AI for Comms Teams: Where It Helps, Where It Hurts, and How to Use It Well
AI isn’t replacing communications teams. It’s just making the wrong workflows slower. The gap between burning time and moving fast comes down to one thing: treating AI like a tool, not a fix. That means knowing where to plug it in, where to keep humans firmly in control, and which common shortcuts quietly drain your team’s efficiency.
How to Chat With AI
Value Your Judgement Over AI’s
From a comms perspective, your most important takeaway is this: AI systems remain unreliable at maintaining judgment across ambiguous, multi-step tasks where context, prioritization, and domain expertise matter. Throughout our testing, we’ve found it struggles with longer, information-dense tasks. It tends to drift off course and miss key insights without a human in the loop.
Large Language Models (LLMs) are stateless, meaning each chat starts as a clean slate. Models rely on their training data and your inputs, they can’t retain new information on their own. Most "memory" features you see work by adding past details back into the prompt.
Because LLMs are stateless, they resend your chat history with every message. The more that history grows, the more the model has to hold at once. There is a finite amount of information an LLM can process at once. That’s your context window. As the chat grows larger and larger, models will start losing threads and missing details; sometimes called context rot. While modern context windows can be quite large, it’s important to keep this in mind.It’s also important to know that cloud apps like ChatGPT, Claude and Gemini have methods for managing ballooning context windows, although this comes with caveats. They often truncate or summarize long chats to preserve context, then use that compressed version to continue the thread instead of failing outright. This is important to keep in mind when using cloud-based models because key details can be lost during the compression process, and that loss can compound as the thread continues, causing it to slowly lose the plot. All this to say, the bigger and messier the conversation gets, the worse problem-solving and inference tend to become.
AI isn’t a second employee. It can’t hold information and nuanced threads like a human can, and trying to treat it as a collaborator will lead to suboptimal performance. Once you treat it as a high-speed, low-judgment tool, you can build effective workflows.
Set Your System Up for Success
The quality of your output is directly tied to the level of supervision you provide. Here is how to set up your system:
While every cloud LLM runs on a hidden system prompt you can't see or change, all of them let you add your own standing instructions on top of it in the application’s settings.
A strong set of custom instructions can soften two major weaknesses, sycophancy and hallucination. This curbs the former by encouraging critical evaluation rather than automatic agreement, and the latter by prompting the model to admit uncertainty and avoid unsupported claims.
Adjust this over time, but start with a baseline that forces critical thinking. From our own internal testing, we’ve noted that instructing the AI to ground any factual claims, cite its sources, and avoid placating the user noticeably improves the output.
Example Base Prompt:
"Don’t tell me what I want to hear. Always ground factual claims empirically with reputable sources. If you are unsure, state that clearly."
Claude, ChatGPT, and Gemini all offer custom instructions that influence their outputs. These settings can be found and adjusted through each platform’s settings menu:
Never Sacrifice Clients’ Security
Never put identifying client data into public LLMs. Anything uploaded to a cloud AI service may be processed outside your direct control depending on your contract and settings. While enterprise plans offer stronger privacy guarantees, it is still inadvisable to upload any data that hasn’t been thoroughly scrubbed of any identifiable client or proprietary information. Once the data is off premises, you are relying on someone else’s controls other than your own.
If your data contains personally identifiable information, confidential strategy, or proprietary information, keep it offline or use a local model. The risk is not worth the convenience.
What to keep off the cloud:
Personally identifiable information: yours or other people's (names, addresses, phone numbers, SSNs, DOBs)
Credentials and secrets: passwords, API keys, access tokens, private keys
Confidential strategy: unreleased plans, positioning, internal roadmaps
Financial data: unreleased earnings, bank/account details, internal projections
Legal material: privileged communications, sealed documents, active litigation details
Anything under an NDA that hasn't been expressly cleared for third-party service providers
Source code or trade secrets you don't own or aren't cleared to share
Health or medical records: yours or anyone else's (HIPAA can apply here, depending on your client relationship)
Anything else that isn't yours to share, or that you'd hesitate to make public
Keep Your Chat on Track
AI has a strong tendency to drift off the original premise in longer interactions. It will creep out of scope, offering suggestions that sound good but derail your task.
Keep chats short: LLMs get less reliable as the context window increases. The model must process the entire chat history (or chunks of it) every time you send a new message. This leads to "context rot" for longer chats.
