Should Annual Report Agencies Use AI? Yes. Here's How to Do It Without Leaking the Numbers

Most annual report teams are already using AI. The question is whether they're using it openly, inside a process the client knows about, or quietly, in a browser tab nobody has approved. The second is where the risk lives.

Our view is simple. Agencies should use AI, and companies should let them. It makes the work faster, it can make the work better, and the audience for annual reports is changing in a way that rewards it. But it needs two hard limits: people own every number and every word, and confidential, unreleased content never goes into a consumer chatbot.

The short answer

Use AI for speed and for checking. Keep people on verification, judgement and feedback. And run anything confidential through a private, contract-covered setup (an enterprise plan, a paid API, or a model hosted in your own cloud account), never through a personal ChatGPT, Claude or Gemini account.

Why agencies and companies should be open to AI

It's faster, and annual reports are a deadline business

A large annual report is a four to five month project that ends in a crunch. Numbers arrive late, sections get rewritten after layout, and thousands of changes land between audited results and dispatch. Much of that time goes on work AI handles well:

• First drafts from approved inputs. A segment review drafted from the business head's interview notes, last year's section and the approved KPI sheet, ready for a writer to rework.
• Benchmarking. Reading thirty peer reports for how they handle capitals, risk or the BRSR, and summarising what the best ones do. These reports are public, so there's no confidentiality issue.
• Interview summaries. Turning a 90-minute recorded conversation with the CFO into a structured set of themes and quotes to verify.
• Format work. Alt text, captions, glossary entries, and first-pass translations for regional-language editions, all of which still need a human check.

None of this replaces the writer. It removes the blank page and the drudgery, so the writer's time goes on argument and clarity.

Quality can go up, not just speed

The checks that slip in the last fortnight are exactly the ones AI is good at. Does revenue in the chairman's message match the MD&A and the financial statements? Is it "net debt" on page 40 and "net borrowings" on page 212? Does every BRSR indicator the narrative mentions match the principle-wise table? Are the same five strategic priorities named the same way everywhere?

A model that reads the whole draft in one pass will flag inconsistencies that a tired proofreader on the fourth round won't. It can also check readability, find undefined abbreviations and hold the draft against the house style guide. Every flag still needs a person to decide what's right, but the net effect is a cleaner report.

Your next reader is probably a machine

Here's the part companies resist most, and shouldn't. More and more of the first reading of an annual report is done by software: analysts' models, data providers, ESG raters, and investors who ask an AI assistant to summarise the report before deciding whether to open it.

This isn't new, only faster. A 2020 NBER study of more than 400,000 US filings found that companies whose filings were downloaded more by machines went on to make them more machine-readable, and even avoided words that algorithms read as negative. Companies are already writing for software, whether they say so or not.

So the "it reads like AI wrote it" worry is aimed at the wrong thing. What matters is whether the content is accurate, clear and consistent, because that's what both a human reader and a machine reader reward. Plain sentences, defined terms, numbers in the text that match the tables, and a logical structure help an analyst's model as much as they help a shareholder. A report written that way, with AI's help or without it, will be summarised correctly. A vague one will be summarised vaguely.

What AI must not do

AI drafts and checks. People decide. In an annual report, that line matters more than in almost any other document, because every page carries the company's name and many carry legal weight.

Two rows deserve emphasis. Numbers: language models still make up figures with total confidence, and a single wrong number in an annual report is a correction, a filing and a credibility problem. Every number must be checked against the audited source by a person, no matter how it was produced. Feedback: when a CFO writes "this doesn't sound like us" or a board member questions a claim, the response needs context, judgement and a relationship. Feeding client feedback into a chatbot and pasting back what it says is how agencies lose clients.

The confidentiality problem nobody talks about

Before an annual report is published, much of its content is confidential. Some of it is more than that. Under SEBI's Prohibition of Insider Trading (PIT) Regulations, information that isn't generally available and is likely to move the share price, such as financial results, is unpublished price sensitive information (UPSI). A draft annual report prepared before the results are out can contain plenty of it.

The PIT Regulations don't stop a company from sharing UPSI with its agency. Sharing with advisors and consultants for a legitimate purpose is allowed. But it comes with conditions: the board's policy defines what counts as a legitimate purpose, recipients are treated as insiders and must be given notice to keep the information confidential, and the company logs who received what in its structured digital database.

