# Why I Stopped Using ChatGPT for Consultancy Work > Why context amnesia made ChatGPT the wrong fit for parallel client work, and how Claude projects, two-way artefacts, RAG, and skills changed how the team delivers. **Type:** Breakdown · **Read time:** 9 min · **For:** Managing Partner, COO · **Published:** 5 Jun 2026 **Video companion:** https://www.youtube.com/watch?v=kMGXw9Hij0k [All resources](https://aigenticlab.com/resources) · [View as HTML](https://aigenticlab.com/resources/why-i-stopped-using-chatgpt-for-consultancy-work) --- ### Two years in ## The problem is not capability For two years I told consultancies to use ChatGPT. I built workshops around it, recorded videos about it, and made it the default. Last month I stopped using it as my primary AI tool. This is what changed, and why it matters for any firm running multiple client engagements at once. If you run a 20 to 50-person consultancy, you already know these tools are capable. The demos are impressive and the benchmarks are fine. The problem is not what AI can do in isolation — it is what happens when AI becomes part of real client delivery. Running engagements in parallel means constantly switching between different standards, tones, industries, and definitions of "good." In professional services that consistency is non-negotiable: the tone you use with clients matters, the structure of your deliverables matters, the format of everything you produce matters. That is exactly where the industry's most popular AI tool quietly breaks down. ### The silent drain ## Context amnesia across engagements We tried to migrate to ChatGPT Teams profiles twice. Both times the team reported that individual personal profiles were more productive than the team environment. That should not happen — it is a signal the tool is not designed for how consultancies actually work. The core issue is what I call **context amnesia**. Every new project starts from zero. You are constantly reintroducing the same information: who you are, what you are working on, your quality standards, what good looks like in your specific context. That is not AI working for your team — that is your team working around the AI. ChatGPT's memory feature was supposed to solve this. It works well enough with a single use case, a single role, and one consistent thread. But across parallel engagements it blends context between clients — you ask a question in one project and it surfaces information from another, Beehiiv newsletter settings showing up inside an unrelated client project. You can now scope memory to a specific project, which should contain this, but that scoping only holds while you are actively inside that project. Step outside it for a quick one-off question and the context bleeding resumes. > You end up spending more time policing the AI's memory than actually using it. ### How the work changed ## Persistent workspaces and two-way artefacts Claude projects function as persistent, dedicated workspaces. Each one keeps its own context, stores uploaded assets, retains custom instructions, and holds a separate memory layer you can inspect and manage directly. When you switch between clients, that context does not drift, blur, or contaminate other workspaces. That alone would justify the switch — but the difference that matters most operationally is how Claude handles artefacts. In ChatGPT, artefacts work in one direction. Content is generated and stays locked inside the conversation. You can download files now, but the workflow still makes you the middleman: download, open, edit manually, re-upload. In Claude, artefacts work in both directions — you open a file inside the interface, inspect it, ask for specific edits, and save the updated version back into the project. A concrete example: in a newsletter project I had a template file loaded and asked Claude to add an introduction section. It edited the file, returned the updated version, and I saved it straight back. The same prompt in ChatGPT produced the introduction as text in the chat, which meant downloading the original file, opening it, locating the right position, and pasting it in manually. More steps, more friction, more opportunity for error. For a consultancy managing multiple engagements, that compounds: you are not starting from zero every time, you are building on what already exists. ### Knowledge that scales ## RAG limits and the KWA foundation Claude retrieves information from project files using RAG (Retrieval-Augmented Generation), and its limits work differently from ChatGPT's. You are constrained by **file size, not file count**. In ChatGPT I was regularly merging smaller files together just to stay within the project's file-number limit — a pointless administrative task with no output value. In Claude you upload your proposal templates, brand guidelines, past deliverables, and methodology documents, and Claude uses them when relevant. Nothing needs to be merged or condensed to fit an arbitrary constraint. This is what the Knowledge layer of a well-designed AI workflow should look like: institutional knowledge that is centralised, accessible, and retrievable. It is the first stage of the KWA framework — Knowledge, Workflow, Agents — that we use to design AI implementations for professional services firms. Most firms stall here because their documents are scattered across drives, inboxes, and individual laptops. Claude projects give you a practical place to start centralising that knowledge without a major systems overhaul. ### Reusable capability ## Skills and the integrations that connect your knowledge Skills are where the efficiency compounds most for teams. They are reusable instruction sets that shape how Claude handles specific tasks. Rather than rebuilding context at the start of every conversation, you configure a skill once and it applies consistently across every chat: - **Sales discovery notes** — defined once, available to the whole team - **Internal communications** — a consistent house style every time - **Research synthesis** — the same method applied across engagements - **Proposal structure** — your firm's format, not a generic one What makes this powerful for consultancies is that the same skills are available behind the API. Connect Claude to your automation layer through n8n or similar and the same institutional knowledge applies — you are building reusable capability that works whether someone uses Claude directly or through an integrated workflow. The current limitation is worth naming honestly: sharing skills across a team still requires each member to upload and manage their own copy. Admins can enable or disable custom skills at the firm level, but there is no internal marketplace yet where approved skills reach all users automatically. That is a gap the business plan should address — but even now the gain for individual workflows is significant. Claude projects also let you enable tools at the project level — web search, extended thinking, and external integrations including Notion and Google Drive — a level of control ChatGPT's project environment does not come close to. This matters because most firms already store their knowledge somewhere: meeting notes in Notion, documents in Google Drive or SharePoint. In practice you can ask Claude to research a content topic, draft a script using your internal ICP and audience documentation, and save the output directly to the right Notion database with the correct properties applied. No copy-paste, no reformatting. Claude is not just generating content — it is building and extending your knowledge base while the work happens. ### The honest scorecard ## Where ChatGPT still wins This is not a clean sweep, and it is worth being direct about where ChatGPT remains stronger. - **Image generation** is considerably better in ChatGPT. We use Gemini for image editing and Sora for visual ideation, but if you need generation inside a single tool, ChatGPT holds the advantage. - **Voice** is still better on ChatGPT. I use it on mobile to talk through a problem like a verbal assistant. Claude has improved recently but is not yet at the same level. - **Brand recognition** matters in client-facing contexts. Most Managing Partners and COOs have heard of ChatGPT and likely used it. Claude is less visible to that audience, which occasionally matters when recommending tools to a client team. I still hold an active ChatGPT subscription. But Claude is where the work happens for me and my team now. ### The bottom line ## What this means for your firm The real cost of the wrong AI tool is not the subscription fee. It is the hours your team loses to reprompting, reformatting, and not trusting the output. That hits margin, utilisation, and the partner economics Managing Partners feel every week — even if they never label it as an AI problem. If you run a 20 to 50-person firm, you do not need a comprehensive AI strategy before you start. You need one AI tool that fits how your team actually works: one that maintains context across engagements, scales with your knowledge base, and reduces friction instead of adding it. That is the difference between AI your team tolerates and AI your team relies on. The 5 Steps for AI Leadership framework starts with Align and Activate for good reason. A tool your team does not trust will never get past the Activate stage, regardless of what it is theoretically capable of. Getting the tooling right is a prerequisite for everything that follows.