Resources · Glossary

AI Automation Glossary

The terms that come up when evaluating AI automation, defined in plain language and linked to where they matter.

Close-up of a printed dictionary page open at the entry for the word artificial

How to use this page

The same words, meaning the same thing

An evaluation goes faster when finance, operations and IT agree on what a confidence score or a three-way match actually is.

These definitions are written for the people who sign off on automation, not for the people who build it. Where a term is specific to Indian tax or payments, such as GSTIN, GSTR-2B, IRN, TDS or UTR, the definition says what it is and why it matters in payables. Tax definitions are practical descriptions, not legal advice; rates and thresholds change, so check the current rules with your tax adviser.

Most terms link to the page where the idea is put to work. The figure below follows one invoice and names the terms it meets on the way.

One invoice, in glossary terms

  1. 01 Classification Settled as a tax invoice before any field is read
  2. 02 Confidence score One per field: the GSTIN scores high, a handwritten quantity low
  3. 03 Three-way match Invoice, purchase order and GRN agree within tolerance
  4. 04 Approval band The value decides who signs, with segregation of duties enforced
  5. 05 Audit trail What was read, which rule applied, who approved, what posted

Each step can stop the invoice and send it to a person with the reason. That is the human-in-the-loop design, and it is why the terms come in this order.

A to W

58 terms, defined

Alphabetical. Abbreviations are filed under the letters most people say out loud.

A

Accounts payable (AP) automation
Software that takes a supplier invoice from arrival to payment with fewer hands: capture, extraction, matching to the purchase order and goods receipt, approval routing, posting to the ERP and the payment run. People keep the exceptions and the approvals above their thresholds. Accounts Payable Automation →
Agentic AI
AI that carries a task of several steps from start to finish. An agentic system reads its inputs, decides the next step against rules and policy, acts through the systems it has been given access to, and checks the result. In business use, what it may do alone and what it must hand to a person are set in advance. Agentic AI →
AI agent
A piece of software given one bounded job, such as reading invoices or verifying a vendor, with the data, tools and permissions that job needs and no more. Several agents can pass work along one workflow, each leaving a record of what it did. AI Agents →
AI orchestration
Coordinating several models, agents and tools so they run in the right order, pass work between them and recover when one fails. Orchestration decides which agent handles which step. It does not do the reading or the deciding itself. AI Orchestration →
Approval band
A value range in an approval policy that decides who must sign off. An invoice of ₹40,000 might need a team lead and one of ₹4,00,000 a finance controller. Bands are often split further by cost centre, category or legal entity, and carry delegation rules for when the approver is away. Business Rules →
Audit trail
The record of what happened to an item and why: what was read, which rule applied, who approved, what was posted and when. A useful one can be queried a year later and answers an auditor without anyone rebuilding events from email. Security →

B

Bank reconciliation
Matching each line on a bank statement to the payment, receipt or ledger entry it belongs to, and explaining every difference. Whatever stays unmatched is a break, listed with a likely cause and an age so someone can clear it. Bank Reconciliation Automation →
Business rules
Written conditions that decide what happens to an item: above this value, route to that approver; more than the tolerance away from the PO price, hold. In an automated workflow, rules are configuration with an owner, versioned so you can tell which rule any past decision followed. Business Rules →

C

CGST, SGST and IGST
The components of GST on most supplies in India. A supply within one state carries Central GST and State GST in equal halves (Union Territory GST in place of State GST in some union territories). A supply between states carries Integrated GST. The place of supply decides which, and an invoice that charges the wrong one has to be corrected by the supplier. FinSuite GST Compliance →
Classification
Deciding what a document is before reading it: invoice, credit note, receipt, statement, purchase order or identity document. It comes first because the type decides which fields to look for and which checks apply. A credit note read as an invoice can pass every arithmetic check and still be wrong. Document AI →
Computer vision
The branch of AI that interprets images. In document work it handles what comes before reading: finding the page in a phone photo, correcting skew and glare, and locating tables, stamps and signatures. Computer Vision →
Confidence score
A model's estimate, between 0 and 1, of how likely one extracted value is to be right. It is scored for each field separately, so a smudged GSTIN does not cast doubt on a clean total. A score only means something against a threshold. DocPro →
Confidence threshold
The score below which a field stops and goes to a person. It is set per field and per document type, so a grand total can be held to a stricter bar than a remarks field. Raising a threshold sends more to review; it does not change what was extracted. Human-in-the-Loop →
Cycle time
The time an item takes from arrival to done, for example from an invoice landing in the mailbox to its posting or payment. Measured by stage, it shows where work waits, which is usually in someone's inbox. AP automation evaluation guide →

