Skip to main content

AI-02 AI automation · Worldwide

AI Business Automation

Somewhere in your business a person spends the day reading documents and typing what they say into a system. That is the work a model does well, provided a person still checks the cases it is unsure about.

  • For input with no fixed shape: scans, emails, messages, agreements
  • A measured pilot before the build, on your own documents
  • Confidence threshold set per field, not one number for all
  • A review queue with a named owner and a service level

Reply within 1 working day · pilot before build, always

P J DivyaCo-Founder, CTO and CMO

Builds the automation, the review step and the day-it-fails path, then measures it.

invoice-run.loglive

ingest --source ap-inbox --count 412

412 documents, 287 distinct layouts

extract --fields vendor,gstin,invno,date,taxable,tax,total

amount 99.2% | gstin 97.4% | vendor 71.8%

threshold --vendor always-review --amount 0.95

338 straight through, 74 to review queue

post --erp --audit-trail on

nothing posted below its own threshold

  1. InputTask inventory, hours counted
  2. AccessData scope written down
  3. BuildModel inside the workflow
  4. OutputSavings you can audit
Invoice entryEmail triageSupport draftsContract termsReconciliation
Pilot set50 to 200 docs
Per use case4 to 8 weeks
AlwaysPilot before build
DesignHuman in the loop

10 Related

What this sits between

The model reads. These decide what happens to what it produces.

  • 01

    Supplier invoice entry

    Three hundred suppliers, three hundred layouts, one set of fields. The highest-volume reading job in most back offices and the clearest case for a model.

    Best payback
  • 02

    Email and ticket triage

    Sorting what arrives into categories, urgency and owner, so a person reads the ones that need them rather than all of them.

    Highest volume
  • 03

    First-draft replies

    A drafted response built from your own knowledge base, which an agent edits and sends. Faster to correct than to compose.

    Assisted
  • 04

    Contract term extraction

    Pulling dates, values, notice periods and renewal terms out of agreements nobody has time to re-read before they auto-renew.

    Risk reducer
  • 05

    Document to record checks

    Comparing what a document says against what the system holds, and surfacing only the differences for a person to resolve.

    Control
  • 06

    Language handling

    Reading and drafting across English and the other languages your documents arrive in, with accuracy measured separately per language rather than assumed.

    Assisted

10 Related

What this sits between

The model reads. These decide what happens to what it produces.

Rules against a model, on the properties that decide it
PropertyRules winA model wins
Input shapeOne template, or a handfulHundreds of layouts, or free text
Failure modeBreaks loudly and predictablyFails quietly and plausibly, so needs review
Running costEffectively zero per itemBilled per unit of text processed
Change handlingA new layout means a new ruleA new layout usually needs nothing
Best forBank statements from three banksInvoices from three hundred suppliers

The combination is usually the answer

A model reads the unstructured document and produces fields; rules then validate them, do the arithmetic and decide what posts. Neither alone is as good as the pair, and asking a model to add up a column is a category error.

10 Related

What this sits between

The model reads. These decide what happens to what it produces.

A worked example on supplier invoice extraction
FieldMeasured accuracyCost of an errorThreshold we set
Invoice amount99.2%A wrong paymentHigh, plus a rules cross-check on the arithmetic
GSTIN97.4%A blocked input tax creditHigh, plus a format and checksum validation
Vendor name71.8%A misposted ledger, easily spottedAlways review, until the accuracy improves
Invoice date98.6%A wrong ageing bucketMedium, with a sanity range check
Line description88%Almost nonePermissive, no review
  • The thresholds are set from the cost of an error, which is a business judgement and yours to make. We supply the measured accuracy and the review load each choice implies.
  • A rules layer sits after the model for anything checkable: format validation, checksum tests, arithmetic that must tie, and range sanity checks. A model should never be the thing that adds up a column.
  • The review queue is a real queue, with a named owner, a service level and escalation, exactly as an approval workflow would have.
  • Where the accepted output posts into the books, the Anything writing into the books inherits the logging control an external audit already tests: an unalterable, timestamped record of each entry and each later change, which is control A.8.15 of ISO/IEC 27001. An automated posting is part of that record.

10 Related

What this sits between

The model reads. These decide what happens to what it produces.

  1. Step 1: Check it is not a rules job

    If every document arrives in one template, rules are cheaper, faster and more reliable. We will send you to rules-based automation where that is the honest answer.

    Week 1
  2. Step 2: Pilot on your own documents

    50 to 200 real items, correct answers defined by your team, accuracy reported per field. The build decision is made here, on a number.

    Weeks 1 to 2
  3. Step 3: Set thresholds per field

    From the cost of an error rather than from the accuracy alone. Expensive fields get conservative thresholds and cheap ones get permissive ones.

    Week 3
  4. Step 4: Wire the queue and the write-back

    Review queue with a named owner and a service level, a rules layer for validation and arithmetic, then the write into your ERP or ticketing system.

    Weeks 3 to 6
  5. Step 5: Monitor for drift

    Accuracy on a rotating sample, review rate, cost per item and failure rate. A rising review share is the earliest warning that something upstream has changed.

