What AI actually does in a CA firm in 2026 (and what it still cannot)

Every CA firm's website in 2026 says something about AI. Very few can tell you what their AI actually did yesterday. This article is the second kind: a plain description of where AI works inside a real Chartered Accountancy practice, what it changed for our clients, and where we deliberately keep humans in charge. I lead Trivida, ADAPT's technology arm, and everything below is running in production today - not on a roadmap.

The honest starting point: most "AI in accounting" is not AI

Let us clear the marketing fog first. A large share of what gets sold as AI in Indian accounting is rule-based automation that has existed for a decade: auto-import of bank statements, scheduled GST filing reminders, template-driven invoices. Useful, but not intelligent.

What changed in the last three years is that machine learning and large language models became reliable enough for finance work where the input is messy: unstructured invoices, badly scanned documents, inconsistent vendor names, free-text narrations in bank statements. That is the layer where real AI earns its keep, because rules break on messy input and models do not.

When you evaluate any firm or product claiming AI, ask one question: what happens when the input is something the system has never seen before? If the answer is "it stops and asks a human", you are looking at a well-designed system. If the answer is vague, you are looking at a demo.

Five places AI does real work in our practice

1. Bank and ledger reconciliation. This is the flagship. One of our manufacturing clients used to have two people spend most of every day matching bank entries, purchase registers and ledgers. Today an automated pipeline ingests the bank feed, classifies each transaction, posts the routine ones, and flags exceptions. The finance team spends about 15 minutes each morning reviewing a short list of flagged items. Everything else posts automatically, reconciliation is effectively real time, and the audit trail is cleaner than it ever was manually. We wrote about this system on the Trivida page; it is the clearest before-and-after we own.

2. GST data matching. GSTR-2B versus purchase register matching is the single most painful recurring task in Indian compliance, because vendor names, invoice numbers and dates never quite agree between systems. Fuzzy matching models handle the near-misses that exact-match rules reject: "Shree Ganesh Traders" and "Sri Ganesh Trader" resolve to the same vendor, transposed invoice digits get caught and paired. The result is more input credit claimed on time and fewer month-end firefights. Given that unclaimed GST input credit is one of the most common leaks we find in Virtual CFO diagnostics, this is money, not convenience.

3. Document extraction. Invoices, salary registers, loan sanction letters and scanned agreements arrive in every format imaginable. Extraction models read them and produce structured entries for review. During audit season this is the difference between a junior spending a week retyping fixed asset registers and spending a week actually auditing them.

4. First drafts of routine analysis and correspondence. MIS commentary, variance explanations, replies to routine departmental notices - AI produces first drafts from the underlying numbers, and a partner edits and signs. The drafting time drops by well over half. The judgment, and the signature, stay human. This matters enough that I will repeat it below.

5. Anomaly detection. Models that have seen a client's transaction history notice what a tired human misses at 7 pm: a duplicate vendor payment, a GST rate that changed on one item, an expense posted to the wrong company in a group. We catch these in review queues now, before they become audit findings later.

What this means in numbers

Across clients where these systems run, the pattern is consistent: daily accounting effort drops from hours to minutes, month-end close compresses from weeks to days, and error rates fall to near zero on the automated portion because software does not get bored on entry number four hundred. The finance headcount does not disappear - it moves up the stack, from data entry to review and analysis. Not one client of ours has automated a person out of a job; several have cancelled plans to hire a third data-entry operator.

What AI still cannot do - and should not

This is the section most AI articles skip, and it is the one that matters most if you run a business.

Judgment under ambiguity. Whether a transaction is capital or revenue in nature, whether a tax position is worth defending, how aggressive to be in a GST refund claim - these are judgment calls with case law, context and risk appetite behind them. A model can summarise the precedents. It cannot own the call. A partner signs our opinions, every time.

Representation. An AI cannot appear before a GST officer, an assessing officer or a tribunal. Litigation strategy in our GST practice is human work, informed by documents that AI helped organise.

Responsibility. An audit report carries a membership number and personal liability. That accountability structure is the entire point of the profession, and no software vendor accepts it. Any firm that tells you AI "does the audit" is describing either a lie or a malpractice claim in progress.

Knowing when it is wrong. Models fail confidently. Every automated pipeline we build has confidence thresholds, exception queues and a human review layer precisely because the worst failure mode is not a wrong answer - it is a wrong answer nobody looked at.

What to ask your CA firm in 2026

If you are choosing or evaluating a firm, three questions cut through the AI marketing quickly:

1. "What runs automatically in your practice today?" Listen for specifics: which process, which client type, what the human review step is. Vague answers mean the AI lives in the brochure.
2. "What would my monthly close look like with you?" A firm using these tools well should be comfortable committing to dates and turnaround times, because machines do not slip deadlines.
3. "Who reviews the machine's work, and who signs?" The right answer names a person. Ideally a partner.

The unfair advantage of accountants who build

Most businesses buying "AI for finance" face a choice between software vendors who do not understand Indian compliance and accountants who do not understand software. The gap between those two is where projects die: the tool is technically fine but books entries the wrong way, or the accountant knows exactly what is needed but cannot build it.

Trivida exists because we got tired of that gap. We are Chartered Accountants who write software, so the systems we build are designed around GST law, TDS mechanics, audit requirements and how an Indian finance team actually works. We implement Zoho One, Odoo and Dynamics 365, we build custom automation where no product fits, and we run the same systems on our own practice first - which is the strongest quality control there is.

If your finance team is drowning in manual work, the first conversation is free. Tell us which process costs you the most hours - reconciliation, GST matching, month-end reporting - and we will tell you honestly whether it needs AI, plain automation, or just a better process. Talk to Trivida, or reach us at info@adaptassociates.com.