Now picture a writer pasting the draft MD&A into a personal chatbot account to tidy it up. The information has gone to a third party that isn't on the database, under consumer terms the company never agreed to. Whether that's a regulatory breach depends on the facts, but it's hard to square with the client's UPSI policy and almost certainly breaches the agency's confidentiality agreement.

Consumer terms matter here. On ChatGPT Free, Plus and Pro, conversations can be used to train OpenAI's models unless the user opts out. Since August 2025, consumer Claude users choose whether their chats are used for training, with five-year retention if they agree. Google says content sent to the free tier of the Gemini API can be used to improve its products and may be read by human reviewers. None of these is a place for an unpublished annual report.

Is an internal RAG tool on an AI company's API the answer?

Mostly yes, with three caveats that teams often miss.

Business products are covered by different terms. OpenAI's business and API data isn't used to train its models by default; API inputs and outputs are kept for up to 30 days for abuse monitoring, and zero data retention is available on approval. Anthropic's commercial terms bar training on customer content, and its consumer changes explicitly exclude the API, its business plans and access through Amazon Bedrock and Google Cloud. Google doesn't use paid Gemini API traffic to improve its products.

Going through a cloud provider adds another layer. On Microsoft's Azure, prompts and completions aren't available to OpenAI, aren't used to train models, and are processed in the geography you choose (unless you pick a "Global" or "Data zone" deployment). Amazon Bedrock doesn't share prompts or outputs with model providers either. For a company that already runs on one of these clouds, this is often the simplest route to something its IT and compliance teams will approve.

Now the caveats.

RAG doesn't keep data private by itself. Retrieval-augmented generation keeps your documents in your own store and pulls out the relevant passages when someone asks a question. Those passages are then sent to the language model with the question. The confidentiality comes from the model's terms, not from the RAG design, so the endpoint still has to be one of the business options above.

Agencies must separate clients. An agency holds several listed companies' UPSI in the same season. One shared index means one wrong query can surface another client's numbers. Each client needs its own store, and retrieval should respect the same access rights as the project folders: only the people on that account can query it.

The tool makes copies. A vector database, prompt logs and cached files are all copies of confidential content. Encrypt them, restrict access to them, and delete them on a schedule, for example once the report is published and the client's retention period ends.

Build the workflow before buying the tool

The firms that use AI well rarely start with technology. They start with a short, written workflow that the whole team follows and the client has seen. It covers six things.

1. Classify the content. Public (peer reports, published filings), confidential (drafts, strategy), and UPSI (unpublished results and anything price-sensitive). Each class gets a list of tools it may enter.
2. Approve the tools. Name the specific enterprise accounts or API-based tools the team may use, and ban personal accounts for client work. A ban without an approved alternative just drives the use underground.
3. Put it in the contract. Tell clients how AI is used, which providers process their data and under what terms, and get their written agreement. Some will want to approve the provider; that's reasonable.
4. Write the prompts down. A shared library of tested prompts for the recurring tasks, tied to the house style guide, gives consistent output and stops everyone improvising.
5. Set review gates. Name who checks AI-assisted drafts, and require a second person on every number. Mark AI-assisted sections in the working file so reviewers know where to look harder.
6. Keep a log. Record which tool touched which document. If the client's compliance team ever asks, the answer takes minutes, not weeks.

Notice what isn't on the list: using AI for everything. Feedback calls, creative direction, management interviews and the final read stay human, because that's where the agency's judgement is, and judgement is what the client is paying for.

What companies should ask their agency

If you're hiring an annual report agency this season, add these questions to the brief:

• Do you use AI on client work, and for which tasks?
• Which tools and providers, under which plans? Do any of them train on our content?
• Where is our data processed and stored, and for how long?
• Is our content kept separate from other clients' in any AI tool you use?
• Who checks AI-assisted drafts, and who signs off the numbers?
• Will you sign a confidentiality undertaking that covers AI tools, so we can record you correctly in our structured digital database?

An agency that answers these clearly is one you can trust with AI. An agency that says it doesn't use AI at all is either behind or not telling you.

Our own approach

We use AI ourselves. It helps us research, draft and check. People review every piece and approve every word before it's published, and we only work from published information. It's the same principle we'd recommend to any agency: let the machine do the reading and the first draft, and let people do the judging.

Companies that insist on no AI at all will pay more and wait longer for reports that aren't better. Companies that allow it without rules will eventually see a draft where it shouldn't be. The sensible position sits between the two: open to AI, strict about where the data goes, and clear that a person signs off every page.