I Tested Five AI Automations for Real Work—Here Is What Actually Saved Time

Everyone talks about AI automation. Few people share honest results. This article is a case study of five AI automation experiments I ran for real work — not demos, not thought experiments. I will tell you the problem, the tools, the workflow, the setup time, the time saved, what went wrong, and whether I still use each one.

Experiment 1: Content research

The problem

Researching articles and market reports takes hours. Finding credible sources, reading them, extracting key data points, and synthesising findings into a usable structure. Most of the time is spent on search and filtering, not analysis.

Tools tested

Perplexity Pro, Claude with web search, and a custom LangGraph research agent.

Workflow

Give the tool a research question. It searches the web, evaluates sources, reads the most relevant ones, extracts key points, and returns a structured summary with citations. I review, verify, and refine.

Time spent setting up

Zero for Perplexity (works out of the box). 2 hours for the LangGraph agent.

Time saved

3–4 hours per research task. What took a morning now takes 20–30 minutes of review and refinement.

What went wrong

The LangGraph agent occasionally hallucinated data points that were not in the source material. It also struggled with paywalled content — inferring article content from headline and snippet. Perplexity Pro was more reliable because it is purpose-built for research with built-in source citation.

Do I still use it?

Yes. Perplexity Pro is my daily research tool. I retired the custom LangGraph agent — the maintenance cost was not worth it when Perplexity already does 90% of what I needed.

Experiment 2: Social-media content

The problem

Creating platform-specific social media content from articles and ideas. LinkedIn needs professional, opinionated posts. Instagram needs visual hooks. YouTube needs video. Each platform has different requirements and audiences.

Tools tested

Claude Sonnet 4 with platform-specific prompt templates, and a Make.com scenario that takes a published article URL and generates social posts.

Workflow

Article published → Make.com reads the article → Claude generates 3 LinkedIn post options, 2 Instagram caption ideas, and a YouTube description → saves to a Google Sheet → I select, edit, and schedule.

Time spent setting up

4 hours for the Make.com scenario and prompt templates.

Time saved

1–2 hours per article in social media content creation. The options give me a starting point; I edit rather than write from scratch.

What went wrong

The generated LinkedIn posts were too generic. They summarised the article instead of extracting a specific, opinionated angle. LinkedIn rewards specificity and personality, not summaries. I had to rewrite most of them significantly. The Instagram captions were fine but lacked visual direction — they did not tell me what image to pair with the text.

Do I still use it?

Partially. I use the Make.com scenario to generate LinkedIn post options, but I heavily edit them. Instagram and YouTube I handle manually because they require visual judgement the AI cannot provide.

Experiment 3: Email and CRM automation

The problem

Sorting, prioritising, and responding to email. 30–60 minutes per day spent on triage alone.

Tools tested

Make.com with GPT-4o for email triage and auto-drafting, connected to Gmail.

Workflow

Email arrives → Make.com reads it → GPT-4o categorises (urgent, newsletter, inquiry, personal, spam) → drafts a response for inquiries → saves draft to Gmail → I review and send.

Time spent setting up

3 hours.

Time saved

20–30 minutes per day. The categorisation alone saves time — I no longer open every email to determine its nature.

What went wrong

Auto-drafted responses were generic and sometimes missed context from earlier in the email thread. For ongoing conversations, the drafted response felt tone-deaf — it did not account for the relationship history or the emotional context. I learned to auto-draft only for new inquiries, not ongoing threads.

Do I still use it?

Yes, for triage. The categorisation and prioritisation are worth the setup. Auto-drafting I use selectively — only for straightforward, first-time inquiries. Anything involving a relationship or ongoing conversation, I handle manually.

Experiment 4: Website creation

The problem

Building and maintaining a professional multi-page website. I needed a personal website with home, about, experience, projects, articles, gallery, and contact pages. Professional web development costs thousands of dirhams.

Tools tested

Kimi K2.7 for planning and content, Qwen 3 Coder for HTML/CSS code generation, MiniMax 5.2 for refinement. Deployed on Vercel.

Workflow

Used AI models to plan the site structure, write all page content, generate HTML and CSS, and deploy. I guided the process, reviewed output, and refined.

Time spent setting up

6 hours total, from planning to deployment. The only direct cost was AED 40 for the domain name.

Time saved

This is not a time-saving automation — it is a cost-saving one. Instead of spending AED 15,000–30,000 on a developer, I spent AED 40 and 6 hours. The website you are reading this on is the result.

What went wrong

The AI models produced inconsistent code quality. Some pages needed multiple iterations. The CSS needed manual cleanup. But the result is a production-quality, SEO-optimised, multi-page website that loads fast and works on all devices. The full story is in my article about building a website for AED 40.

Do I still use it?

Yes. I maintain and update the website using the same AI-assisted approach. Changes that would have taken a developer days now take hours.

Experiment 5: Image and video production

The problem

Creating visual content for articles, social media, and marketing materials. Professional photography and design are expensive. Stock photos are generic. AI image generation offers a middle ground.

Tools tested

Midjourney for concept images, Flux Pro for article headers, Stable Diffusion locally for mockups, and DALL-E 3 for quick illustrations. For video: Runway Gen-3 and Pika for short clips.

Workflow

For articles: write the article → identify the visual concept → generate 3–5 image options with Flux Pro or Midjourney → select and crop → use as article header. For social media: generate platform-specific visuals.

Time spent setting up

1 hour for prompt templates. No complex setup — these are web-based tools.

Time saved

2–3 hours per article (compared to sourcing or commissioning images). For social media, 1 hour per post set.

What went wrong

AI-generated images have limitations. Hands are still problematic. Text in images is unreliable. The style can look generic if prompts are not specific enough. Video generation is improving but still produces short, low-resolution clips that need editing. For professional brand visuals (property photography, headshots, event coverage), AI is not a substitute for real photography.

Do I still use it?

Yes, for specific use cases. Article header images, social media concept visuals, and mood boards. For anything requiring real people, real properties, or brand-specific photography, I use a camera or hire a photographer.

The honest summary

"The automations that stuck are the ones where AI did the boring part and I kept the interesting part. The ones that failed tried to automate the interesting part — and produced generic, lifeless output."

The pattern

Looking across all five experiments, the pattern is clear. AI automation works best when:

• The task is repetitive and well-defined (email triage, research gathering)
• The output is a starting point, not a final product (social media drafts, website code)
• A human reviews and refines before the output reaches its audience
• The volume justifies the setup cost (daily email, weekly articles)

AI automation fails when:

• The task requires taste, personality, or relationship awareness (social media voice, email tone)
• The output is used without review (hallucinated data, generic content)
• The setup is more complex than the task it automates (over-engineered agent chains)
• The volume is too low to justify the investment (one-off tasks)

The bottom line

AI automation is not magic. It is a set of tools that, used correctly, can save meaningful time and money. Across five experiments, I saved approximately 8–10 hours per week and AED 15,000+ in web development costs. But the savings came with a cost: setup time, maintenance, and the discipline to review every AI output before it goes live.

The most valuable lesson is this: automate the process, not the judgement. Let AI gather, sort, draft, and generate options. Make the final decision yourself. That is where the line between useful automation and expensive failure lives.