Every week there’s a new AI tool with a product hunt launch, a Twitter thread claiming it’ll “replace your entire marketing team,” and a waitlist with 50,000 people on it.
We’ve tested a lot of them. Most are interesting. A handful are genuinely transformative. Many are wrappers around GPT-4 with a nicer UI and a $49/month price tag.
This is an honest account of what we actually use, what we’ve retired, and why — with no affiliate arrangements or sponsored placements.
The Core Stack
Claude (Anthropic) — Daily Driver for Writing and Reasoning
We use Claude for the majority of our text-based work: drafting client communications, writing long-form content, analysing documents, code review, and complex reasoning tasks.
What sets it apart: the context window is large enough to paste entire client briefs or codebases, the instruction-following is precise, and the writing output has noticeably less “AI voice” than most alternatives — which matters enormously when you’re producing content for clients.
For anything requiring careful reasoning, nuanced writing, or working with large documents, Claude is our first choice.
ChatGPT (OpenAI) — Rapid Iteration and Brainstorming
We keep ChatGPT in the stack primarily for brainstorming sessions and rapid first-draft iteration. The interface is familiar to most team members and the turnaround speed for quick tasks is excellent.
Custom GPTs have also proven genuinely useful — we’ve built internal GPTs pre-loaded with our brand guidelines, client context, and style preferences for specific content types.
Make (Integromat) — The Automation Backbone
If Claude is our thinking layer, Make is our plumbing. We use it to connect tools, automate repetitive workflows, and build the pipelines that power our client automation projects.
Make’s visual flow builder makes complex multi-step automations comprehensible — you can look at a flow and understand what it does. The OpenAI and Anthropic API modules integrate cleanly, making it straightforward to inject AI into any workflow step.
We’ve built client automations in Zapier too, and while it’s more accessible for non-technical users, Make handles complexity and conditional logic better at scale.
Perplexity — Research and Current Information
For research tasks requiring current information — market data, recent news, competitor research, industry trends — Perplexity is more reliable than asking a base LLM. The cited sources also make it practical for fact-checking.
We use it as a research starting point, not an endpoint. It’s excellent for getting oriented in a topic quickly and identifying primary sources to investigate further.
Cursor — AI-Assisted Development
For our development work, Cursor has meaningfully changed the speed at which we write code. The ability to have a conversation about a codebase — “why is this component re-rendering on every keystroke?” — and get contextually aware answers is genuinely different from asking ChatGPT with pasted snippets.
Tab completion for code has also become something team members report feeling dependent on in a relatively short time.
Descript — Video and Podcast Editing
For any project involving video or audio content, Descript’s AI editing is remarkable. Editing by transcript rather than waveform, removing filler words automatically, and the overdub feature for minor corrections have cut our video production time significantly.
Not a tool we use on every project, but on projects where we do, the time savings are substantial.
The Ones We Tried and Retired
Jasper
We gave Jasper a genuine trial for content production. The template library is extensive and the team-focused features are well-designed. But the writing output felt template-bound — recognisably AI, competent but generic.
For the subscription cost relative to what Claude produces with a good prompt, it didn’t make sense to keep.
Midjourney
We still use Midjourney selectively for concept generation and moodboarding in brand projects. It’s powerful for visual ideation.
We retired it as a production tool after too many client situations where AI-generated imagery caused complications — difficult to source-verify, style inconsistency across a project, and client discomfort with using AI-generated visuals in final deliverables.
It remains useful as an ideation tool, not a production tool.
Copy.ai
Similar evaluation to Jasper — genuinely useful for teams without strong writing skills needing to produce volume, but for an agency with good writers using AI as a multiplier rather than a replacement, the output ceiling was too low.
What We’ve Learned About AI Tool Adoption
The best AI tools amplify existing skill — they don’t replace it. Team members who were already strong writers get dramatically faster and better with AI assistance. Team members who struggled with writing still struggle — the AI output still requires judgement to evaluate and refine.
Workflow integration matters more than raw capability. A slightly less capable tool that fits naturally into existing workflows gets used. A more capable tool that requires context-switching and extra steps often gets abandoned.
Prompt libraries are underrated. We maintain a shared Notion database of prompts that have produced excellent results for specific task types. New team members can immediately benefit from prompts refined over months without having to rediscover them.
The cost adds up faster than expected. At agency scale, AI tool subscriptions across a team of 8 people add £600–900/month before API costs. Audit your stack quarterly and be ruthless about tools that aren’t pulling their weight.
The Honest Take
AI tools are not magic. The teams seeing the biggest productivity gains are the ones who have invested in building good workflows, maintaining prompt libraries, and training their people to use AI tools thoughtfully rather than expecting the tools to do the thinking for them.
The tools that earn a permanent place in our stack do one thing consistently: they give us back time to spend on the work that actually requires human judgment, creativity, and relationships. That’s the bar.
Everything else is a subscription waiting to be cancelled.
