03 Design
Why set the threshold per field?
Because the fields are not equally dangerous. One threshold across a document either wastes review effort on the safe fields or lets the risky ones through, and both are avoidable.
| Field | Measured accuracy | Cost of an error | Threshold we set |
|---|---|---|---|
| Invoice amount | 99.2% | A wrong payment | High, plus a rules cross-check on the arithmetic |
| GSTIN | 97.4% | A blocked input tax credit | High, plus format and checksum validation |
| Vendor name | 71.8% | A misposted ledger, easily spotted | Always review, until accuracy improves |
| Invoice date | 98.6% | A wrong ageing bucket | Medium, with a sanity range check |
| Line description | 88% | Almost none | Permissive, no review |
Step 1: Measure per field
On 50 to 200 of your own documents, with the correct answer defined by your team rather than by us.
PilotStep 2: Price the error
What a wrong value actually costs, which is a business judgement and yours to make. We supply the accuracy and the review load each choice implies.
DecisionStep 3: Add the rules layer
Format checks, checksums, arithmetic that must tie and range sanity tests. A model should never be the thing that adds up a column.
BuildStep 4: Staff the review queue
A real queue with a named owner, a service level and escalation, exactly as an approval workflow would have.
BuildStep 5: Watch the review share
A rising proportion below threshold is the earliest warning that the document mix or the provider model has changed.
Monitor
04 Governance
The two rules that follow the data
Both are cheaper designed in than retrofitted, and both are questions an audit or procurement team will eventually ask in writing.
- Personal data. Documents contain names, phone numbers, addresses and bank details, and sending them to a model is processing them. The Digital Personal Data Protection Rules, 2025, notified on 13 November 2025 by the Ministry of Electronics and Information Technology, apply, with the main obligations commencing on 13 May 2027.
- The accounting record. Anything posting entries is part of the accounting system and inherits the Rule 3(1) audit trail obligation administered by the Ministry of Corporate Affairs, in force since 1 April 2023, with the eight financial year retention Section 128(5) of the Companies Act, 2013 requires.
- 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.
05 Reference
Terms used on this page
- 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.
- 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, required of company accounting software since 1 April 2023.
- Data Protection Board of India
- The adjudicating body under the Digital Personal Data Protection Act, 2023. The Schedule to that Act sets fixed rupee ceilings rather than a share of turnover: up to ₹250 crore for failing to take reasonable security safeguards against a personal data breach, up to ₹200 crore for failing to notify the Board and the people affected, and up to ₹50 crore for a breach of any other provision. That is why consent, retention and access control are build decisions here rather than paperwork.
06 Questions
AI automation in Howrah: FAQs
The variables are the same wherever you sit: the number of use cases in production, the monthly volume of each, and how much review the confidence threshold leaves in place. There is no city premium, because the pilot, the review queue and the monitoring are the same pieces of work in Howrah as anywhere else. We quote after the pilot. Every quote states IncorpX professional charges; model usage is billed by the provider at actuals, and any government fees are billed separately at actuals.
4 to 8 weeks including the pilot. The pilot is the first two weeks and it is where the decision gets made, because there is no point building against an accuracy figure nobody has measured.
No. Document sampling, evaluation, integration, reviewer training and monitoring all run over calls and screen shares from our base in Chennai, Tamil Nadu, with the documents moving over an encrypted channel.
When the input has no fixed shape. If every invoice arrives in one template, rules are cheaper, faster and more reliable. If they arrive from three hundred suppliers in three hundred layouts, that is where a model earns its cost.
Supplier invoice and purchase order entry, email and ticket triage, first-draft support replies, contract term extraction, and comparing documents against a system record. All five are high-volume reading tasks with a checkable answer.
Always, on 50 to 200 of your own documents with the correct answer defined by your team, and reported per field. An extractor at 95% overall can be 99% on the amount and 70% on the vendor name, and only the per-field view tells you what to review.
They go to a person, by design. Anything below the confidence threshold enters a review queue with a named owner and a service level, which is what turns a model that is right most of the time into a process that is right almost always.
Yes, and the audit trail obligation follows it there. The proviso to Rule 3(1) of the Companies (Accounts) Rules, 2014 has required company accounting software to keep a non-disableable edit log since 1 April 2023.
It is mapped before the pilot runs. The Digital Personal Data Protection Rules, 2025 were notified on 13 November 2025, with the main obligations commencing on 13 May 2027. We redact what the task does not need and record where the rest goes.
Then we say so and you have spent a pilot fee rather than a build. That happens often enough that we treat it as a normal outcome, and you keep the test set either way.
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 Dhanush Prabha, our CTO, not a form queue. Reply usually within one working day, in your time zone.
Nebin Binoy 
