Document intake
Turn source documents into proposed structured records, then review missing values and ambiguous evidence.
Explore this workflow ↗ Documents. Decisions. Done with care.
AgenticBack Office
Turn repetitive administrative work into structured, reviewable workflows. Give agents defined tools, keep exceptions visible, and put people at the decisions that matter.
Explore the workflow ↓Discuss your process
Documents → structured workExceptions → human reviewDecisions → evidence
Research & technical references
Inspect the claim ledger ↗Source logos identify cited publications. They do not indicate endorsement, partnership, certification, or product validation. τ-bench is a paper hosted on arXiv.
What does it mean in practice?
An agentic back office uses AI to interpret administrative work and choose among defined tools, while preserving system access controls, exception handling, and human responsibility. Useful outputs include structured work items, proposed actions, and evidence for review.
Anthropic distinguishes predefined workflows from agents that choose their own processes and tools. That is a useful design distinction: a stable rule can stay a rule, while an agent handles interpretation within an explicit set of operations.[BL-CLM-15]
NIST’s Generative AI Profile supports risk management across the AI lifecycle. For a back office workflow, our proposed implementation turns that principle into concrete checkpoints: source review, access boundaries, exception handling, evaluation, and an accountable owner.[BL-CLM-13]
Follow one piece of work
From invoice to reviewable work.
Choose a stage to follow a fictional invoice. Watch the missing purchase-order reference become an exception that requires a person’s attention.
Interactive walkthrough / fictional data
Invoice workspace · simulation
Invoice / INV-1042
From document to decision.
SupplierExample Supplies
Amount$240.00
Purchase-order reference: missing
Simulated educational flow. No live document processing, sign-in, connector, or mainframe session.
Step 01 / Read document
Open a fictional invoice and preserve its source reference.
The source document remains available so a reviewer can inspect the original evidence.
What comes out
An intake record linked to the document
Start with a process you can evaluate.
Internal requests
Give staff a front door for requests, queues, and approvals, with clear ownership of each transition.
Explore this workflow ↗Legacy workflows
Connect a defined transaction to an easier interface while preserving its business meaning.
Explore this workflow ↗One document. A reviewable workflow.
Turn an invoice into structured work.
Select a field to follow it from the source document into a structured record. A missing purchase order creates an exception. Resolve it in the demo review step before preparing the handoff.
01 / Example invoice
Example Supplies / Fictional document
INVOICE
The original invoice stays unchanged. A review decision is attached to the work record.
02 / Structured work item
Invoice number → reference
A stable reference connects the source document to the work item.
{
"reference": "INV-1042",
"supplier": "Example Supplies",
"amount": 240,
"currency": "USD",
"purchaseOrder": null,
"status": "needs_review"
}This record is assembled from fixed example values. The demonstration does not extract or validate a real invoice.
03 / Resolve the exception
Keep the decision with a person.
04 / Prepare the handoff
Handoff blocked: the purchase-order exception needs review.
Interactive illustration with fictional records. All changes stay in this preview; no document service, purchasing system, or payment API is contacted.
Questions worth asking.
What is an agentic back office?
An agentic back office uses AI to interpret administrative work and choose among defined tools, while preserving system access controls, exception handling, and human responsibility. Useful outputs include structured work items, proposed actions, and evidence for review.
How does it differ from conventional automation?
Conventional workflows follow predefined paths. Agents can choose their next step based on the task and intermediate results. Use fixed rules where they suffice; introduce agent decisions only where interpretation or variation justifies the added complexity.
Does an agent approve payments or sensitive decisions?
Authority comes from the organization’s approved process. A workflow can prepare information and request approval; it should not invent authority. This demonstration never makes a payment or changes a production record.
Do these studies measure Papyrus performance?
No. The cited research and guidance explain design principles and evaluation approaches. Papyrus performance must be established against the customer’s representative tasks and acceptance criteria.
Evidence, with boundaries
Read the sources. Inspect the claims.
Source review: October 9, 2026. Research findings, government guidance, and vendor documentation are identified separately.
Generative AI Profile · AI 600-1
Published / updated 2024-07-26
NIST AI 600-1 is a cross-sectoral generative AI profile supporting risk management throughout the AI lifecycle.
Voluntary risk-management guidance. It is not a certification or a finding about Papyrus.
Research benchmarkτ-bench · Yao, Shinn, Razavi & Narasimhan
Published / updated 2024-06-17
τ-bench evaluates tool-using agents against goal database states and tests consistency across repeated trials under domain policies.
The 2024 benchmark uses simulated users and domain environments. Its results are not Papyrus results or a measure of current models.
Vendor engineering guidanceBuilding Effective Agents
Published / updated 2024-12-19
Anthropic distinguishes workflows with predefined code paths from agents that dynamically direct their processes and tool use.
Engineering guidance published in 2024. The article notes that the tooling landscape has since changed.
Bring one real back office workflow.
Start with the documents, systems, exceptions, and approval requirements. Decide what to automate after you understand how the work succeeds.