Ingest Agent
Email inbox listener, SFTP and shared folders, PDFs, Excel and CSV, images. It watches the channels rather than waiting to be fed.
Across every stage of the platform
Agents do the work.Rules do the posting.
DataTwin is a multi-agent system: specialised agents with exclusive duties, cognitive interpretation of policies and contracts, humans at the control gates, and deterministic rules wherever something becomes an entry in your books. Here is exactly where each one operates.
- 01
Multi-agent framework
Specialised agents with exclusive duties. Plug-and-play for new rules and geographies.
- 02
Cognitive + logical
Interprets policies and contracts. Real-time decisions and escalations.
- 03
Human-in-the-loop
Intervention only at control gates, with an audit-ready accountability trail.
- 04
Elastic scalability
Compute and agents auto-scale for month-end and quarter-end peaks, so a first pass returns a number in about two minutes.
- 05
Continuous learning
Feedback loops from exceptions and audits. RAG keeps rules current.
The architecture
Not one model. A team of specialists.
A single general-purpose model asked to do everything is neither accurate nor auditable. DataTwin runs specialised agents, each with a narrow remit and a defined output, coordinated by an orchestration layer that decides what runs, in what order, and when a human is needed.
Orchestration Agent
Sits above everything else. Sequences the specialists, classifies what each finding is and how severe it is, routes work to the right agent or the right person, and holds the accountability trail that ties every agent action and human decision together.
01
Intake & understanding
Getting data in, whatever shape it arrives in, and working out what it actually is.
Data Hygiene Agent
Duplicate checks by reference and checksum, multi-page document structure, multilingual documents, handwritten pages, document classification.
Registry Agent
Determines the vendor, item, tax, cost centre and company each record belongs to: the identity resolution that everything downstream depends on.
02
Judgement & control
Interpreting the rules that govern a transaction, and spotting what should not be there.
Governance Agent
Interprets company policy, location-specific policy, accounting policies and principles, reporting standards, and the agreements and contracts that set the terms.
Risk Intelligence Agent
Pattern recognition, fraud scenario recognition, segregation-of-duties reasoning and contextual risk scoring, running continuously rather than at audit time.
WF & Compliance Agent
Tax jurisdiction rules, reverse charge, exemption and certificate handling, HSN/SAC classification, withholding, cross-border and taxability mapping.
03
Resolution & reporting
Reconciling, explaining, correcting and telling people what happened.
Recon Agent
Fusion across vendor, PO, GRN and bill, and beyond AP into timesheets, expenses and gateway settlements. N-way, not two-way.
Investigation Agent (RCA)
Deep-dives a flagged anomaly, finds the root cause, whether fraud, error or process gap, collects the evidence and recommends the corrective action.
Accounting Agent
Expense and payment accounting, accrual handling, cost and management accounting. Produces the correct treatment, for a human to release.
Remediation Agent
Fixes data gaps, resubmits, and closes the loop so the same exception does not return next cycle.
KPI & Communication Agent
Touchless rate, exception rate, cycle time, SLA monitoring, exception heatmaps by vendor, and proactive alerts on breaches and holds.
Underneath the agents
Three engines the agents call on
Reading what arrives
DataTwin OCR Models
Purpose-trained for finance documents rather than general text. Invoices in any layout, multi-page and multi-language documents, handwritten annotations, payment screenshots and gateway reports, at line level, not just header.
Knowing the rules that apply
DataTwin RAG Models
Retrieval over your own contracts, price agreements, tax positions and policies, so an agent reasons against your terms as they stood on the relevant date, not a generic interpretation. This is how rules stay current as agreements change.
Autonomous Data Engineering & Recon Engine
AUDERE
The part that models a source it has never seen, builds the schema, applies the SCDP transformations and runs the reconciliation. It is why a new entity or distributor format does not need a mapping project first.
The vocabulary here, from SCDP and N-way to RAG and full population, is all defined in plain language.
In practice
Four places the intelligence does the work
Concretely, on the platform, today, not on a roadmap. These are the same four capabilities summarised on the Platform overview, described here with the engines behind them.
It models new sources for itself
A distributor with their own claim format, a bank whose statement layout matches nobody else’s, an entity from an acquisition. AUDERE reads it, infers what each field means and builds the acquisition model. No mapping spreadsheet.
It builds the schema from your requirement
Describe what needs reconciling and against what. The engine elicits the detail, then constructs the schema, rules and checks included, for entities it has never handled. SCDP makes that possible: transformation is declared, not coded.
