Every agency claims to use AI. Most are using it for content drafts and calling it automation. We've been running an AI-first delivery pipeline for 18 months. Here's the honest inventory: what we use every day, what we tried and dropped, and what's overhyped for our use case.
What we use every day
1. GPT-4 / Claude for first-draft content
Every blog post starts as an AI draft. The senior writer takes that draft and turns it into something worth reading — the 20% that matters (positioning, voice, claims, the part an AI would get wrong). The 80% that's repetitive (research, structure, fact-stating) is AI. Volume: 80-120 blog posts per month across all clients.
2. Puppeteer + Lighthouse for technical audits
Every client site gets crawled weekly. We auto-detect indexation issues, broken links, redirect chains, schema validation, and Core Web Vitals regressions. A human reviews the diff against last week, not the full audit. Volume: 200+ sites crawled weekly, 3-5 regressions flagged per week that need human attention.
3. Ahrefs + Surfer for keyword research
Ahrefs gives us the SERP analysis, keyword difficulty, and content gap data. Surfer gives us the NLP-driven content structure that matches what Google rewards. Together they produce a content brief in 5 minutes that used to take 30-60 minutes manually.
4. Schema generation via templates
Every page on a client site gets the right schema markup (LocalBusiness, FAQ, HowTo, Article, Product, Review). It's a templated system with AI-assisted field population. We don't hand-write JSON-LD anymore.
5. Looker Studio for dashboards
Every client gets a live Looker Studio dashboard pulling from Google Search Console, Google Analytics 4, and our internal rank tracker. Clients can self-serve or wait for the monthly call.
6. White-labeled PDF report generation
The monthly client report is auto-generated on the 1st of each month from the same data as the dashboard. White-labeled with the agency's logo, brand colors, and template. 30 minutes of human review per client, not 2 hours of manual report building.
7. Resend for transactional email
Contact form submissions, audit completions, lead notifications. Resend delivers. We tried SendGrid, Mailgun, and Postmark — Resend is the cleanest developer experience.
8. Puppeteer for the audit tool
The /audit page runs 20+ checks in 60-90 seconds: Puppeteer crawls the site, Lighthouse runs, Ahrefs queries, the report is generated as a PDF and emailed. The audit tool is also our lead magnet — 100+ free audits shipped, 12-15% conversion to paid.
9. Klaviyo for email automation
For clients who need email + lifecycle (which is most of them). Klaviyo is the best ESP for e-com. Their AI subject line tool is genuinely useful — 15-25% lift on open rates for our clients.
10. Hotjar for behavior analytics
Session recordings, heatmaps, funnel analysis. Especially useful for CRO audits and understanding why a page converts (or doesn't).
11. Clearscope for content optimization
NLP-driven content scoring. Tells us what entities, terms, and topics to include in a piece to rank for a given query. We use it as a final QA pass on every published page.
12. Linear + Notion for project management
Linear for sprint planning and the operational backlog. Notion for client-facing documentation and SOPs. Both are overkill individually; together they're the minimum viable operations stack for a small team doing a lot of work.
What we tried and dropped
AI for client communication
Tried GPT-4 to draft client emails. The output was technically correct but missed the nuance of client relationships. Reverted to human-written emails for any client-facing communication. The 5 minutes saved wasn't worth the trust cost.
AI for design (Midjourney + DALL-E)
Tried for social media graphics and blog hero images. Quality was inconsistent. Clients noticed. Reverted to a mix of stock photography (for body content) and a human designer (for hero/branded visuals). For one-off social posts where quality tolerance is lower, AI is fine.
AI for sales calls
Tested Gong + AI-driven call summaries. Useful for note-taking, useless for actual sales. Sales is a human conversation. The summaries are great, the 'AI coach' features are noise.
AI for code generation
Use Copilot for boilerplate, but for the parts of our codebase that matter (the SEO engine, the audit tool, the deployment pipeline), the code is hand-written and reviewed. AI-generated code in production systems is a maintenance debt we don't want.
AI for strategy
Tried AI to draft quarterly strategy decks. The output was generic and missed the client's actual market. The senior strategist's 4 hours of work is worth 40 hours of AI output we'd have to edit anyway.
AI for paid ads creative
AI-generated ad creative works at low budget for testing. Doesn't scale to the level our paid media team operates at. For $5K/mo budgets, AI creative is great. For $50K/mo budgets, you need a human creative director.
What's overhyped
1. 'AI agents' that run your whole agency
The pitch: 'an AI agent handles your entire SEO workflow end-to-end.' The reality: 80% of an SEO workflow is the strategic 20% (positioning, voice, the part that requires understanding the client's business). AI agents are great at the boring 80%. They fail at the strategic 20%. Which is the part that matters.
2. 'Fully automated content' with no human touch
Works for the first 6 months. Decays after that. Google has gotten much better at detecting content that lacks E-E-A-T signals, and AI content without human refinement gets hit by Helpful Content updates. The hybrid model wins.
3. 'Just use ChatGPT for everything'
ChatGPT is a great general-purpose tool. It's not the right tool for any specific job. We use specialized tools (Surfer for SEO, Klaviyo for email, Puppeteer for crawling, etc.) because they do their specific job better. The 'one AI to rule them all' pitch is for buyers, not operators.
The bottom line
AI is the operating system, not the product. We use it to ship faster, cheaper, and at higher quality than the agencies that don't. The clients get more for less. The work is better. The humans spend their time on the 20% that actually moves the needle.
If you're an agency founder or a marketing leader, the question isn't 'should we use AI' — it's 'which parts of our delivery should we automate, and which should stay human.' The agencies that figure that out will be the ones standing in 5 years. The ones that don't will be selling hours.