Chartered accountants, lawyers and consultants live in documents — 80-page agreements, GST notices, audit reports, board minutes. AI can genuinely halve the time you spend reading and summarising them. It can also get you into serious trouble if you paste the wrong thing into the wrong place. Both are true. Here's how to get the first without the second.
What AI is genuinely good at in a professional practice
- Summarising: "Give me the five commercial risks in this agreement, with clause numbers." Claude does this well and cites the clause it's referring to.
- Structuring: turning a messy client email thread into a clean list of facts, dates and open questions.
- First drafts: an engagement letter, a reply to a routine notice, a checklist for a filing — from your own template.
- Explaining: "explain this clause to a client who isn't a lawyer" — plain-English versions in seconds.
- Numbers: reasoning models (DeepSeek R1) walk through calculations step by step, so you can audit the logic.
The two questions to ask before you paste anything
1. Would I be comfortable if this text appeared in a data-breach report? Names, PAN/Aadhaar numbers, bank account details, medical or salary data, unreleased financials — these should not go into a prompt in raw form. Redact first. Most tasks don't need the identity to be useful: "Client A", "Company B", "₹X" work fine.
2. Is the answer going to depend on current law or rates? Tax thresholds, filing dates, penalty rates change. A model answering from training data alone can be confidently outdated. Use a web-grounded model (Gemini in Kryotta) and check the source it cites — or better, verify against the official notification. AI is a research assistant, not a source of law.
A safe workflow that actually saves time
- Redact the document (client names, IDs, account numbers).
- Summarise with a clear prompt: "You are helping a CA. Summarise this GST notice: what is alleged, the amount, the response deadline, and what documents are requested. Bullet points, cite paragraph numbers."
- Cross-check anything factual or numeric with a second model — the Compare arena shows both side by side; disagreements are where you should look harder.
- Draft the response from your template, then edit like you would a junior's draft.
- Verify legal or numeric claims against primary sources before anything leaves your office.
Prompts that work well
- "List every date and deadline mentioned in this document as a table: date, what happens, which clause."
- "Compare these two versions of the agreement and list every substantive change. Ignore formatting."
- "Rewrite this paragraph for a client with no legal background, 3 sentences, no jargon."
- "Check this computation step by step and tell me if any step is wrong: …" (use a reasoning model).
What NOT to do
- Don't upload a full KYC folder "to save time".
- Don't ask AI "is this legal?" and act on it without checking — ask "what should I verify?"
- Don't let a client-facing document leave with a model's confident-but-unverified figure.
- Don't use a shared personal account for client work; use a workspace where memory is scoped and you control who's in it.
Why a workspace, not five apps
For a practice, the useful bits are boring: workspaces with seats, memory that stays inside the workspace, every answer labelled with the model that produced it, and one bill in rupees. You get Claude for the drafting, Gemini for the "what's the rule now", and R1 for the numbers — without documents scattered across tools.
Used this way, AI doesn't replace professional judgement. It gives you more time to exercise it.
