AI for Company Secretaries: Board Packs, Minutes and Filing Reminders

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI for Company Secretaries: Board Packs, Minutes and Filing Reminders.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI for Company Secretaries: Board Packs, Minutes and Filing Reminders.

A company secretarial practice can use AI to draft minutes from the secretary's notes, assemble and check board packs against an agenda, and build a deadline tracker that sends reminders. Every output gets a named reviewer; the secretary controls what the record says, and the chair approves the minutes.

Company secretarial work is a good fit for careful AI use because most of it is structured, repetitive and templated. It is also a bad fit for careless use, because the minute book is a legal record and a filing missed is a penalty. So treat AI as a drafting and checking assistant with a named sign-off at each stage, and adopt the three jobs in order of risk, lowest first.

Follow me on Instagram@sagnikteaches

Order of adoption: tracker, then packs, then minutes

The three jobs carry different risks, so adopt them in order of how easily you can check the output.

Connect on LinkedInSagnik Bhattacharya
JobWhat AI doesRisk if wrongCheck
Filing trackerTidies the list, drafts reminders, flags gapsMissed deadline, penaltyDates confirmed from source by a person
Board packsCompares pack to agenda, drafts cover notes, builds indexMissing or wrong paper at the meetingSecretary reads the pack contents list against the agenda
Draft minutesTurns notes into minute formInaccurate legal recordSecretary checks every decision and name; chair approves

A practice that starts with minutes will be impressed in week one and frightened in week six. Starting with the tracker lets the team learn the failure modes of the tools on a task where a wrong answer is caught by a calendar. See how to tell if a process is ready to automate for a short test to run before each job.

Subscribe on YouTube@sagnikteaches

The filing tracker: AI tidies, a person owns the dates

Most practices run their deadlines from secretarial software, a spreadsheet, or both. AI can help in three ways: cleaning a messy spreadsheet into one structure, drafting the reminder messages, and flagging gaps. It should not be your source of statutory dates. Filing periods and their rules come from the law, the registry's guidance and each company's year end, and a language model can state a plausible rule that is wrong or out of date.

A tracker structure that works

ColumnExample (illustrative)Source of truth
Company and numberClient A, [registration number]Registry record
Year end31 MarchLatest filed accounts
Filing typeAnnual accounts; annual confirmationYour software or registry
Due dateCalculated by software or checked by a personLegal period applied to year end
Internal target dateDue date minus 14 daysPractice rule
Owner and reviewerInitialsPractice allocation
Status and date filedDraft prepared; awaiting director sign-offUpdated by owner

Paste your existing mess (three spreadsheets, a mailbox export and someone's notebook list) and ask the assistant to merge it into this table and list every row with a missing or conflicting year end. That list is the real value. Illustratively, a practice with 140 companies might find 11 rows with no year end and 6 duplicates, which is exactly the sort of gap that causes a missed filing.

The Complete Excel Course with AI Integration thumbnail

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Reminder messages

With the tracker in place, draft three standard reminders: 30 days out, 14 days out and 3 days out. Keep them short and specific, with placeholders in place of names. For example:

Subject: [Company] annual accounts - sign-off needed by [date]
Hello [first name],
We need your approval of the draft accounts attached by [internal target date] so we can file before the deadline on [due date].
Please reply "approved" or send changes. If we do not hear by [date], I will call you.
[your name]

Run the reminders through your email or workflow tool, not through the chat assistant itself. If you want automation, a connected workflow such as an email rule driven by the tracker's due-date column is easier to audit than a chat session. A related way of thinking about reminders and scheduling is in whether AI can manage your calendar and book meetings.

Board packs: assembling, checking and the cover note

A board pack is a bundle of papers with an agenda: previous minutes, management accounts, a CEO report, committee reports, resolutions to approve, and any papers for decision. The work is chasing, collating and checking, and AI helps with the last two.

  1. Agenda to checklist. Paste the agenda and ask for a table: each item, the paper expected, who owns it, the deadline and status. You update status as papers arrive.
  2. Completeness check. When the pack is assembled, paste the contents page and ask the assistant to compare it with the agenda and flag missing papers, mismatched titles and items with no paper. Do it before the pack goes out, not after.
  3. Consistency check. Ask for contradictions between papers: a figure in the finance paper that differs from the CEO's report, a decision date that appears twice in different forms, a resolution wording that differs from the earlier draft.
  4. Cover note. Ask for a one-page summary of what the board is being asked to decide at each item, in neutral wording. The secretary edits it, because the cover note shapes how directors read the pack.

A consistency check in action (illustrative)

A small manufacturer's board pack has 11 papers across 74 pages. The assistant is given the text of the finance paper and the CEO report and asked to list every number that appears in both with a difference. It returns two: monthly revenue is $412,000 in one and $421,000 in the other, and the headcount at the start of the quarter differs by two. Both turn out to be transposition and timing differences, fixed in ten minutes. Without the check, the chair would have found one of them in the meeting.