D

Deployment model
Where the software runs and who controls the infrastructure. PTAS AI offers cloud in a region you choose, private cloud in your own account, and on-premise or air-gapped deployment where supported. The products and controls are the same in each. Security →
Document AI
AI applied to business documents: classifying them, reading their text and layout, extracting fields and tables, and judging how sure it is of each value. Intelligent document processing is the same idea with validation and workflow built around it. Document AI →

E

E-invoicing (GST)
The Indian requirement for businesses above a notified turnover threshold to register their B2B invoices on a government Invoice Registration Portal. The portal validates each invoice, returns an IRN and a signed QR code, and passes the details on to the GST system. From an in-scope supplier, a B2B invoice without an IRN is not a valid tax invoice. GST Compliance Automation →
ERP (enterprise resource planning)
The system of record for finance and operations: general ledger, payables, receivables, purchasing, inventory and often payroll. SAP, Oracle, Microsoft Dynamics and NetSuite are ERPs. Tally, QuickBooks and Zoho fill the same role in many smaller businesses. Integrations →
Exception
An item a rule cannot settle on its own: a price outside tolerance, a missing goods receipt, a possible duplicate, a field below threshold. It goes to a queue with the reason attached. The exception rate and its causes are the most useful numbers to have before automating a process. Human-in-the-loop guide →

G

Generative AI
AI models that produce new text, images or code from a prompt. In finance and document work they help with reading unstructured text and drafting, and they need controls: a generated value must trace back to its source, or be left blank and flagged. Generative AI →
GRN (goods receipt note)
The record that goods arrived: what, how many and when, raised by the receiving team. In a three-way match the GRN is the evidence that the goods you are paying for actually arrived. For services, a service entry sheet plays the same part. Procure-to-Pay →
GST (Goods and Services Tax)
The indirect tax India has charged on the supply of goods and services since 1 July 2017. It is levied at each stage of supply, and a registered business can usually set the GST it paid on purchases against the GST it charges on sales. GST Compliance Automation →
GSTIN
The 15-character GST identification number a business holds for each state it is registered in. It is built from a two-digit state code, the business's ten-character PAN, an entity number, the letter Z by default and a check digit. Its format, status and match to the vendor master are among the first checks on a purchase invoice. FinSuite Vendor Management →
GSTR-2B
A monthly statement the GST portal generates for each registered business, listing the input tax credit available from what its suppliers reported. Unlike GSTR-2A, it does not change once generated for a period. Reconciling the purchase register to it before filing shows which invoices can be claimed and which suppliers have not filed. GST automation guide →

H

Hallucination
When a generative model produces a value or a statement its source does not support. In document work the control is easy to state: every extracted value must trace to the page, and a value the agent cannot read is left blank and flagged rather than filled in. AI Security and Governance →
HSN and SAC codes
Classification codes on GST invoices. HSN (Harmonised System of Nomenclature) codes classify goods and SAC (Services Accounting Codes) classify services. The code determines the GST rate, and the number of digits that must be shown depends on the supplier's turnover. Checking each line's code against the rate charged catches a common class of tax error. FinSuite GST Compliance →
Human-in-the-loop (HITL)
A design in which people review what an automated system is unsure about, or what policy says a person must decide, before it acts. Done well, the reviewer sees only the field in doubt and the reason, the correction is logged, and the review rate becomes something you tune. Human-in-the-loop guide →

