AI Automation · Agents
Autonomous AI Agents
Named agents own a narrow job each: read the document, check it against policy, decide, act, and escalate the rest to a person.
What it is
Sixteen agents in three teams
Each one owns a narrow job and reports into the workflow. An agent here is not a chatbot with system access; it is a worker with a defined input, a defined output and a defined limit on what it may do alone.
Document
Invoice Agent
Reads the header, the line table and the tax block, and checks the arithmetic before anything moves.
Document
Receipt Agent
Turns photographed, folded and faded expense receipts into claim lines.
Document
Bank Statement Agent
Parses statements from any bank and proves the running balance row by row.
Document
Purchase Order Agent
Reads orders and shapes them for the match that follows.
Document
KYC & Identity Agent
Extracts identity fields, cross-checks them and redacts on export.
Document
Custom Document Agent
Built for a document type of yours: delivery notes, contracts, claims, certificates.
Finance
Accounting Agent
Codes each line to a ledger account and cost centre, with the evidence attached.
Finance
AP Agent
Runs the payables cycle from capture to posting and hands the exceptions to a person.
Finance
Matching Agent
Two- and three-way match at line level, across units of measure and part codes; six-way where the data exists.
Finance
Vendor Agent
Onboards suppliers, verifies them at source and scores the risk.
Finance
Reconciliation Agent
Matches payments to bills and explains what is left over, item by item.
Finance
Compliance Agent
Checks the tax treatment per line and isolates credit you cannot claim.
Workflow
Validation Agent
Applies your business rules before a document is allowed to move.
Workflow
Approval Agent
Routes by value band to the approver the policy names, with delegation.
Workflow
Posting Agent
Writes a clean entry into the system of record and links back to the source.
Workflow
Exception Agent
Escalates what needs a person, with the reasoning attached.
Where it applies
Where an agent earns its keep
The test is simple. If the work arrives in volume, follows a policy that exists on paper, and stalls when a person is busy, an agent will take it. If the work is a judgement call, it stays with the person who is accountable for it.
Payables
Invoice intake
Invoice, Matching and Approval agents in a chain. The person sees the mismatches, not the invoices that matched.
Treasury
Bank reconciliation
Statement and Reconciliation agents run daily, so month end inherits a short list of named breaks instead of a spreadsheet.
Procurement
Vendor onboarding
Vendor and KYC agents verify GSTIN, PAN and bank details at source before the first invoice is accepted.
Tax
GST through the month
The Compliance Agent checks HSN rates and GSTIN status per line as invoices arrive, so reconciliation to 2B is a check rather than a hunt.
How PTAS AI does it
What each one adds
The handovers are the interesting part. Every agent picks up what the last one established instead of starting from the document again, and the item carries its whole history with it.
One invoice, five agents, no handover meeting · Illustrative example
Narrow jobs are deliberate. An agent that does one thing can be tested, measured and replaced; one that does everything cannot.
Narrow jobs
An agent that does one thing can be measured.
Explicit handovers
Each agent carries forward what the last established.
Escalation built in
What should not be decided alone, is not.
How they group
Three teams, one chain
Document agents live in DocPro, finance agents in FinSuite, and the workflow agents belong to the platform both run on.
Controls
What they are not allowed to do
An agent that is wrong does not waste a minute; it posts an entry. So the limits are enforced by configuration, and a change to a limit is itself recorded.
Confidence thresholds
The score above which an agent may act alone is yours to set, per agent and per workflow.
Mandatory human gates
Value bands, document types or counterparties that always need a person, enforced rather than suggested.
Reasoning on every action
What the agent saw and which rule it applied, stored with the outcome and queryable later.
Narrow scope by design
An agent owns one job, so a failure is contained to that job and fixable without touching the rest.
The products
Where the agents live
This page is the roster. The product pages show what each agent returns, with illustrative output, and the solution pages show what a chain of them does to a process.
DocPro
The six document agents, with illustrative output and field-level confidence.
Read more →FinSuite
The finance and workflow agents, module by module.
Read more →Custom Document Agents
How a document type of yours gets an agent of its own.
Read more →From Document to Finance Action
How the document and finance agents hand over on one invoice.
Read more →FAQ
Questions Teams Ask
Why so many agents instead of one that does everything?
Because a narrow agent can be tested, measured and replaced, and a broad one cannot. When an invoice posts wrongly you want to know which step failed and fix that step. One agent covering the whole cycle turns every failure into an investigation of the whole cycle.
Do we have to use all of them?
No. Most teams start with two or three, usually Invoice, Matching and Approval, and add others as the workflow extends. Each agent works on its own, and the chain lengthens as you go.
Can we add an agent for work specific to us?
Yes. Custom agents are part of standard onboarding. You define the steps, the systems and the guardrails in business terms, and the agent is built against those. The runtime, the controls and the audit trail are the same as for the standard set.
What stops an agent acting outside its remit?
Its scope is configuration, not a hope about behaviour. An agent has defined inputs, defined systems it may write to, and a defined confidence and value band inside which it may act alone. Anything outside those escalates instead of proceeding.
How does an agent know when to ask a person?
Two ways, and both are configuration. A field below its confidence threshold sends the document to review with that field named. A rule that fails, such as a duplicate, an unmatched line or a vendor outside policy, sends it with the rule named. Review is a control the agent uses, not a sign that it failed.
Keep reading
Related Capabilities
AI Automation
The section this page belongs to, and the other five capabilities.
Read more →Agentic AI Platform
The runtime these agents run on, and how it is governed.
Read more →Agentic AI Architecture
The technology underneath, for a CIO who wants the mechanism.
Read more →Accounts Payable Automation
The first chain most teams run, described as a solution.
Read more →RAG and Knowledge Systems
How agents consult policies, master data and history before they decide.
Read more →Next step
Start with two or three
Invoice, matching and approval is the usual first chain. The rest extend from it once that one holds on your own documents.