Sensitive data exposure
Personal, financial or engagement information may be uploaded without an approved data-handling path.
Alert: information location may not be centrally confirmedSINGHI & CO. · PROPOSED BY THIRDEYE DATA
Governed AI for
everyday audit work
See the same audit query in two ways: today’s independent AI process and the proposed secure, repeatable and measurable ThirdEye AI process.
Demonstration only: all data, names, outputs and measured results shown here are fabricated. Benefits are illustrative planning assumptions to be validated during the PoC.
ThirdEye AI
Secure audit helper
ExcelLedger
PDFEvidence
DOCMethod
✓Approval
Shared query examplesStarting library, not a limit
Initial usersSmall, controlled Phase 0
Hours per yearIllustrative at 200 users
Governed data locationProposed Azure environment
1
As IsPersonal AI tools and repeated manual work
2
What we proposeOne governed audit helper
3
To BeSource-linked output with human approval
4
Business valueMeasured time, cost, security and quality
AS IS · CURRENT WORKING MODEL
Useful AI, but outside one firm-wide control
Teams may use ChatGPT, Claude or Gemini on personal or separate accounts. The work can be helpful, but the firm has limited visibility over files, prompts, answers and access.
! Current-session control alerts
- Files may contain personal, financial or engagement information.
- Office identity and engagement permission are not checked here.
- Prompt, source, calculation and approval are not in one firm record.
- Numeric results and citations require complete manual validation.
Files uploaded manually
This screen illustrates a separate general AI session. It does not represent any specific vendor’s exact interface or policy.
AI
Outside firm-wide workflowChatGPT-style sessionPersonal or separately managed login · illustrative
✦
How can I help with the audit query?
Select a query, review the uploaded file names and send the prompt.
CAAudit professionalChooses a general AI tool
Uploads again
GPTPersonal login
ClaudeSeparate history
GeminiSeparate settings
Copies output
Working fileAnswer copied manually?
HigherGovernance exposureCurrent unmanaged pattern
Central visibilityLow
Engagement separationNot enforced
Source traceabilityInconsistent
RepeatabilityPrompt dependent
No common identity control
Without central identity, access may continue after a role change or exit unless every tool is updated separately.
Alert: joiner and leaver control remains manualNo engagement data wall
A general AI login does not automatically understand which engagement records a person may access.
Alert: separation depends on individual careAnswers may lack evidence
A confident answer can still be incomplete or wrong when the source is missing, unclear or outdated.
Alert: professional validation takes additional timePrompts differ by person
Two people may ask the same audit question differently and receive different outputs.
Alert: firm methodology is difficult to standardiseNo complete audit trail
The firm may not have one record of the question, files used, response, edits and approval.
Alert: review evidence remains fragmentedNumbers can be unreliable
Language models can explain calculations but should not be the only engine used for depreciation, variance or Benford testing.
Alert: numeric results need reproducible rulesCost rises with every licence
At ₹1,750–₹2,600 per person each month, licence spend increases as more people join.
Illustrative: ₹42–₹52 lakh yearly at 200 usersILLUSTRATIVE QUERY
“Review the annual journal dump and show entries posted on national holidays, the largest posters and unusual number patterns.”
journal_dump.xlsxholiday_list.xlsxapproval_policy.pdf
16 hoursCurrent preparation and checking
vs
4 hoursIllustrative governed workflow
12 hours potentially returned for review
THIRDEYE AI WORKSPACE
The same audit question, now inside a governed workflow
Select a use case and run the governed workflow. The platform verifies access, processes approved files, applies repeatable calculations and prepares an evidence-linked result for professional review.
AI
ThirdEye AI WorkspaceGoverned query and prompt workflow
✓ Office identity✓ MFA verified✓ Meridian FY26 access
Proposed Azure India environment
Reference files
Ready to analyse
The demo will show how the selected files, prompts, calculations and review steps are handled.
Figures shown inside Query Studio are fabricated illustrations. Actual time, usefulness and accuracy will be measured during the PoC using agreed representative data.
APPROVED QUERY & PROMPT LIBRARY
Senior knowledge becomes a repeatable working method
Each use case can use several prompts, calculation rules and review checks. Access depends on role and engagement permission.
92550+
The nine shared queries are the starting point, not the limit
New approved queries can be added as teams use the platform. The same method, access rules and review controls continue as the library grows.
TO BE · PROPOSED WORKING MODEL
Files stay governed while the audit team works faster
ThirdEye AI brings approved data, reusable query flows, reproducible calculations and human review into one controlled environment.
UPSTREAMAuthorised source information
XLSLedgers & registers
PDFInvoices & policies
DOCMemos & methods
EMLApproved emails
PPTPrior presentations
TXTNotes & extracts
AZURE INDIA ENVIRONMENTProposed governed boundary
◆
Microsoft Entra IDOffice identity · MFA · role and engagement groups
ACCESS VERIFIED
AI
ThirdEye AIOne secure audit workspace
Approved query libraryReliable calculation rulesDocument understandingSource retrievalConfidence checksAudit logging
Encrypted engagement filesEngagement AEngagement BEngagement CSeparate access walls
DOWNSTREAMReview-ready audit outputs
01Exception schedule
02Recomputed values
03Source-linked synopsis
04Memo or presentation draft
05Working paper
✓Review & approval trail
AI
ThirdEye AI preparesExtracts, calculates, compares and drafts
CA
Professional reviewsChecks sources, assumptions and exceptions
M
Manager approvesConfirms completeness and response
P
Authorised professional concludesResponsibility always remains human
SECURITY & PRIVACY
Access before intelligence
The proposed design checks identity and engagement access before any information can be retrieved. Important activity is recorded for review.