Confidentiality note: board packs commonly include price-sensitive or personal data. Use a business-tier account, redact names where you can for checks that do not need them, and follow the client's own rules. Long packs also need care with file limits: Claude accepts PDFs of up to 1,000 pages but analyses charts and images only in files of 100 pages or fewer, reading the text alone above that, so split a long pack and check tables by eye. The method for chunking documents safely is in summarising bundles and transcripts with AI safely.

Draft minutes from notes: the routine in detail

Minutes are the highest-value and highest-risk task. The rule that makes it workable is that AI drafts from your notes, and you decide what the record says. The secretary writes brief notes in the meeting (decisions, who proposed and seconded, votes, actions, declarations of interest) and AI turns them into the practice's minute format.

Set the house style once

Create a project in your assistant (ChatGPT Business and Claude's Team plan both let colleagues share a Project; Claude's individual plans cannot share one) holding a sample of approved minutes with the names redacted, a list of your standard phrases ("It was resolved that", "The board noted"), and a rule sheet: past tense, no verbatim discussion, record decisions and reasons, list actions with owner and date. Keep one project per client so house style and previous minutes stay together.

You are drafting board minutes from the secretary's notes below. Follow the house style in the project files.
Rules:
- Record only what is in the notes. Do not add discussion, reasons, names or figures that are not in the notes.
- Where the notes are unclear or a decision lacks a proposer or outcome, write [CHECK: what is missing] in the text.
- Number resolutions as in the notes. Put actions in a table: action, owner, due date.
- Start with attendance and apologies exactly as listed, and declarations of interest if any.
Notes:
[paste notes]

The [CHECK] instruction does the most important job: it converts gaps into visible flags instead of plausible filler. A realistic output contains four or five such flags in a 40-minute meeting, and each one sends you back to the notes or, if you recorded the meeting, to the recording, or to a quick message to the chair.

What to fix in a first draft

  • Tidied-up decisions. Notes say "agreed to look at the lease options"; the draft says "resolved to renew the lease". That is a legal difference. Fix to what happened.
  • Imported names. The assistant fills a missing proposer with a name mentioned elsewhere. Remove it, and check.
  • Smoothed votes. A split decision becomes "unanimously approved". Check against the notes.
  • Invented background. A sentence of rationale that nobody said. Delete it.

Because the same prompt can produce different wording each time, expect variation, and treat each draft as one version of a first draft; why AI gives different answers each time explains why and how to reduce it with fixed instructions and examples.

If you use a transcript

Some practices record the meeting and transcribe it. Agree this with the chair first, record the agreement and a deletion date for the recording, and use a business-tier tool whose terms you have read. A transcript helps you check details, but it makes the draft longer and tempts the assistant to include discussion. Keep the instruction "record decisions, not discussion" in the prompt, and use the transcript as a check against your notes, not as the primary input. The question of recording meetings is covered for lawyers in AI note-takers in client meetings, and its consent and deletion points carry across to board meetings.

The review and sign-off routine

The routine is what lets partners trust the output. Make it a short checklist attached to each job, with names and dates.

  1. Secretary checks every resolution against the notes, and every name, date and number against the source.
  2. Second person reviews anything marked [CHECK], and any resolution that affects the register, share capital, directors or banking.
  3. Draft goes to the chair with changes visible. The chair approves, and approved minutes are signed in the usual way.
  4. File the notes, the AI draft and the final minutes together, with initials and dates. Delete working copies from the AI tool on the schedule in your policy.

The tools should never touch the signed record. They help produce a draft; the final minutes are a human document, and the practice's file shows who approved them.

Worked example: a minute-and-pack routine for a 40-client practice

Say a practice of three secretaries services 40 private companies, each with quarterly board meetings: 160 meetings a year. Before AI, each set of minutes takes about 75 minutes to draft and 15 to review, 90 in all, or 240 hours a year. The new routine keeps notes as before (the same 20 minutes in the meeting), drafts in about 10 minutes using the house-style project, and review takes about 25 minutes because the secretary now checks more carefully. That is 35 minutes of drafting and review against 90, a saving of about 55 minutes a meeting or 147 hours a year, before counting the time cost of building the style projects.

Those figures are illustrative and the saving is not guaranteed; measure your own over ten meetings. The cost side is modest. ChatGPT Business is $25 a seat a month monthly or $20 billed annually, with a two-seat minimum; three seats on the annual rate is $720 a year. Add a few hours to set up each client's project and house style. If the practice charges a fixed fee per meeting, the saving is margin; if it bills time, decide how to share it with clients before the next fee review.

