Skip to content
ThetaLayer

Solutions · Finance operations

Know what your AP agents posted, changed and paid

Invoice, vendor-master and payment agents work across a shared mailbox, an AP tracker and the ERP. ThetaLayer reconstructs every run across all of them, checks each posting, bank change and payment against its mandate, and tells you what each completed invoice cost.

Evidence from
  • AP mailbox audit
  • AP tracker workbook
  • ERP change documents
  • Approvals
  • Payment runs
Evidence chain
Invoice €4,312.50 · AP agent
Posted invoice to ERP
Agent reported
timeout
→
ERP change log says
invoice posted NL01-5200004276 +2.4s
ContradictedLinked by business key · 50%
Posted invoice to ERP · retry
Agent reported
success
→
ERP change log says
invoice posted NL01-5200004277 +2.3s
ConfirmedLinked by business key · 75%
Same supplier invoice posted twice — both paid
Duplicate posting · high severity · routed to AP owner for review
Illustrative · demo environment, simulated data

The workflow

One invoice, six steps, four systems

An AP agent's work is spread across systems that each see only their own part. ThetaLayer stitches the pieces into one run — and records which system vouches for each step.

  1. 01
    AP intake agent
    Supplier email read
    Recorded in
    Mailbox audit log

    Message opened by the agent's identity. Message ID hashed; body and subject dropped.

  2. 02
    AP intake agent
    Logged in AP tracker
    Recorded in
    Workbook activity log

    Row appended by the agent's service principal. Cell values never leave your network.

  3. 03
    AP intake agent
    Vendor checked
    Recorded in
    ERP vendor master

    Lookup against the vendor record. New vendors or changed bank details go to the vendor-master agent.

  4. 04
    AP intake agent
    Posted or parked
    Recorded in
    ERP change documents

    Up to €10k posted; above it, parked for approval. Linked on hashed vendor + invoice number.

  5. 05
    Finance controller
    Approved
    Recorded in
    Approvals system

    Decision, approver identity and time — often hours or days after the agent's run ended.

  6. 06
    Weekly payment run
    Paid
    Recorded in
    ERP payment run

    Invoice cleared for payment. Late evidence is linked back to the original run automatically.

Every step is linked to the run with an explicit method and confidence. Uncertainty is shown, never hidden.

Direct reference
95%
Business key
85%
Time window
70%

What ThetaLayer shows you

The failures that cost money, found in the evidence

What ThetaLayer surfaced in an AP demo environment: three agents, two agent platforms, one ERP. Every finding comes with the run, the records and the link behind it.

Ledger · Duplicate posting

The same invoice, posted twice and paid twice

The agent reported a timeout and retried. The ERP shows the first attempt posted after all — and so did the retry. Both documents went out in the payment run.

Evidence
Agent: timeout → ERP: posted · Payment run: two payments, one invoice
Mandate · Approval bypass

Invoices over €10k posted without approval

A new agent version misread “12.450,00” as 12.45 and posted four invoices above the threshold directly, skipping park-and-approve.

Evidence
ERP: posted, never parked · Approvals: no request on record
Mandate · Vendor bank change

Bank details changed without approval — two to the same IBAN

Three vendor bank-account changes had no approval request. Two different vendors were moved to the same new IBAN: the pattern behind invoice-redirection fraud.

Evidence
ERP vendor master: bank field changed · Hashed IBAN matched across vendors
Ledger · Unattributed write

Agent credentials used at 02:13 on a Sunday

ERP postings were made interactively with the AP agent's technical user — with no agent run behind them. Someone else used the agent's identity.

Evidence
ERP: dialog posting by agent user · Runtime: no matching run
Scorecard · Cost regression

€0.70 → €2.40 per completed invoice after a version change

A new agent version added a retry loop on multi-line invoices. Cost per completed invoice more than tripled, and the scorecard showed it in the first week, not at quarter end.

Evidence
Runtime spans: repeated extract and validate calls · Tokens priced per model
Scorecard · Routing

Send standard PO invoices to the cheaper agent

A vendor agent completes standard PO invoices for €0.57 against €1.85 at a similar completion rate — but completes only 64% of multi-line invoices without a takeover. Route by invoice type, not by vendor.

Evidence
Scorecard by invoice type · Saving labeled as modeled

Examples from our demo environment with simulated data. The control plane isn't told where the problems are; it finds them in the evidence.

Scorecard

Catch a cost regression in its first week

Cost per completed invoice is tracked by agent, version and invoice type. When a release changes the economics, the scorecard shows when, by how much and which runs caused it — down to the retried model calls.

  • Version-by-version cost per completed invoice
  • Per-run and monthly spend caps with burn-down
  • Modeled savings always labeled as modeled
Scorecard
Cost per completed invoice · weekly
Cost / completed
€1.22
Straight-through
79%
Evidence coverage
99.7%
v2.4.0
AP agent · DACHAP agent · Benelux
€0.70 → €2.40 per completed invoice after v2.4.0 shipped. Staying on v2.3.0 would have cost ~€406 less.
Illustrative · demo environment, simulated data

Mandate

Test a control against history before anyone enforces it

Write down the rules your AP policy already contains — approval thresholds, bank-detail approvals, spend caps, identity — and ThetaLayer checks every historical run against them. Change a rule and the whole history is re-checked.

