> ## Content Index > Fetch the complete content index at: https://www.appliedcomms.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Five layers of AI tooling, and the comms job each one is for - URL: https://www.appliedcomms.ai/ai-tool-layers-for-comms/ - Published: 2026-08-18T06:00:20.000Z - Updated: 2026-08-18T06:00:20.000Z - Description: The problem is not that comms professionals are using AI badly. It is that most of them are using one layer of it and assuming that is all there is. - Author: Michael MacLennan - Tags: Guides Most communications teams have found one way of working with AI and stopped there. They open a chat window, type a request, copy the answer out. It works, so it sticks. The chat window is real and useful. It is also one layer of a deeper stack. The jobs that change how a comms team actually operates, the ones that give you back hours rather than minutes, mostly sit on the layers above it. The problem is not that comms professionals are using AI badly. It is that most of them are using one layer of it and assuming that is all there is. Here are the five layers, what each is for, and where each one stops being the right tool. The examples under each are jobs I have run myself. ## **Layer 1: The free chat window** This is the fresh conversation: the likes of ChatGPT, Claude, Gemini and Copilot as we knew them a couple of years ago. One thread, one task. You ask, it answers, you take the answer somewhere else. The comms job it suits: quick drafts, rewrites, summaries, sense-checks, a thinking partner for a tricky paragraph. Most tools have now bolted on memory and other layer-two features (see below), but there is still something to be said for a chat that starts entirely fresh, without any baggage. For example, last month, before a 30-minute call with a founder I was advising, I used a chat window to pressure-test how they might frame their offer. I drafted a handful of questions to ask, then tried two or three ways of saying the same thing out loud until one sounded right, and it became the jumping-off point for exactly where we needed to get to. Twenty minutes, disposable, gone by the afternoon. That is layer one doing exactly what it is good at. Where it stops: this kind of chat window (still very much available if you choose not to log in to a tool such as ChatGPT) has no memory between conversations and no access to your material. Every session starts cold, so you re-explain your brand, your tone and your context every time. That re-explaining is the tax you pay for staying on layer one. It leads to lengthy, detailed explanatory prompts that are about as fun to write as they are to read back, and for a lot of teams it quietly cancels out the time the tool saved. And even if you do log in, so that there is a general memory and the ability to search through past conversations, the results are often scattergun and lead to a bumpy, frustrating experience. It is like working with a distracted colleague who almost remembers your name and may recall what you are referring to, but might also rummage through your shared history and emerge with something wholly inappropriate. The results are more likely to be worth it than with no memory at all, but nowadays we can do far better. And on that note. ## **Layer 2: The configured workspace** A workspace with memory: Claude Projects, ChatGPT Projects and custom GPTs, Google's Gemini Gems, and similar. You load it once with your knowledge, your tone of voice, your house style and your standing instructions, and every conversation inside it starts pre-briefed. The comms job it suits: anything recurring. A client account, a campaign, a publication, a named executive whose voice you write in. I loaded a workspace once with a brand's guidelines, voice and visual identity, then had it produce [a complete new-business pitch suite](https://www.appliedcomms.ai/claude-skills-projects-business-pitch/) in a single sitting: a speaking proposal, a twelve-slide deck, a twelve-week content calendar and the outreach email to go with it. Ninety minutes end to end, most of it spent waiting rather than working. The writing came back around ninety per cent of the way there; the visual design needed real manual effort and was the weakest part. The point of the layer holds either way: the context was done once, not re-explained for every asset, which is what made tailoring the whole suite worth the time at all. Where it stops: it is still a conversation. It will give you a better reply, in your voice, but you still take that reply somewhere else and do the next step yourself. ## **Layer 3: The 'desktop' agent** This is where I have found things particularly interesting in 2026, and where I have [personally spent most of my time](https://www.appliedcomms.ai/claude-cowork-review/) over the past few months. AI that works alongside you with access to your files and apps: Claude Cowork, which moved from research preview to general availability in April 2026 and is now rolling out [beyond the desktop to mobile and web](https://www.wired.com/story/shut-those-laptops-anthropic-puts-its-claude-cowork-agent-on-your-phone/), is the clearest current example. It reads your documents, drafts, saves, and runs multi-step jobs rather than just answering questions. Tellingly for our field, Anthropic's own usage data suggests [more than 90 per cent of Cowork use is not software work](https://venturebeat.com/technology/anthropic-brings-claude-cowork-to-mobile-and-web-as-usage-data-shows-most-users-arent-coding), which is exactly why comms teams should be paying attention. The comms job it suits: the whole task, not just the thinking part. I pointed a desktop agent at a directory of forty-odd documents I had been avoiding, gave it a single