as lights-out.
- 02Headcount-reduction projections derived from task automation rates without accounting for the control roles that replace them.
- 03Any claim of unattended operation in a CPS 230 material process without a named control framework.
- 01A Big Four audit opinion, or an APRA guidance note, that describes what an acceptable unattended control environment looks like.
- 02A published lights-out case in a regulated function with the exception rate, the control design and the auditor's view on record.
- 03Delegated authority for agents actioning financial systems — the Auth Broker pattern — accepted by at least one AU bank's risk function.
- 01Run the lab's own analytics production dark for two quarters as the reference case; publish exception rate and control design.
- 02Watch for the audit/regulatory trigger; when it fires, take the reference case to the banking sector owner as a proposal.
- 03Keep the 'human-approved is not lights-out' distinction in every client conversation.
Signals · 9 in this cluster
What the cluster is made of.
Every item carries its source, tier and sightings. Detector-found signal sits beside human drops; downstream they are indistinguishable except by provenance.
Lab analytics production run dark since June: 3.8% exception rate, one human on exceptions
The lab's standard measurement reports have been produced unattended for eleven weeks. Exceptions route to one person; 3.8% of runs needed a human. No control review yet, so tried tier despite the duration. The reference case the field would be built on.
extracted claimAnalytics report production can run unattended for a quarter at an exception rate under 5%.
Big Four hiring 'AI Process Assurance' partners in Sydney and Melbourne
Partner-level assurance roles for automated processes. Argus inference: the audit profession is preparing to opine on unattended processes, which is the precondition for the trigger.
'If the agents do the close, who do I fire?'
Asked by a CFO in a workshop, half-joking. Answered with the control-roles paper. Logged because the question reveals the headcount frame the field will be bought on and the frame the evidence does not support.
'The autonomous enterprise: finance without people by 2028'
A large consultancy's position piece projecting lights-out finance within two years. No control framework, no auditor, no regulator mentioned. Kept as the competitor position our evidence contradicts — an Argus positioning delta.
Analyst survey: 0 of 120 finance functions report lights-out operation; 38% report agent-prepared journals
Survey of 120 finance functions. Human-approved automation is common; unattended operation is nonexistent. The number that anchors claim 2.
'Our close went from 8 days to 3. Here is what the auditors said.'
Mid-size company CFO on an agent-prepared close: 80% of journals agent-drafted, every one human-approved. Auditors accepted it because the approval step is unchanged. Asked about lights-out, said 'not while I sign the accounts'.
extracted claimAgent-prepared, human-approved close is in production and cuts close time by more than half; lights-out is refused by the signer.
Control Roles After Automation: Why Task Automation Rates Overstate Headcount Effects
Firm-level study of 40 automation programmes: for every three doer roles removed, one control or exception role is created, and it is more senior. The paper behind the 'uncounted control roles' claim.
Major ERP vendor ships 'autonomous close' agents with mandatory approval gates
Agents for journal preparation, reconciliation and variance analysis. Every action lands in an approval queue; the vendor's own docs say the gates cannot be disabled. The product is human-approved automation with a lights-out name.
extracted claimERP vendors are shipping approval-gated automation under lights-out branding.
APRA CPS 230 FAQ: no position yet on unattended automated processes
Updated FAQ addresses AI in material services only to say existing operational-risk requirements apply. No definition of an acceptable unattended process. The absence is the gate.
Claims · 4 supporting, 1 refuting
The atoms.
A document cannot go stale; an assertion can. Claims are immutable and stamped with the extractor that produced them, so staleness, diffs and the graveyard operate at claim level.
No regulated function anywhere has a published lights-out case with the auditor's view on record; every case retains a human on approval.
Agent-prepared, human-approved finance close is in production at mid-size scale and cuts close time by more than half.
Task-level automation rates overstate the headcount effect because the control roles that replace the doers are not counted.
The lab's own analytics production can run unattended for a quarter with a human on exceptions only, at an exception rate under 5%.
Unattended operation of a material process is achievable under CPS 230 with current controls.
Position history · the diff is the product
2 validation runs against a fixed brief. Confidence 36% → 40%.
Lab's own analytics production is the reference case, running dark since June. Headcount claims overstated by uncounted control roles. Red team's marginal-value argument is the open question.
- The lab's own analytics production can run unattended for a quarter with a human on exceptions only, at an exception rate under 5%.
- Task-level automation rates overstate the headcount effect because the control roles that replace the doers are not counted.
- c-dark-factory-3 ↓ 0.22 → 0.15
Scoring · ordinal bands
Agents propose. A named human commits.
Uncommitted scores are visibly marked and never leave the building. Bands, not point estimates — false precision is the tell that a number was generated rather than derived.