    Ongoing

Two people were keying invoices from about three hundred suppliers. They now review the exceptions and handle the difficult vendors, and month-end closes four days earlier.

Accounts payable leadMulti-location retailer

10 Related

What this sits between

The model reads. These decide what happens to what it produces.

  • Personal data. Documents contain names, phone numbers, addresses and bank details, and sending them to a model is processing them, so the General Data Protection Regulation and its equivalents apply. We map the fields, redact what the task does not need, and record where the rest goes.
  • The accounting record. Anything posting entries is part of the accounting system and inherits the logging control an external audit already tests, retained for as long as the record-keeping rules where you are incorporated require.
  • What we put in writing before a pilot. Which provider, which region, whether inputs are used for training, what retention applies, and what never leaves your systems at all.

10 Related

What this sits between

The model reads. These decide what happens to what it produces.

The savings figure in the report is measured, not modelled. The same team produces the before number and the after number.

  • P J Divya

    Co-Founder, CTO and CMO

    Builds the automation and the human review step, and instruments both so the saving is auditable.

  • Chetan Ravi

    Founder and CEO

    Counts the hours before, agrees the task inventory, and signs off the number afterwards.

  • Harith, Compliance Expert at ProTax

    Harith

    Compliance Expert

    Checks the record the automation leaves and whether it satisfies the books it feeds.

  • Harsh Jain, Legal Expert at ProTax

    Harsh Jain

    Legal Expert

    Sets what the model may read, and reviews accountability when an automated step gets it wrong.

10 Related

What this sits between

The model reads. These decide what happens to what it produces.

  • Working hours

    Our day runs on UTC+5:30. The overlap window with your team is written into the scope rather than assumed, and everything outside it runs asynchronously.

  • How we communicate

    One written update a day on the channel you already use, a standing weekly call inside the overlap window, and a named person to escalate to. Nothing important is agreed only on a call.

  • Who you contract with

    ProTax Consultancy, the company behind ProTax, named in the agreement with its registration number. The governing law and the forum are agreed before you sign, not after a dispute.

  • Currency and payment

    Invoiced in your currency or in ours, your choice, and settled by bank transfer. Milestones are tied to deliverables you can see, never to elapsed time.

  • What you own

    Copyright in everything built for you is assigned on final payment: source code, design files, prompts, configuration and documentation. Third-party licences are listed by name so nothing is a surprise later.

  • Where data sits

    You choose the region your data is stored and processed in, and the answer is written down before the build starts, with who at ProTax can reach it and for how long.

Offset from our working day standard time

  • London-5:30
  • Dubai-1:30
  • Singapore+2:30
  • Sydney+4:30
  • New York-10:30
  • San Francisco-13:30

ProTax is a brand of ProTax Consultancy. The contracting entity, the governing law and the invoicing currency are all named in the proposal before you sign anything.

09 Questions

What this sits between

Intelligent document processing
Written IDP: extracting structured fields from documents that arrive in no fixed layout.
Confidence threshold
The level below which an output is sent for human review instead of being accepted. Set per field, from the cost of an error.
Review queue
Where below-threshold output goes, with a named owner, a service level and escalation.
Straight-through rate
The share of items processed with no human touch. The number that decides whether the automation pays back.
Schema validation
Checking output has the expected shape and types before it is accepted, so a malformed answer fails loudly.
Drift
Behaviour changing over time, either because the provider updated the model or because the incoming documents changed.
Audit trail
A non-erasable log of each entry and each later change with its date. Control A.8.15 of ISO/IEC 27001, and the first thing an external audit asks to see.
GDPR
The EU and UK data protection regime. Article 83 sets administrative fines of up to €20 million or 4% of worldwide annual turnover, whichever is higher, which is why consent, retention and access control are build decisions here rather than paperwork.

09 Questions

What this sits between

Using a model to do the reading, sorting and drafting a person currently does by hand, with a review step for the cases the model is unsure about. It differs from rules-based process automation in that the input is unstructured: a scan, an email, a message, a paragraph.

Next step

Name the job your team does by reading and retyping.

We will pilot it on your own documents and come back with a measured accuracy figure per field, a cost per item, and the review load it implies.

Read by P J Divya, our CTO, not a form queue. Reply usually within one working day, in your time zone.

FROM THE PROTAX LIBRARY

Recent Articles & Guides

Practical reading on business records, tax filings and the decisions that come with running a company.

Newsletter

Stay ahead on compliance, tax & business updates

Get selected business and compliance updates by email. Once a month, no spam.

  • 100% privacy
  • 1 email / month
  • Unsubscribe anytime
Contact ProTax
Chosen by individuals and businesses

Get Help With Your Business

Tell us what you need help with. Our team will review your enquiry and explain the next steps.

Free Consultation No Obligations Expert Advice
FREE Consultation No obligation, no lock-in

Talk To ProTax

Share Your Business Requirement

Instant Response 100% Confidential Expert Advice
FREE Consultation No obligation, no lock-in

Request a Free Quote

Share Your Business Requirement

Instant Response 100% Confidential Expert Advice
The reading and typing work, handled A model drafts a person approves the uncertain