AI in Company Secretarial Work: What Actually Works in 2026 (and What Doesn't)

Half of LinkedIn says AI will replace the company secretarial profession. The other half says AI cannot read a board resolution. Both are wrong, and the truth — like most truth in 2026 — is in the middle. This article is what we have actually learned shipping AI-assisted CS workflows over the last eighteen months.

What AI is genuinely good at now

• Draft generation from structured prompts. Board notices, agenda items, MGT-14 covering letters, DIR-3 KYC chasers, replies to standard SEBI queries. Given a template and a fact pattern, a CS-trained model writes a clean first draft in seconds.
• Summarising a filing. An AOC-4 of a peer company, a fifty-page LODR disclosure, an SEBI enforcement order — summarise the points, extract the dates, list the principal officers named. Hours of reading collapsed to minutes.
• Reading the gazette. MCA notifications, SEBI circulars, ICSI guidance — read, tagged, and surfaced to the team in the categories that matter to them.
• Finding the precedent. “Show me the last three SEBI orders involving Section 11B and listed company XYZ” — the kind of query that used to take a junior associate an afternoon, now answered in seconds.
• Director and company search. Cross-reference DIN, PAN, CIN across MCA, exchanges, and SEBI orders. Indispensable for due diligence and KYC.

What AI is still bad at

• Judgement. Whether a transaction is material, whether a director's interest needs disclosure, whether a specific RPT requires audit-committee approval — AI can frame the question, but the call belongs to the qualified CS.
• Edge cases. The unusual scheme of arrangement, the unique cross-border transaction, the litigation-driven board change — these still need a human who has seen the pattern before.
• Originating board strategy. AI can document a decision; it cannot make one for a board.
• Politically sensitive drafting. The reply that needs to land carefully with a regulator, the disclosure that has to be technically accurate without inviting follow-up — still a human craft.

The honest workflow that wins

1. AI drafts. CS reviews. The cycle that compounds productivity is not “AI does it” — it is “AI drafts in 30 seconds, CS reviews in 5 minutes.” The 5-minute review is non-negotiable.
2. Use a CS-trained model, not a generic one. The difference between a generic LLM and one fine-tuned on Indian CS material is night and day on Section 117(3) versus Section 179(3), or LODR Regulation 30 versus Regulation 30A.
3. Keep a structured log of AI-assisted artefacts. When the auditor or regulator asks how a particular draft was produced, you should have the answer in three clicks.
4. Pick tools that integrate, not chatbots that sit alone. A draft inside the workflow tool beats a draft pasted from a chatbot every time. The latter creates version drift.

What this means for the profession

AI does not replace CS. It changes the shape of the job. The mid-level associate whose role was 60% drafting and 40% judgement is moving towards 20% drafting and 80% judgement. The implication is structural: practices need fewer juniors per partner, and the juniors who remain need to upskill into judgement work faster than they did in 2018.

The practices that are growing fastest in 2026 are the ones that have:

• Adopted AI-assisted drafting deeply;
• Restructured pricing away from per-hour and toward per-engagement;
• Pushed the senior associates and partners into more advisory work because the routine drafting is no longer a bottleneck.

How Delta Filings approaches this

We built Delta Filings around the principle that AI should write the first draft of any reply, letter, or note a CS would otherwise write from scratch — but never the final word. Every drafted artefact in the tool is reviewable, editable, and version-controlled, with the model's reasoning visible to the CS. The CS-trained model knows Indian law; the workflow knows the firm's style. The combination is what turns the chart above from a benchmark into a daily reality.

The takeaway

AI in CS work in 2026 is not a future thing. It is a present thing being used badly. The practices that will define the profession five years from now are the ones building real AI-assisted workflows now — not the ones running ChatGPT in a separate tab.