Photools cut invoice
processing time by 92% and
ended duplicate payments

A photography equipment retailer was keying 2,500 vendor invoices a month into Tally by hand. Silicon Power Solutions replaced that with an Amazon Bedrock pipeline that reads, validates and posts each invoice in under three minutes — and pays for itself in four.

RESULTS AT A GLANCE

Measured over 12 months post go-live Before After Change
Time to process one invoice 20–30 min 2–3 min −92%
Data entry error rate 5–8% <1% −95%
Duplicate payments Recurring Zero −100%
Vendor matching accuracy 85% 98%+ +13pt
Invoice-to-payment cycle 5–7 days 1–2 days −4 days
Audit preparation, per year 100+ hrs 20 hrs −80%
Annual labour cost ₹96 L ₹11 L −₹85 L
Return on investment - 4 months payback

IN PLAIN TERMS

What actually changed, without the jargon

Photools gets around 2,500 vendor invoices a month by email, covering cameras, lenses, lighting and accessories. Every one of those used to be typed into Tally by hand — checking the vendor, the amounts, the GST rate, and whether it had already been paid before. That took 20–30 minutes per invoice and mistakes were common: wrong amounts, wrong GST rates, and the same invoice occasionally getting paid twice.

Here's what we built instead :

  • The invoice arrives by email as a PDF, PNG, JPG or scanned file.
  • A computer reads it automatically — pulling out the vendor, invoice number, amounts, GST and line items.
  • An AI checks everything at once : Does this match our vendor list? Do the numbers add up? Is the GST rate correct? Have we already paid this invoice before?
  • If everything checks out, it posts straight into Tally — no typing required.
  • If something looks off or unclear, it goes to a person to check, so nothing risky happens automatically.

The result, in plain terms :

  • An invoice that took 20–30 minutes now takes 2–3 minutes.
  • Data entry mistakes dropped from up to 8% of invoices to less than 1%.
  • Duplicate payments — a recurring problem before — have stopped completely.
  • Getting audit-ready went from 100+ hours a year to about 20.
  • Altogether, the system paid for itself in about 4 months and now saves roughly ₹85 lakhs a year in labour alone.

In one sentence : we replaced slow, error-prone manual invoice entry with an AI assistant that reads each invoice, checks it against what we know is true, and only asks a person when something's genuinely unclear.

THE CUSTOMER

Photools
INDUSTRY Photography equipment retail and e-commerce
HEADQUARTERS [CITY], India
OPERATIONS Online store and physical showroom
VOLUME 2,500 vendor invoices per month
FINANCE SYSTEM Tally
DELIVERED BY Silicon Power Solutions

THE CHALLENGE

Growth in orders meant growth in headcount

Every vendor invoice at Photools arrived by email as a PDF and was typed into Tally by hand. Across camera, lens, lighting and accessory brands — and two sales channels — the finance team was spending roughly 1,250 hours a month on data entry alone.

  • 20–30 minutes per invoice covering entry, vendor matching, line-item and GST verification, and cost-centre allocation, with throughput capped at 50–100 invoices per person per day.
  • A 5–8% error rate — transposed amounts, wrong GST rates against HSAC codes, vendor name typos — each costing around ₹3,000 to unwind once rework, payment delay and late fees were counted.
  • No duplicate detection. The same supplier appeared under several name formats across the e-commerce and showroom invoice streams, so duplicates passed into Tally and got paid.
  • Manual GST validation ran at a 2–3% error rate, creating penalty exposure, and audit trails were incomplete enough to cost 100+ hours a year in audit preparation.
  • A 5–7 day payment cycle strained supplier relationships and forfeited early-payment discounts on invoices that would otherwise have qualified for 10-day terms.

Month-end, quarter-end and festive-season peaks produced backlogs the team could not absorb. Invoice volume scaled with retail growth; finance headcount could not.

THE SOLUTION

An event-driven invoice pipeline on AWS

Silicon Power Solutions built a serverless pipeline that takes an invoice from the inbox to a posted Tally entry without a human touching it — unless the system decides a human should. Amazon Bedrock does the reasoning; every figure it produces is checked against a system of record before anything is posted.

