The AmplifierReal cases · Professional services · Small practice
Dev, 29: a two-partner CA practice in GST season
Two hundred clients, the same forty questions, and a fear of being the accountant who trusted a chatbot with a filing.
Planned for release phase 2. The page is complete as a draft; interactive parts run in the browser and do not yet save to an account.
In brief
The situation. Dev and his partner run a chartered-accountancy practice in Indore. Every filing season, the same thing: two hundred clients, the same forty questions by WhatsApp, and twelve-hour days spent answering them instead of doing the work that needs a CA. He has read the stories about professionals citing rules a tool invented. He would rather be slow than wrong, and has been both.
Step 1 — Define the End State
"During the season, no client waits more than a day for an answer to a routine question, no answer goes out that a partner has not seen, and the partners' evenings go to the twenty clients whose situations are actually complicated. A client should feel the practice got faster and no less careful." Audiences: the routine client, the complicated client, his partner, and the tax officer who reads a filing.
Step 2 — Map the Flow
His list: sort the forty questions into routine and not, write answers to the routine ones once, build a way to send them, keep partner review on everything. AI review adds: the answers must carry the date and the rule version, because rules change and a stale answer is a wrong one; and a log of which answer went to whom. It suggests an AI chatbot for clients. He declines flatly: cost of being wrong is high and the client cannot evaluate the output.
Step 3 — Assign Tasks
Deciding what counts as routine: his and his partner's. Drafting the forty standard answers: AI drafts; a partner verifies each against the rule text, line by line. Classifying an incoming question as routine or not: hybrid — the tool suggests, a junior confirms. Answering the complicated twenty: humans only. Sending: a junior, from the verified bank.
Step 4 — Select the Tool
A conversational tool for drafting, with the rule texts pasted in rather than trusted from memory; a document tool that holds the verified answer bank; no client-facing AI. Frequency is seasonal but recurring, so the bank is built to be reused next year with a date stamp.
Step 5 — Prepare and Direct
Criteria: every answer cites the section it relies on and the date it was checked; no answer contains the words "generally" or "usually"; a partner's initials on every entry before it is used; any question the classifier is unsure about goes to a human by default.
Steps 6 and 7 — Execute and Evaluate
The drafts are good and, in six of forty, wrong — a superseded threshold, a rule that changed in the last budget, one outright invention. The criterion catches all six because a partner checks each against the text. Dev notes: the tool was confidently wrong in exactly the way he feared, and the framework made that a Tuesday afternoon rather than a notice from the department.
Step 8 — Integrate and Stress-Test
Season starts. Small fix: two answers too long for WhatsApp; shortened. Structural: the complicated clients are still waiting, because the juniors escalate too much — the "unsure goes to a human" rule is right and the threshold was set too cautiously, a Step 5 criterion needing tuning. Adjusted after a week's log.
What changed
Routine answers go out in hours, verified. Partners' evenings go to the twenty. Next season the bank is re-verified against the new rules in a day, not rebuilt. Dev is still slow where slow is right, and no longer where it isn't.
Paste the rule text in; never trust the tool's memory of a rule. "No 'generally', no 'usually'" flushes out the hedging that hides an invented answer. Decline client-facing AI where the client cannot evaluate the output.
A composite case, not a client engagement; the person is invented, the situation is not.