
In corporate communications, the first useful thing an AI assistant does is not write a finished announcement. It gives you a place to begin. It can turn a long technical brief into an organised list of claims, flag missing stakeholders, or reconstruct a scattered email chain into a timeline that people can check.
That speed is useful. But a clean summary can still be wrong in the way that matters most: it can make an exploratory conversation sound like an approved contract, an old quote sound current, or a possible roadmap sound like a commitment. The prose may be fluent while the status is false.
I treat generative models as drafting and triage tools, never as the source of record. Before a sentence leaves my desk, I want clear answers to three questions: What exactly are we claiming? What source supports it? What is its legal and commercial status?
Start with structure, not a paragraph
A communications brief is rarely a clean set of approved facts. It may combine verified information, market commentary, open decisions, and aspirational language. If you ask a model to write an announcement immediately, it may smooth those categories into one convincing paragraph. Uncertainty disappears from the prose even when it has not disappeared from the underlying evidence.
I get a better first pass by asking the model to act as a parser. It extracts each assertion into a table with the claim, its supporting document, the document’s author or owner, and whether written clearance exists. That changes the task from “make this sound exciting” to “show me what we can prove.”
The model can help identify missing sources, inconsistent dates, and repetitive language. A person still has to decide whether a claim is accurate, approved, and appropriate for public release. A polished draft can create pressure to send; a visible list of open checks makes it easier to pause.
Also Read: Open code, unowned risk: The governance gap in business-critical email
Audit the verbs of commitment
In corporate communications, verbs such as “launched,” “partnered,” “selected,” “acquired,” and “secured” carry weight. They describe different states of affairs. A model that treats them as stylistic variations can turn an exploratory discussion into an announcement of a completed relationship.
An AI-assisted draft can draw on genuine historical correspondence and recast it as a current partnership. The source material may exist; the implication may not. The fix is not another prompt. It is to return to the primary correspondence, check the relevant agreement and approvals, and remove the unsupported assertion.
That discipline applies to Votee AI’s own communications too. The HKMA’s published list names Votee AI among the technology partners selected for the first GenA.I. Sandbox++ cohort; the SFC’s published annex records the same cohort. Similarly, our CTO is listed as a contributor to OWASP’s Agentic Skills Top 10 (AST10), a framework documenting security risks in AI agent skills. “Selected” and “contributed” are different claims, and each should match exactly what the public record shows.
The governance question for agents
The industry is moving from scripted chatbots towards systems that can plan and use tools across multiple steps. That can make software more capable, but it also makes permissions, oversight, and accountability harder to reason about.
An agent with access to company data should not be treated as safe simply because its output sounds plausible. Systems need limits on what an agent can access and do, clear records of the actions it takes, and a way for people to review or stop consequential steps. A practical “maker-checker” approach can help: one component gathers information or prepares a draft, while a separate control process checks the action against defined rules and routes higher-risk decisions to a person.
Also Read: AI is a tool. Stop treating it like a strategy
Review should scale with risk. An internal recap email does not need the same scrutiny as a public statement, a regulatory filing, or an earnings disclosure. But each process should make clear which facts are verified, who approved them, and what remains uncertain.
Use saved time to check the record
The useful promise of AI in communications is not that it removes human judgement. It is that it can reduce the manual work of sorting, collating, and summarising information from different documents.
The important question is what we do with the time saved. I use it to trace claims to their sources, ask subject-matter colleagues what a document actually authorises, and check whether a draft reflects the decision that was made, not merely the decision someone hopes will be made.
A strong draft is not the one that sounds most certain. It is the one whose confidence matches the evidence behind it. In communications, the source check is not a final proofreading step. It is part of the work.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post The AI draft is the easy part. The source check is the job appeared first on e27.
