The AI Tools We Actually Use in Client Work (And the Ones We Dropped)

After a year of integrating AI into production platforms and client workflows, we have a clear picture of what's genuinely useful and what's impressive in demos but falls apart in practice. Here's the honest breakdown — the tools we kept, the ones we dropped, and why.

The Evaluation Framework

We evaluate every AI tool on three criteria: reliability (does it produce consistent output under real production conditions?), integration quality (can it be wired into an existing workflow without significant custom engineering?), and actual value delivered (does it save meaningful time or create output that wouldn't exist otherwise?). A tool that passes one or two of these criteria but fails the third doesn't make the cut for client work.

What We Actually Use

Claude (Anthropic) — Writing, Strategy, Code Review

For long-form content generation, strategy documents, and code review, Claude produces the most consistently useful output we've tested. The key differentiator is context handling — Claude maintains coherence across long, complex prompts where other models drift or contradict themselves mid-response.

We use it for: first drafts of blog posts (refined by humans, not published as-is), technical documentation, analyzing audit data and producing written recommendations, and reviewing code for logic errors before shipping.

Where it falls short: anything requiring real-time data (it has a training cutoff), visual tasks, and multi-step agentic workflows that require browsing. For those, we layer in other tools.

ElevenLabs — Voiceover and SFX

For video content that requires narration, ElevenLabs is production-ready for 80% of use cases. We integrated it directly into ClipMe for automated voiceover generation. The text-to-speech quality on their curated voices is genuinely difficult to distinguish from a professional voice actor on most content types.

Where it works: informational content, marketing narration, explainer videos. Where it doesn't: comedy (timing is off), highly emotional content, and regional accent accuracy at the edges.

The SFX generation is underrated — you describe a sound and get a high-quality result back. Useful for short content elements; degrades past 22 seconds.

Cursor — Development Acceleration

For code, Cursor (VS Code + Claude/GPT backend) dramatically accelerates repetitive development tasks: writing boilerplate, generating test cases from existing patterns, refactoring functions to match a new interface, and explaining unfamiliar code. We use it on every project.

The danger with Cursor is over-reliance: AI-generated code is confidently wrong often enough that every output needs human review. Cursor is a multiplier for a developer who knows what correct looks like — not a replacement for knowing what correct looks like. Teams that use it without that foundation ship bugs faster.

Perplexity — Research and Fact-Finding

For research tasks that require current data — competitor analysis, industry statistics, recent platform changes — Perplexity outperforms general-purpose LLMs because it cites sources and pulls live data. We use it for initial research on any topic requiring facts rather than reasoning.

We verify anything that will appear in client-facing deliverables, but Perplexity narrows the search space dramatically compared to open-ended web searches.

What We Dropped

AI Image Generation for Client Work

Midjourney and similar tools produce impressive images, but not for service business marketing. The images look AI-generated to anyone paying attention — which is increasingly everyone — and fail the authenticity test that matters most for local service businesses. Real photos of real work outperform AI-generated visuals for trust and conversion every time.

We use AI image generation for internal mockups and ideation, never for client-facing marketing materials.

Automated Social Media Caption Generators

We tested several dedicated social media AI tools (Jasper, Copy.ai, several others). The output is generic, brand-agnostic, and immediately recognizable as AI-generated. For a client whose differentiator is voice and authenticity, posting AI captions is actively damaging. We dropped these entirely in favor of Claude with detailed brand context prompts — which produces better output with more customization.

AI Chatbots Without Human Escalation

Fully automated customer service chatbots — without a clear path to a human when the bot fails — create more damage than they prevent. We've seen businesses deploy chatbots that confidently provide wrong information about pricing, availability, and services, creating support issues that cost more to resolve than the chatbot saved. AI chatbots are useful for qualification and triage; they need human escalation built in from the start.

The Honest Take

AI tools in 2025 are powerful accelerators for specific tasks and unreliable replacements for human judgment. The businesses getting the most value from AI are using it to handle high-volume, lower-stakes tasks — first drafts, research, boilerplate, qualification — while keeping humans in the loop for anything that touches brand voice, creative judgment, or client relationships.

The businesses getting burned are the ones treating AI output as finished product. The output is always a starting point. What you do with it is still the differentiator.