Tech Tips
AI Usage – Safety & Compliance Risk
Guidelines and examples highlighting where AI answers can hide compliance risk, and how to verify before relying on them.
Anonymisation is not deleting names.
Even if you remove names,
• Exact amounts
• Precise dates
• Rare fact patterns
• Section combinations
can re-identify a client matter.
Safer workflow before using tools like ChatGPT:
Step 1: Abstract the identities
- Client A / Company B / Vendor X
• Avoid city + industry combination if niche
Step 2: Neutralise the numbers
- ₹10,47,32,890 → “~₹10 Cr” or “₹X”
• Exact turnover → “mid-sized entity”
Step 3: Generalise the trigger
Instead of:
“Notice under Sec 148 dated 12/01/2026 for AY 2019-20”
Write:
“Reassessment notice for a past assessment year”
Step 4: Ask for structure, not solution
Request:
• Draft framework
• Checklist
• Risk factors
• Points to verify
Not:
“Rewrite this exact client reply”
Treat AI like:
A public strategy room
Not a confidential case file.
Professional judgment remains yours.
Before accepting an AI answer, set this rule:
“If anything is unclear, ask clarifying questions before answering.”
Why this matters:
• Reduces assumptions
• Improves relevance
• Cuts rework during review
Better inputs improve outputs.
Final judgment still rests with you.
Using AI for summaries?
Safe workflow:
• Use AI to *draft structure*
• You verify numbers, sections, clauses
Never ask AI:
“Summarise impact for my client” without checking the source document.
AI sounds confident — even when wrong.
Never paste client data by default.
Before typing into any AI chat, pause and ask:
“Would I paste this into an email to a stranger?”
If the answer is no,
rephrase the question without identifiers.
Before trusting an AI answer, ask this one line:
“List assumptions, uncertainties, and areas where this answer may be wrong.”
Why this works:
• AI normally hides doubt
• This forces it to expose weak spots
• You instantly know what to verify
Treat AI as an assistant, not an associate.
Judgment remains yours.
We saw why one safety question matters.
Now, a real compliance example.
Ask AI:
“Is Form 15CA required for all foreign remittances?”
AI’s first answer (often simplified):
“Form 15CA is required for foreign remittances made by residents.”
Now apply the safety line:
“List assumptions, uncertainties, and areas where this answer may be wrong.”
AI then surfaces what actually needs checking:
• Certain remittances are exempt under Rule 37BB
• Small-value payments may not require Form 15CA
• Nature and purpose of remittance matters
• Bank and remittance category differences apply
Lesson:
An apparently correct AI answer can still hide *compliance risk*.
The second question turns the answer into a *verification checklist*.
AI gives speed.
That one question gives safety.
Judgment remains yours.
GST Example (Real Compliance Risk)
Ask AI:
“GST rate on works contract for government buildings?”
AI’s first answer (often simplified):
“12% GST applies to works contracts provided to the government.”
Now apply the safety line:
“List assumptions, uncertainties, and areas where this answer may be wrong.”
AI then exposes things to check:
• Nature of building (residential vs commercial)
• Government or public use
• Specific notifications & amendments
Lesson:
A neat one-line answer can hide invoicing and tax risk.
The second question shows where your judgment is needed.
Speed + Safety = Smart AI use.