Five ways a secretarial AI routine goes wrong

  • Using personal accounts. A secretary pastes board material into a free account. Stop it with an approved-tool list and a simple approval step; the AI tool approval process shows one.
  • Letting the draft become the record. A rushed review approves wording nobody read. Use the [CHECK] flags and a minimum review time, and insist on a second pair of eyes for resolutions on directors, shares and banking.
  • Trusting a date the tool computed. Always compute deadlines in software or a checked formula, and confirm against the registry's own guidance.
  • Forgetting the client's rules. Some boards forbid AI or recording. Keep an opt-out flag on the client record and honour it.
  • Uncontrolled sharing. A shared project or link reaching the wrong person is a confidentiality breach inside your own walls. Check sharing settings when you set up each client project.

Turning minutes into an action tracker

Once the minutes are approved, the action table is the most useful by-product. Ask the assistant to extract every action from the approved minutes into a table (action, owner, due date, meeting date) and merge it into a running list per company. Before the next meeting, ask it to produce the "matters arising" list: actions still open, grouped by owner, with the date each was first recorded.

Illustratively, a board that met four times in a year might carry 23 actions, of which 17 are closed, four are open and two have no owner recorded. The two with no owner are the useful finding, because those are the actions nobody will do. Send the list to the chair a week before the meeting. Checking is simple: every row must point back to a numbered minute. Never let the assistant create an action that is not in the approved minutes, and review any row where the due date was inferred (for example "by the next meeting") by writing the actual date.

Secretaries whose clients sit under different legal systems

The workflow above is the same wherever you practise, but the content differs. A practice serving companies set up under different legal systems should keep one house-style project per system, with that system's terms for officers, resolutions and filings. Depending on the system, the statutory side of the job may centre on written consents in place of meetings, share or stock ledgers and annual meeting notices, or on statutory registers of members, charges and beneficial owners. Whichever applies, load your own templates and let the assistant adapt them, and rely on your own checked sources for the rules, not on the assistant's recall.

If you would like help mapping your own minute, pack and filing steps and choosing where to start, that is what my AI implementation consultation covers; start from the contact page. Otherwise begin with one client's tracker this week, and measure the number of gaps it finds before you touch minutes.

Setting up the first client project in an hour

To make the minute routine concrete, here is how a practice might set up a first client in about an hour. Create a project named for the company and add four items: three previous sets of approved minutes (names redacted if the client prefers), the constitution or articles in summary form with the quorum rule and the officers' titles, the house style sheet, and a list of standing agenda items. Add the prompt above as the project instruction, with a line stating the quorum so that the assistant can flag a meeting where the attendance list suggests the quorum was not met. Then test with a past meeting: feed the secretary's old notes, compare the draft with the signed minutes, and count differences.

In an illustrative test of a 40-minute meeting with six resolutions, a good result is a draft that matches the signed minutes in structure and resolution numbering, differs in phrasing in a handful of places, and raises three or four [CHECK] flags where the old notes were thin. If it invents anything, tighten the rules and re-run. When two consecutive tests are clean, use the routine on a live meeting, with the secretary's full review and the chair's approval exactly as before. Keep the first live use to one friendly client who understands what you are doing, and say so in advance.

The same set-up works for pack checks and for filing reminders: a project per client, a short instruction sheet, a test on a past example, and a named reviewer. Because each step is small and checkable, the practice builds trust in the tools through evidence, not through promises, and partners can see for themselves where the output is good enough and where it is not.

Questions company secretaries ask about AI

Can AI write board minutes?

It can draft them from your notes or an approved transcript, in your house style. The secretary still decides what is recorded, checks every resolution and name, and the chair approves the final minutes. Treat the output as a first draft that saves typing, not a record that is ready to sign.

Is it safe to put board papers into an AI tool?

Board material is usually confidential and sometimes price-sensitive. Use only a business-tier account with terms you have read, check the client's own rules on third-party tools, and redact where you can. If a client forbids AI, record that on the file and honour it.

Should AI track statutory filing deadlines?

Use it to build and tidy the tracker, not as the source of dates. Deadlines come from the law, the registry and the company's own year end. Calculate them in your secretarial software or a checked spreadsheet, and have a person confirm each date.

Do we need the board's consent to record a meeting for transcription?

Treat it as something to agree in advance, and record that agreement in the minutes or the board's procedures. Rules on recording differ and some boards prefer no recording at all. Ask your adviser, and agree a deletion date for the recording.

What is the right order to adopt this in a practice?

Start with the filing tracker, because the output is easy to check, then board pack assembly, then minute drafting. Minutes carry the most risk because the document becomes the legal record, so earn trust on lower-stakes tasks first.

Further reads

Sources: OpenAI's business pricing page (ChatGPT Business seats and shared projects), Anthropic's Claude pricing page, and the Claude help centre articles on projects and file uploads, checked 4 November 2026; deadlines and meeting details described are illustrative. Check your registry and governing documents for real dates and rules.

Want your secretarial routine mapped before you automate it?

On a 1:1 call we can walk through your minute, pack and filing steps, choose the first one worth automating, and set the review points so partners are comfortable signing.

Book a 1:1 call with me