I

IDP (intelligent document processing)
Software that turns unstructured documents into validated, structured data. It classifies the document, extracts fields and tables, scores its confidence in each, checks the values against rules and reference data, and sends what it cannot settle to a person. OCR, by contrast, stops at the characters. Intelligent Document Processing →
Input tax credit (ITC)
The GST a business paid on its purchases, which it can set against the GST it owes on its sales. To claim it, the business generally needs a valid tax invoice, the goods or services must have been received, the invoice must appear in GSTR-2B because the supplier reported it, and the tax must have been paid to the government. Some purchases are blocked from credit altogether. GST automation guide →
Integration method
How one system exchanges data with another. PTAS AI names four: Native, where a supported connector exists for your edition; API, calling the interfaces the system exposes; File, structured files on a schedule; and Custom, orchestration across systems or a bespoke build. The method for each system is confirmed in discovery. Integrations →
IRN (Invoice Reference Number)
The unique 64-character reference the Invoice Registration Portal returns when an e-invoice is registered, with a signed QR code. The QR carries the key invoice details, so comparing it with the printed invoice shows whether the PDF you received matches what was reported. GST Compliance Automation →

K

KYC (know your customer)
Verifying the identity of a customer, vendor or other counterparty before doing business with them, and keeping the evidence. In India it commonly involves a PAN card, passport, voter ID or masked Aadhaar, read, cross-checked against what the applicant entered, and checked for expiry. KYC & Identity Agent →

L

LLM (large language model)
A generative model trained on large amounts of text to read and produce language. LLMs can follow instructions and make sense of unstructured text, which is useful inside automation, but on their own they do not check an answer against your records. The validation and review around them do that. Generative AI →

M

Machine learning
Methods that learn patterns from examples. Classification, extraction and confidence scoring rely on machine learning; business rules are written by people, and a sound automation uses each for what it does well. Machine Learning →

N

Non-PO invoice
An invoice with no purchase order behind it: utilities, rent, subscriptions, professional fees. With no PO or goods receipt to compare, it is checked on vendor, tax, duplicates and coding rules, and its approval follows the approval band. FinSuite AP Automation →

O

OCR (optical character recognition)
Software that converts an image of text into machine-readable characters, in reading order. It tells you which characters are on the page; it does not know which of them is the invoice total. Intelligent document processing builds on it. How to evaluate IDP →

P

PAN (Permanent Account Number)
The ten-character alphanumeric identifier the Income Tax Department issues to individuals and entities. It is embedded in every GSTIN a business holds, so a vendor's PAN and GSTIN can be checked against each other.
Place of supply
The GST rule that decides where a supply is treated as made, and so whether it carries CGST and SGST (same state) or IGST (different states). For goods it usually follows the delivery location; services have their own rules. Checking it against the tax type charged is a routine line-level test.
Procure-to-pay (P2P)
The whole purchasing chain: requisition, approval, purchase order, goods receipt, invoice, matching, payment and the record of each. Automated well, each step carries the reference of the one before, so an invoice can be traced back to the order that justified it. Procure-to-Pay →
Purchase order (PO)
The buyer's formal order to a supplier, stating items, quantities, prices and terms. It is the commitment an invoice is checked against: a two-way match compares the invoice with the PO, and a three-way match adds the goods receipt. Purchase Order Agent →

R

RAG (retrieval-augmented generation)
A way of making a generative model answer from your own material. Before it responds, relevant passages are retrieved from a controlled source such as policies, contracts or past cases, and the answer is generated from them, ideally citing the source. It reduces the need to check the answer. It does not remove it. RAG →
Redaction (field-level)
Masking specific fields, such as an identity number or a bank account, when a record leaves the platform, set per field and per destination. A KYC reviewer and a billing system can then receive different versions of the same record. DocPro Security →
REST API
A common way for systems to talk over the web: one system sends a request to a documented address and gets structured data back, usually JSON. DocPro accepts documents and returns records this way. REST API →
RPA (robotic process automation)
Software bots that repeat the clicks and keystrokes a person would make on a screen. RPA works on stable, rule-bound screens and struggles when the screen or the document changes, because it follows a script and does not read the page. Agentic AI vs RPA →

S

Segregation of duties (SoD)
A control that splits a process so no single person can complete a risky transaction alone. In payables, whoever creates or changes a vendor cannot approve that vendor's invoices, and whoever approves cannot release the payment. FinSuite →
Six-way match
An extension of the three-way match, run where the data exists: the invoice is checked against the vendor's identity, the purchase order, the goods receipt, the quantities and pricing, the contract terms, and the tax, bank and duplicate checks. A check with nothing on file to compare against is skipped, and the skip is recorded. FinSuite →
Straight-through processing (STP)
An item that goes from arrival to completion without a person touching it, because every check passed and no policy required a sign-off. The share of items that go straight through is the other side of the touch rate. DocPro + FinSuite →