It reasons through the hard reconciliations
Identifiers that disagree across systems. A fact that only exists inside a clause. An answer that depends on context spread across records. A rules engine stops there; reasoning resolves those faster than a manual queue.
It reads the documents that hold the terms
The governing fact in finance is usually in a document: a clause, a certificate, a roster, a remittance. OCR and RAG let an agent cite the term that governed a transaction, which is what a counterparty will accept.
Easier to judge on your own data than from a description.
Closed-loop assurance
Stop it at source. Investigate what got through.
Two directions running at once, one preventing and one correcting, each feeding the other.
Preventive
Applies policy before anything posts
- Rules and policies applied before the entry is created
- Errors stopped at source: two-way and three-way match, approval limits
- Continuous preventive checks on master data, thresholds and cut-offs
- Eligibility and tax treatment determined at bill creation, not at close
Agents: Governance, Risk Intelligence, WF & Compliance
Corrective
Deep-dives whatever still gets flagged
- Anomaly investigated after it is raised, not filed for later
- Root cause identified: fraud, error or process gap
- Evidence collected and a corrective action recommended
- The fix feeds back as a new preventive rule, so it does not recur
Agents: Investigation, Recon, Remediation
Exception to remediation, end to end
- 01
Detect
Governance, Risk and Compliance agents flag the anomaly as it appears.
- 02
Classify
Orchestration assigns a type and a severity, so response matches the risk.
- 03
Route
Sent to remediation, or to a named human where judgement is required.
- 04
Remediate
Data gaps fixed, corrections prepared, the item resubmitted.
- 05
Close & learn
Resolved, and the outcome feeds back to improve the next cycle.
Which of your exceptions never get closed? That is usually where we start.
Human-in-the-loop
People at the control gates, not in the queue
AI handles the volume. Humans are involved where judgement, authority or accountability genuinely requires a person, and every one of those decisions is logged alongside the agent actions around it.
Providing missing inputs
A human enriches the packet where something simply is not in the data: a missing PO number, a tax certificate, vendor KYC.
Correcting the AI’s course
Reviewers accept, adjust or reject agent recommendations, such as an exception waiver or a policy override, and the correction is captured.
Approver guidance
The agent recommends approve or reject with its reasoning attached. The person still decides, but not from a blank screen.
Exception arbitration
Where two agents reach different conclusions, a human resolves the conflict rather than the system silently picking one.
Escalation handling
Complex or high-value cases route to Controllers for sign-off, with thresholds you set rather than ones we assume.
Policy calibration
Reviewer corrections feed back into the Governance and Compliance agents, so the same judgement does not have to be made twice.
How other finance teams have set their control gates.
The line we do not cross
Nothing posts to your booksbecause a model was confident.
Agent reasoning
Evidence
Versioned rules
Your books
Every value that becomes an entry is produced or gated by explicit, inspectable, versioned rules. You can read the rule, see which version produced a finding, and trace it to the source record and the document behind it. Where an agent contributed reasoning, that reasoning is recorded as evidence supporting the decision, never as the decision itself.
This is a deliberate architectural boundary, not a limitation we are working to remove. A finance platform that cannot explain why a number is what it is has not automated your close; it has made your close harder to defend.
Every finding we return can be traced to the rule and the document behind it. Test that on your data.
What changes
Traditional operations vs agentic operations
Error handling
Traditional
Manual detection, high rework
With DataTwin agents
Preventive governance plus auto-remediation minimises errors before they post
Compliance
Traditional
Policy checks inconsistent, heavy audit dependency
With DataTwin agents
Continuous cognitive validation with audit-ready evidence attached
Cycle time
Traditional
Slow approvals from exceptions and rework
With DataTwin agents
Faster invoice-to-pay, with human-in-the-loop only where it is needed
Risk management
Traditional
Fraud and risk spotted late, reactively
With DataTwin agents
Risk Intelligence Agent monitors patterns in real time
Workload
Traditional
High manual effort across the team
With DataTwin agents
Agents absorb the routine work; people move to judgement
Transparency
Traditional
Limited visibility, delayed reporting
With DataTwin agents
KPI dashboards with proactive communication on breaches and holds
Audit
Traditional
Heavy, after-the-fact correction
With DataTwin agents
Ongoing assurance in near real time, evidence assembled as work is done
Scale
Traditional
Volume spikes overwhelm systems and teams
With DataTwin agents
Compute and agents auto-scale; new geographies plug in as new agents
The fastest way to judge any of this is to point it at your own data.