  • See how many runs a rule would have stopped
  • Separate agent, person and credential in every check
  • Draft first, enforce later at the tool boundaries you choose
Mandate · AP Intake Agent
v1 · L3 act within limits
  • Allowed tools
    9 tools
    100% pass
  • Approval above
    > €10,000
    4 failed
  • Bank details need approval
    bank_account
    3 failed
  • Spend cap per run
    €2.50
    22 failed
  • Acts under own identity
    required
    pass
Edit a rule and every historical run is re-checked against the new version.
Illustrative · demo environment, simulated data

Mandates that fit this work

Your AP policy, written as rules an agent can be checked against

Most of these rules already exist in your finance policy. ThetaLayer turns them into versioned mandates and checks every run, past and future, against them.

  • Versioned: edit a rule and every historical run is re-checked
  • Detection first: breaches become exceptions for a named owner
  • Enforcement later, only at the tool boundaries you choose
Mandate · AP Intake AgentL3 · act within limits
  • Approval above threshold
    amount > €10,000 → park + approve

    Invoices above the threshold must be parked and approved before they're posted.

  • Bank details need approval
    vendor.bank_account

    No bank-account change without an approved request from the vendor master data steward.

  • One posting per invoice
    unique(vendor, invoice no.)

    Checked on the hashed business key, so duplicates surface before the payment run.

  • Allowed tools
    mail · sheets · erp · approvals

    Only the tools the workflow needs. Anything else is an exception for the owner.

  • Spend caps
    per run · per month

    Runaway runs and monthly overruns are flagged, with a burn-down the owner can see.

  • Acts under its own identity
    service principal required

    No delegated clerk credentials, so the ERP can tell agent from human.

Example mandate · adapted to your policy during the pilot

Metrics on the scorecard

The numbers behind every AP agent decision

Defined once, versioned, and traceable to the runs and records behind them — comparable across agents, vendors and versions.

M01Cost per completed invoice
All agent spend — model tokens, platform credits and allocated platform fees, labeled as allocated — divided by invoices the agent completed.
M02Straight-through rate
Share of invoices completed with no human takeover, escalation or failed run.
M03Takeover rate
Share of runs a person had to finish, by agent, version and invoice type.
M04Approval compliance
Share of invoices above the approval threshold that were parked and approved before posting.
M05Duplicate-payment exposure
Value of invoices posted more than once for the same vendor and supplier invoice number, and whether they were paid.
M06Evidence coverage
Share of agent-claimed actions confirmed by an independent record in the ERP, workbook, mailbox or approvals log.

Every metric is defined, versioned and traceable to its source records. Allocated platform fees are labeled as allocated.

Who it's for

Built for the people who own AP outcomes and controls

CFO

“Is our agent spend producing finished invoices — and are payment controls holding?”

Cost per completed invoice next to approval compliance and duplicate-payment exposure.
Head of Shared Services / AP

“Which agent should handle which invoices, and where does my team step in?”

Straight-through and takeover rates by agent, version and invoice type, with exceptions routed to owners.
Internal Audit

“Can we evidence what agents posted, changed and paid, and who approved it?”

Run-level evidence chains with link method and confidence, plus an audit trail of every review decision.
Chief AI Officer

“Which finance agents earn their cost, across vendors?”

One scorecard for in-house and vendor agents, with routing and renewal decisions backed by cost per outcome.

How a pilot works here

Our flagship pilot

Accounts-payable invoice processing is the workflow we know best. In six weeks we connect read-only to one agent runtime plus your ERP, AP mailbox and approvals, baseline cost per completed invoice, and draft the mandates your AP policy already implies.

Pilot scope · Finance operationsSix weeks · read-only
Workflow
AP invoice processing
Agent runtime
One agent runtime or vendor agent platform
Systems of record
ERP · AP mailbox · approvals
Access
Read-only. No changes to your agents.
Success criteria
e.g. a target reduction in cost per completed invoice at equal output, or a defined set of payment-control findings
  1. Week 0
    Connect
  2. Week 2
    Baseline
  3. Week 4
    Mandates
  4. Week 6
    Decide
Which ERPs and agent platforms can ThetaLayer read from?
ThetaLayer is built to read ERP change documents and payment runs (for example from SAP or Oracle), mailbox and workbook audit logs from Microsoft 365 or Google Workspace, approvals, and agent traces from in-house runtimes or vendor platforms such as Microsoft Copilot Studio. Connector availability for your stack is confirmed during pilot scoping.
Do invoice numbers, IBANs or email contents leave our network?
No. The collector runs inside your network. Supplier invoice numbers, IBANs and message IDs are hashed before export, and email bodies, subjects and cell values are dropped. The control plane can still link records — and spot the same IBAN on two vendors — by comparing hashes.
Does ThetaLayer block postings or payments?
Not at the start. ThetaLayer begins read-only: it detects and flags rather than enforces. When you're ready, enforcement can be added at tool boundaries you choose — for example, blocking agent postings above the approval threshold.
How is cost per completed invoice calculated?
Agent spend for the period — model tokens priced per model, vendor platform credits, and platform fees allocated and labeled as allocated — divided by the invoices the agent completed. Every figure links back to the runs and records behind it.

Book a Demo

See an AP workflow reconstructed end to end.

Book a 30-minute walkthrough of the invoice demo — duplicate postings, approval bypasses and cost per completed invoice — and scope a six-week pilot on your own AP agents.