T

TDS (tax deducted at source)
Income tax the payer deducts when making certain payments, such as contractor charges, professional fees or rent, and deposits with the government on the payee's behalf. The rate depends on the provision of income tax law that applies, so a payables workflow often has to work out TDS before it releases a payment.
Three-way match
Checking an invoice against its purchase order and goods receipt before paying it: the right items, at the agreed price, in the quantity actually received. A two-way match leaves out the receipt. Units of measure and pack sizes are where most manual matching stalls. A six-way match adds the vendor, contract, tax, bank and duplicate checks. FinSuite AP Automation →
Tolerance
The difference a rule allows before a check fails, such as a small percentage on price or a unit on quantity. Set well, tolerances let rounding differences through and stop real overcharges. Business Rules →
Touch rate
The share of items that needed a person at any step, best split by the reason: low confidence, a failed check, or a policy that requires approval. It is a more honest measure of automation than the share of documents processed. Human-in-the-loop guide →

U

UPI (Unified Payments Interface)
The real-time payment system run by the National Payments Corporation of India and used from phone apps for everything from reimbursements to supplier payments. A UPI receipt carries the amount, payer, payee, a transaction ID and a UTR: the fields that tie it to a claim or a bank line. Receipt Agent →
UTR (Unique Transaction Reference)
The reference the banking system assigns to a transfer such as NEFT, RTGS, IMPS or UPI. It appears on both the payment confirmation and the bank statement, which makes it the cleanest key for matching one to the other. Bank Statement Agent →

V

Validation
Checking an extracted value before anyone relies on it: recomputing totals, checking a date's order or an identifier's format, matching a vendor to the master, applying your rules. A field can be read with high confidence and still fail validation, and it is held when it does. Document Data Extraction →
Vendor master
The reference record for every supplier: legal name, GSTIN, PAN, bank account, payment terms and contacts. Duplicates and stale bank details here undermine matching, tax checks and payment controls, so changes to it, bank details above all, need verification and a sign-off. Vendor Management Automation →

W

Webhook
A message one system sends to another the moment something happens, instead of waiting to be asked. DocPro can send one when a document finishes processing, so your system receives the record without polling for it. Webhooks →
Workflow automation
Automating the steps between people, teams and systems: routing, approvals, reminders, escalation and status updates, with a clock on each step. It moves the work along. Document AI and business rules decide what the work is. Workflow Automation →
Workflow orchestration
The layer that runs a multi-step process across systems and agents, keeps its state, retries what failed and knows what is waiting on whom. It is what lets a long process stop, wait for a person and resume, with a record of every step. Workflow Orchestration →

FAQ

Terms people often confuse

What is the difference between OCR and intelligent document processing?

OCR turns an image into characters and stops there. Intelligent document processing classifies the document, finds the fields that matter, scores its confidence in each, checks the values against rules and reference data, and routes what it cannot settle to a person. What comes out is validated data.

What is the difference between agentic AI and RPA?

RPA replays a fixed script of clicks and keystrokes, so it breaks when the screen or the document changes. An agentic system reads the input in context, applies rules, acts through APIs and checks the result, and hands anything it cannot settle to a person with the reason attached.

What is the difference between GSTR-2A and GSTR-2B?

Both list the input tax credit suppliers have reported for you. GSTR-2A updates continuously as suppliers file. GSTR-2B is generated once for each period and does not change afterwards, which makes it the statement to reconcile the purchase register against before filing.

What is the difference between two-way and three-way matching?

A two-way match compares the invoice with the purchase order: items, quantities and prices. A three-way match adds the goods receipt, so you only pay for what was actually received. Invoices without a purchase order are checked on vendor, tax, duplicates and coding rules instead.

What is the difference between three-way and six-way matching?

A three-way match checks the invoice against the purchase order and the goods receipt: the right items, at the agreed price, in the quantity received. A six-way match adds three more sources: the vendor's identity against the master, the contract terms, and the tax, bank and duplicate checks. Three-way is the standard; six-way also catches the payment to a changed bank account or the invoice submitted twice, which a three-way match lets through.

From terms to your process

Find out which of these your process needs first

The assessment asks fourteen questions about how work moves today and names the two places we would start. It takes about seven minutes.