Amazon SES → Amazon S3
  • Receipt rules pull attachments straight to encrypted S3 storage. Only PDF, PNG, JPG and TIFF up to 20 MB are accepted; file signatures are checked against declared types and scanned before storage.
Amazon Textract
  • Vendor, invoice number, date, amounts, GST rate, line-item tables, tax breakdowns and payment terms — returned as structured JSON with a confidence score on every field.
Amazon Bedrock — Claude 3.5 Sonnet
  • Vendor names normalised against the master list, line items recomputed against totals, GST rates validated against the HSAC code, duplicates detected across recent history, and a plain-language audit explanation generated for every decision.
Retrieval-augmented generation
  • Vendor master, current GST rate tables, cost-centre mappings and recent invoice history are retrieved at the moment of validation. Onboarding a new brand supplier is a data update, not a model retrain.
Amazon DynamoDB
  • Extracted fields, verdicts, model reasoning and audit metadata, encrypted with customer-managed keys and 35-day point-in-time recovery. This is the record auditors now read instead of a filing cabinet.
AWS Elastic Beanstalk
  • Extracted fields, verdicts, model reasoning and audit metadata, encrypted with customer-managed keys and 35-day point-in-time recovery. This is the record auditors now read instead of a filing cabinet.
Choosing the model was a cost decision, not just a quality one

Four foundation models were benchmarked against the same validation set. Claude 3.5 Sonnet reached 97.5% accuracy at $0.045 per invoice; Claude Opus reached 98.2% at $0.082. The 0.7-point gain did not justify an 82% higher unit cost at Photools' volume, so Sonnet became the default and Opus is invoked only for complex multi-page and multi-currency invoices.

Guardrails

Model output is schema-validated with enum enforcement before it can reach Tally. Every assertion is cross-checked against the vendor master and GST tables rather than trusted. Low-confidence results route to a human review queue, and a monthly governance audit samples 50 validation outcomes across vendor tiers and cost centres.

  • Amazon Bedrock
  • Amazon Textract
  • Amazon SES
  • Amazon S3
  • Amazon DynamoDB
  • AWS Elastic Beanstalk
  • AWS KMS
  • AWS CloudTrail
  • Amazon CloudWatch

THE OUTCOME

1,125 hours a month went back to the business

The finance team now spends 125 hours a month on invoices instead of 1,250. That time moved to vendor reconciliation and stock planning — work that was previously deferred indefinitely.

What is measured, and how
  • Processing time — CloudWatch tracks start-to-finish duration per invoice, alerting above a five-minute average.
  • Accuracy — finance reviews a monthly sample against manual verification; anything below 98% is investigated.
  • Validation pass rate — the share of invoices clearing without human review, held at 85–90%.
  • False positives — flagged by finance when the system escalates unnecessarily; above 3% triggers prompt refinement
  • Cost per invoice — Cost Explorer spend divided by CloudWatch invoice count, monitored for month-on-month drift above 20%.
  • Payment cycle — the timestamp delta between S3 upload and Tally posting.

Silicon Power Solutions transformed our order operations with AWS generative AI. Cutting processing time from 45 minutes to under 3 while achieving 100% pricing consistency across 10 brand catalogues has completely removed our growth ceiling.”Kanti Khemani, Director, Photools

THE ECONOMICS

₹13.5 lakhs a year replaced ₹1.65 crores

Running the pipeline costs roughly ₹21,000 a month in AWS services at Photools' volume, plus retained finance capacity for exception handling. The manual process it replaced cost approximately ₹1.65 crores a year in labour, rework and audit effort.

  • Unit cost is measured, not modelled. $0.08–0.12 per invoice, split roughly $0.03 Textract, $0.04 Bedrock, $0.02 storage and compute.
  • Consumption pricing removes the over-provisioning premium. Cost follows invoice volume, so festive-season peaks cost proportionally and quiet months cost less — no idle capacity.
  • No custom model was trained. Using managed foundation models removed GPU training spend, MLOps tooling and data science headcount entirely, and cut deployment from months to weeks.

WHAT WE LEARNED

Three things worth carrying into the next deployment

Benchmark models before you pick one

A formal comparison across four models found a 3x spread in cost for a four-point spread in accuracy. Reaching for the most capable model by instinct would have meant overpaying 82% per invoice on routine work.

Never let the model be the last word on a number

Bedrock output is recomputed and cross-checked against the vendor master and GST tables before posting. This caught extraction errors that were internally plausible and would otherwise have gone through cleanly.

Explainability paid for it self

Generating an audit summary for every decision is what collapsed audit preparation from 100+ hours to 20. A conventional OCR pipeline would have delivered the speed and none of that saving.

NEXT STEP

Processing invoices by hand?

Silicon Power Solutions builds generative AI document pipelines on AWS for finance and operations teams across India. Most deployments reach production in weeks, not months.