Impact
committed · AWEvery client runs the functions in scope; the firm runs one itself.
Timeline
committed · SKThe technology is closer than that; the audit and regulatory definition is what sets the date.
TAM
agent-estimatedAgent-estimated on back-office BPO substitution globally; the consulting slice is not separable. Uncommitted.
Cost
committed · JPThe reference case is two quarters of the lab's own production run dark, plus an assurance review.
Cost of being wrong
committed · LFA lights-out failure in a client's material process is a reportable operational-risk event under CPS 230.
Demand
committed · CDCFOs ask about close automation; nobody asks for lights-out and the ones who hear the phrase step back.
Workforce readiness
agent-estimatedNo delivery team has designed a control environment for an unattended process; the reference case is the lab's own. Agent-estimated.
Relevance · per vertical
Why it matters here, or explicitly does not.
Ranking is per vertical, not global. Sector owners commit notes against agent drafts.
Finance close and reconciliations are the natural first function; CPS 230 makes unattended operation of anything material a defined risk without a defined control.
Mechanism · Agent-prepared, human-approved close as step one; lights-out only after the audit trigger fires.
Straight-through claims is the closest thing to a dark function in a served vertical, and the exception rate is already measured.
Mechanism · Extend STP to medium-complexity claims with sampled review; exception analytics as the control evidence.
Unattended decision-making about citizens is the Robodebt problem restated; the adoption gate here is political, not technical.
Mechanism · Would need a legislative basis for automated decisions with a human-in-the-loop guarantee; none exists in a usable form.
The firm's own analytics production is the reference case and the cheapest place to learn what dark operation costs.
Mechanism · Standard report production unattended; exceptions to a human; exception rate published as the evidence.
Red team · the strongest case against
The strongest case against: the field is a relabelling of process automation that has been happening for twenty years, and 'dark' adds nothing except a headline that scares the buyer. The economically meaningful step — human-approved automation — is already table stakes, and the last step to lights-out has almost no marginal value because approvers are cheap and auditors are not.
- —The marginal value of removing the last approver is small; the marginal cost in audit and regulatory exposure is large. The field may have no economic reason to exist.
- —Every vendor and every client already does the first 80%; the lab is watching a gate nobody needs to pass.
- —Government's adoption gate is political and permanent after Robodebt; the field is not watchable there, it is closed.
Source diversity
- Practitioners / CFOs30%
- Vendor / consultancy25%
- Regulatory / assurance20%
- Internal / Engel25%
A field supported by one epistemic community is a flag, not a finding.
Cross-pollination · typed joins
Connected, not merely similar.
Enabling, compounding, substituting, blocking. A satisfied dependency trigger is a far stronger signal than semantic proximity.
Trigger · A dark-harness software function runs unattended for 30 consecutive days with an error rate at or below the human baseline and a published control design; that is the template a non-engineering function would copy.
When the trigger fires, this field is resurfaced automatically. Watchable rather than parked.
Trigger · x-auth-broker-poc or a successor is accepted by an AU bank's risk function for an agent actioning a financial system under delegated authority; Argus tracks APRA and Big Four publications for the first audit opinion on an unattended material process.
When the trigger fires, this field is resurfaced automatically. Watchable rather than parked.
Lights-out software delivery is the same problem in the function where the controls are best understood; it goes first.
An agent actioning a ledger needs delegated, scoped, expiring authority; without it there is no dark anything in finance.
The reference case is the cleanest ROI measurement the firm can run on itself.
Dark functions are the mechanism by which labour pricing changes; this is the concrete end of that field.
Share graph
Provenance running forward.
Discovery, not accountability. No counts, no rankings, no rollups to managers.
Convergence · who else is here
- JPJun Park · Measurement (Nightingale)2 drops
- CDClaire Dubois · Sector owner · Banking1 drop
- SKSam Kowalczyk · Research engineer · SDLC1 drop
- LFLena Fischer · Red team & assurance1 drop
- RMRohan Mehta · Exec sponsor1 drop
- MLMarcus Lee · Delivery lead · Telco1 drop
Several people’s drops meet here. An informal working group already exists and probably does not know it.
Lineage
What this field produced, and what it killed.
Experiments, recommendations and graveyard entries stay attached. The reasoning that killed a claim is the reusable asset.
No experiments, recommendations or graveyard entries yet. That is what a candidate looks like.
Open questions · return to the pile
Every run leaves a record. Separately, its question either closes or returns to the pile with notes — which is what the next person proposing the same thing will see.
- 01What is the marginal value of removing the last approver, in any function, net of the audit exposure it creates?
- 02What exception rate and control design would a Big Four auditor accept for an unattended close, and has anyone asked them?
- 03Is government permanently closed to this after Robodebt, or is there a legislative form that opens it?