✓GovernedDefence in depth
MFAIdentity check
RBACRole access
IndiaData location
LogsActivity record
LIVE ACCESS TEST
Can Asha open another engagement?
AMAsha MehtaAssigned: Meridian FY26
Requests Aster Retail FY26
?Access checkWaiting to test
Audit log will appear after the test.
The architecture is designed to reduce and control the risk of data compromise. It does not remove residual risk or professional responsibility.
ILLUSTRATIVE BUSINESS VALUE
Time saved becomes additional capacity for professional work
Adjust the number of users to see the planning model. Actual benefits will be validated with Singhi & Co.’s real utilisation data during the PoC.
Planning assumptions45 working weeks · ₹500 blended hourly value · 2,040 hours per FTE
Potential time returned each year9,000hours
Routine queries27,000per year
Capacity equivalent4.4FTE
Time value₹45.0 Lper year
QUERY CAPACITY
Same time can support more review-ready queries
In a five-hour work block, the same professional could prepare 10 queries today or review approximately 30 machine-assisted queries under the illustration.
ANNUAL LICENCE PLANNING
Central economics improve as adoption grows
₹19–₹29 Lillustrative yearly difference at 200 users before implementation cost and subject to actual quotations and usage
EXAMPLE BY AUDIT QUERY
Preparation time can reduce while human review remains
As IsTo Be
≈9,000hours returned yearlyAt 200 users under the base illustration
4.4 FTEcapacity equivalentReleased for investigation, review and advisory work
≈₹22 Lrework valueIllustrative half an hour avoided per person per week
₹10–₹13 Lper 50 licences retiredIndicative yearly licence saving after proven coverage
≥95%grounded answer targetPoC target for correctly sourced knowledge answers
0fabricated citation targetPoC target for source-linked knowledge answers
Same dayleaver access removalTarget operating control after identity integration
100%seeded anomaly targetPoC target on an agreed test dataset
01
Responsible AI leadershipA visible and controlled approach to AI adoption within professional services.
02
Firm knowledge becomes reusableApproved queries and methods become available according to role and permission.
03
Future external offeringA proven internal platform can later support an external service, subject to a separate decision.
04
Aligned with India’s AI directionA governed India-based foundation can evaluate relevant national AI initiatives as they develop.
PHASED IMPLEMENTATION
Benefit is measured before each scale-up
The platform starts with a small group and grows only after security, usefulness and business value are demonstrated.
ID
Phase-wise identity foundationMicrosoft Entra ID is introduced gradually with the user rollout
Phase 010–15 pilot identitiesSSO · MFA · initial role and engagement groups
Phase 150 governed identitiesRole groups · joiner, mover and leaver process
Phase 2100 governed identitiesAdditional teams · stronger access policies where licensing permits
Phase 3150–200 governed identitiesAutomated lifecycle · periodic access review
10–15Selected users50First teams100Wider adoption150–200Planned scale
FABRICATED SAMPLE DATA
Download the audit demonstration pack
These files contain no real engagement or personal information. They support the examples shown in the Query Studio.
ZIP
Complete sample audit pack
Six datasets, reference notes, a sample email and the query guide used in this demonstration.
CSV
Fixed asset registerDepreciation recomputation · 160 demo assets
↓ CSVInventory purchasesQoQ and YoY price variance · 600 rows
↓ CSVSales transactionsPrice variance and trends · 720 rows
↓ CSVJournal entry extractHoliday, poster and Benford examples · 2,000 rows
↓ CSVStatutory dues registerDue date and deposit delay analysis
↓ CSVScrap sales extractVendor-wise price and volume trends
↓ TXTAudit methodology notesReference source for grounded answers
↓ EMLSupporting emailFabricated approval and context
↓ MDQuery library guidePrompts, steps and review checks
↓FINAL COMPARISON
The proposed change in one view
ThirdEye AI reduces repeated preparation and strengthens control. Audit judgement and final responsibility continue to remain with authorised professionals.
AreaAs IsTo Be with ThirdEye AIIllustrative value
AI accessPersonal or separate accountsOffice identity, MFA and role accessSame-day leaver control target
Engagement dataUploaded separately to general toolsGoverned Azure India environment with engagement wallsCross-engagement test must fail
Prompt methodRecreated by each personApproved reusable query and prompt library9 examples at start, then grows
Numeric analysisMay depend on a language-model answerReproducible calculation rules with an explanationSame data and rule give the same result
EvidenceMay be copied without a complete source trailSource-linked output and “no reliable source” response≥95% sourced-answer PoC target
Time≈30 minutes per routine query≈10 minutes including professional review≈9,000 hours/year at 200 users
CapacityPreparation absorbs professional timeMore time for investigation, review and advisory work≈4.4 FTE capacity equivalent
Licence planning₹42–₹52 lakh/year at 200 users≈₹23 lakh typical recurring case₹19–₹29 lakh indicative difference
AccountabilityHistories and approvals remain fragmentedQuestion, source, answer, edit and approval loggedComplete activity-record target
THE BUSINESS CASE
Less repetitive preparation. More professional review.
A small controlled start can prove security, usefulness and measurable value before wider adoption.
6 weeksPhase 010–15Initial users₹8.5 LPlanning reference