Reset frequently: Don’t let one long thread become a mess of tangents. Start fresh for distinct tasks.
Keep your goal in mind: Always keep in mind your actual goals for your task. The LLM doesn’t have your context, it is guessing, and usually guesses poorly in ways that create extra work or branching threads. Be clear and specific about what you want it to do.
Trim the Fat Yourself
You might think you can fix AI’s tendency to ramble by adding constraints like "Be concise" or "Max 10 words." Don’t.
Because LLMs predict the next token (a small unit of text that a large language model processes) based on the previous one, they tend to struggle with brevity. Consecutive tokens are based on probability distributions that favor complete, safe, and expansive sequences. When you force extreme brevity, the model often performs poorly, dropping nuance and essential context.
The solution to this? Don’t use AI to write your final punchy copy (It can’t distinguish between "fluff" and "essential nuance"). Use it for raw drafts, then let a human trim the fat.
Always Double-Check Sources
AI in 2026 can competently source and structure information, but it’s not always good at reading and synthesizing it properly.
LLMs will often miss context on a source, leading to an incorrect inference on your question, or misattribute information entirely. It presents these errors with high confidence because the syntax is correct, even if the facts are wrong.
The Rule: Always check its sources for factual claims you plan to use downstream. Treat AI research as a starting point or an annotated bibliography, not a finished analysis. You must verify every claim against the primary source before it goes into any client deliverable.
Choosing the Right Tasks for AI
When to Use AI
Look for projects that fall into repetitive, easily identifiable patterns. AI shines here where human judgment is less critical, and you can accept that some portion of the final product may be wrong, require reworking, or need additional human attention.
Starting Research: Models like Claude can generate beautiful initial research documents quickly. However, they miss leads and lack depth. It’s effective at generating the structure for your research task, though you will need to fill out what it misses.
Web Research Synthesis: Mature models often pull better, more synthesized results than a raw Google search for broad topics.
Rubberducking (The Sounding Board): Use AI to challenge your assumptions. If you use it this way, tune your system prompt to force it to critique your ideas rather than agree with them. Sycophancy is a major problem; you need a devil’s advocate, not a yes-man.
Document Generation: Use AI to help turn a messy brain dump into a structured document. If your information is present and accurate yet disorganized, a thoroughly grounded model/prompt can be effective at organizing it. It can also generate spreadsheets and PDFs that follow a strict template. (Note: For purely repetitive data entry, a simple script is still faster and more reliable)
When Not to Use AI
These are areas where AI introduces more risk than value for comms teams.
Planning from Scratch: AI is not nearly as good at initial brainstorming as talking to your colleagues. In a small team, the collective brain of people who know the client intimately is faster and higher quality than prompting an LLM to generate mediocre ideas that you then have to filter. Use AI for variation or validation research on a seed idea, not for generating the seed itself.
Outsourcing Unfamiliar Tasks: If you don’t know what "good" looks like, you can’t verify the AI’s output. Outsourcing tasks you don’t understand is a massive risk for your clients because AI often makes confident, yet incorrect, statements. It’s particularly insidious when it misrepresents or misattributes its sources.
Client Data with Personally Identifiable Information: Anything involving client data that can’t be stripped of identifying information. Big privacy concerns. Enterprise plans offer stronger privacy, though still shouldn’t be trusted with client data or sensitive internal strategy.
Judgment Calls and Long, Multi-Step Tasks: If you need AI to do a big multi-step task, it requires hand-holding and a hypercritical eye to prevent drift and scope creep.
Unified Brand Voice: AI will drift without supervision or a heavily bounded task architecture. It usually won’t nail a specific brand voice consistently because it lacks the human ability to understand why a certain tone works. It can help you draft ideas, but the final polish must be human.
AI Art: Avoid AI to generate art for your clients. Even if the image is technically good, and even if the client doesn’t care, the public at large hates AI art.
Realistically, most people aren't going to follow every AI best practice all the time. Sometimes you're on a deadline. Sometimes you need a starting point. Sometimes AI is just the fastest way to get a task going. The throughline is knowing when convenience is worth the risk.
A good rule of thumb is to ask yourself: What happens if this is wrong?
If the answer is "not much," AI can probably help you save some time. If the answer is "this could hurt a client relationship, expose sensitive information, or create a bigger mess later," that's a sign you should slow down and rely on human judgment. AI can be a useful tool for comms teams, but it's still just a tool. Use it to help you move faster, not to outsource your cognition.