positioning note, and let it read every file, build a folder structure from what it learned, move everything into place and then draft a fifteen-section playbook off the result. It filed everything correctly, with no errors or hallucinations, and, importantly, it asked before it acted. I would firmly recommend making sure it shares a plan for you to approve before you ever let it move, delete or fundamentally alter a document. The shift was less about the AI writing better and more about it clearing the friction that had stopped me starting at all. Where it stops: it acts on your behalf, so it needs supervision. This is the layer where governance starts to matter in a way it does not at layer one. An agent that can save and send is an agent that can save and send the wrong thing; and though the best-in-class models have grown more reliable, you still want to be paying attention and leading the way. ## **Layer 4: Building the tool** Making the thing you wished existed: Claude Code, and the various app builders. Not using a tool, but producing one. A comms professional with no coding background can now build a working internal tool in an afternoon. That is more or less what I did with [Comms With AI](https://www.appliedcomms.ai/comms-with-ai-claude-code-build/), then poured in far more time once I realised the idea had [award-winning legs](https://www.commswith.ai/updates/ai-comms-awards-2026-result/). The comms job it suits: a repeated need that no off-the-shelf product quite serves. The Comms With AI [template library](https://www.commswith.ai/library/) was built by a communicator rather than a developer. Studying machine learning and AI last year gave me some very limited coding skills, but nothing that would ever have let me create something like this. A [stakeholder-review checker](https://www.appliedcomms.ai/ai-stakeholder-comms-review-tool/) and a [LinkedIn content tool](https://www.appliedcomms.ai/claude-artifact-app-builder-linkedin-content-creator/) came the same way, each built to scratch a specific itch. Where it stops: building is the easy part now; deciding what is worth building, and maintaining it once it exists, is not. A tool nobody owns becomes a liability. There is a reason people are still paying their subscriptions for the leading software solutions, which are regularly updated and less likely to fall over. This layer rewards judgement about which problems are worth solving in software and which are not, and that judgement is now the scarce thing, not the code. ## **Layer 5: The autonomous agent** You hand over a brief and the tools to do it, and the agent runs to completion and reports back, often after hours rather than minutes. Long-running, mostly unsupervised. With each new frontier model the number of hours it can run unsupervised climbs; a recent Claude Sonnet has been reported to sustain focus on a single task for [30 hours straight](https://www.anthropic.com/news/claude-sonnet-4-5). The comms job it suits: large, well-defined, low-ambiguity jobs you do not want to babysit. Across Claude and ChatGPT I now have a batch of scheduled activity that runs on the dot each week. A scheduled agent sweeps Comms With AI every Friday morning: it checks the copy and any new pages against house style, [stress-tests key pages against three reader personas](https://www.appliedcomms.ai/icp-agent-ai-target-reader/), and returns a ranked list of what to fix and why. I set the brief once; it runs on a schedule and reports back, and I decide what to act on. Work like monthly analytics reporting or a bulk content audit fits the same shape, because the brief can be specified tightly in advance. Where it stops: this is the real frontier, and comms agencies are mostly not at this point yet, though that is changing quickly, and it pays to experiment and find where scheduled autonomous tasks can make a real difference. That weekly sweep works precisely because the brief is tight and the domain is narrow. Loosen either and a human comes straight back into the loop. The further a job sits from a tightly specified brief, and communications work usually sits some distance from one, the more supervision it needs. Treat layer five as real but early. ## **The actual skill: matching the job to the layer** The mistake is not picking the wrong tool. It is not realising there are layers at all, and so trying to do a layer three job in a layer one window, or reaching for layer four when a configured workspace would have done. Two rules of thumb. First, the value climbs as you go up the stack, and so does the need for judgement and governance. A team moving up the layers should be having the AI-policy conversation in step with the move, not after it. Second, do not climb for its own sake. Plenty of good comms work lives happily on layers one and two. The point is to climb deliberately, when a job calls for it, rather than staying on layer one because it is the only layer you know is there. The teams that get real value from AI are not the ones using the cleverest tool. They are the ones who can look at a piece of work and know which layer it belongs on. --- ## **Related reading** - [AI Agent-Driven Communications: A Practical Framework](https://www.appliedcomms.ai/ai-agent-driven-communications-practical-framework/). The five-phase Comms With AI Operating System that sits alongside these five layers. - [How Project Folders Supercharged My AI Comms Workflow](https://www.appliedcomms.ai/project-folders-ai-comms-workflow/). Organising the workspace that layers 2 and 3 depend on. Putting this into practice: [Comms With AI training](https://www.commswith.ai/training/) helps teams move up the layers, and [Deploy Comms With AI](https://www.commswith.ai/deploy/) governs the move.