AI Automation FAQs — 30 Questions Answered

Five to seven business days end-to-end if you have credentials ready (API keys, OAuth, sample data) on day one. Three days of build, one day of test with real payloads, one day for you to dogfood it before we flip it live. If you have to chase a vendor's IT team for an API key, the clock pauses — that wait is the bottleneck in 80% of projects, not the building.

Day 0 you fill the brief. Day 1 we send a Loom walking through scope, edge cases, and what we're not doing. Day 2-3 you approve, send credentials, and we kick off. Day 4-10 we build in a shared Notion with daily 60-second progress videos. Day 11-14 you run live with real data while we sit on standby. Total: 14 days for one workflow, 21-28 for a bundle of three. No discovery calls longer than 45 minutes.

Fixed-price per workflow or per site. Hourly billing rewards slow building — it's bad for both sides. A 5-step n8n automation is a fixed scope: trigger, transform, branch, action, log. We quote it, write it, ship it. If scope changes mid-build, we send a change order with a new fixed number before continuing.

One workflow, one website, or one chatbot. We don't do "an hour of help" gigs — context-switching costs more than the hour pays. If you need a 20-minute tweak to an existing automation we already built, that's included in a 30-day handover window. After that we move you to a small retainer.

Both. Every build includes a 30-day bug-fix window. After that you pick: no retainer (we're on standby at hourly), light retainer (4 hours per month, monitoring + small tweaks), or growth retainer (16+ hours, new flows and optimization). Most clients pick light retainer for the first quarter then upgrade once the wins compound.

Yes. /case-studies has anonymized writeups of shipped projects with screenshots, n8n graph exports, and outcome metrics. For specific stacks (HIPAA, Shopify, healthcare scheduling) we send a 5-minute Loom walking through a live build we own. We do not share client production URLs without their consent.

Zapier is rented infrastructure with a polished UI; n8n is self-hostable infrastructure with a less polished UI. Pick Zapier if you have one or two simple flows, no engineers, and zero interest in hosting anything. Pick n8n if you have more than five flows, any need for branching/loops, or you want to avoid per-task pricing that compounds. Above ~50K tasks/month, self-hosted n8n on a cheap VPS beats Zapier's bill by roughly a factor of 10.

Yes, but do it one flow at a time. Export the Zapier zap as a screenshot map of steps, rebuild in n8n in parallel, run both for 48 hours with a comparison dashboard, then disable the Zapier version. Never bulk-migrate — you'll lose tribal knowledge about edge cases that lived in step descriptions. A typical 12-zap migration takes us 3 weeks.

n8n Cloud is fine until you hit one of three walls: you need IP allowlisting for an enterprise API, you need to run more than ~10 concurrent executions, or you want to write custom nodes in TypeScript. Self-host on a small Hetzner VPS with Docker + Caddy + Postgres takes 90 minutes to set up. We do this for every client who plans to run 10+ flows.

Three layers: workflow-level error trigger (catches uncaught exceptions, sends to Slack), node-level "continue on fail" with a branched error path that logs to a Postgres table, and a daily cron that re-runs anything in the error table within a retry window. Never just rely on n8n's default retry — it doesn't handle webhook idempotency, so you can double-charge a customer if you're not careful.

Queue mode separates the n8n main process from worker processes using Redis as the broker. You need it when a single execution can take longer than your webhook timeout, when you process more than ~1K executions/hour, or when you want zero-downtime restarts. For a one-VPS setup with light traffic, regular mode is fine — don't over-engineer.

Yes — there are first-party LLM nodes plus generic HTTP request nodes for anything not yet wrapped. The real reliability question is: what do you do when the LLM returns malformed JSON? We always parse with a try/catch in a Code node, fall back to a smaller model on first failure, and log the bad output to Postgres for prompt iteration. Don't trust LLM output to be schema-valid without verification.

Three usual culprits: your page lacks structured data that maps to the entities Google's ranking models care about (Product, Offer, FAQPage), your content reads like sales copy instead of answering a question directly in the first 80 words, and you have no external mentions on sites that LLMs treat as authoritative (Wikipedia, Reddit, niche industry forums). Fix all three before blaming the algorithm.

SEO optimizes for the ten blue links — keyword targeting, backlinks, page speed. AEO (Answer Engine Optimization) optimizes for LLM chat answers — Claude, ChatGPT, Perplexity, Gemini citing you in a generated paragraph. GEO (Generative Engine Optimization) is a near-synonym for AEO that some agencies use to sound new. They share fundamentals but the surface area is different: AEO rewards definitions, lists, comparison tables, and clear authorship.

Only if you sell your content as the product (paywalled news, paid research). For a service business, blocking them is shooting yourself in the foot — you want LLMs to scrape, index, and cite you. Add an llms.txt file explicitly listing what you allow them to use. The ROI of one citation in a Perplexity answer outweighs the marginal cost of being scraped.

Two to twelve weeks depending on freshness signals and how the four engines index. Perplexity tends to update fastest (often within 7 days of publishing). ChatGPT browse and Gemini take longer because they pull from Bing and Google Search indexes. Claude with web search now picks up new content within days. Track citation appearance in a sheet weekly — if nothing shows in 8 weeks, the issue is content quality, not patience.

llms.txt is a draft standard (proposed by Jeremy Howard) for a markdown index file at /llms.txt that tells LLMs which of your pages are canonical, citable, and structured for ingestion. It's not yet enforced by any major engine, but Claude, Perplexity, and a few others have started honoring it. Cost to add: 30 minutes. Upside: small but compounding. Yes, add one.

More than ever, because schema is how an LLM disambiguates entities. Without it, "Bali" could be the island, the type of cat, or a restaurant name. With Place + geo coordinates, the model knows. Schema is also a freshness signal: dateModified in JSON-LD is a heuristic LLMs use to prefer your version over a stale Reddit thread.

Use the model that wins on your specific evals, not the one with the loudest marketing. For tool-use heavy bots (booking, CRM updates) Claude tends to lead on instruction following. For long-context document Q&A, Gemini's 2M context is hard to beat. For raw price-per-token on simple replies, Haiku and Gemini Flash are both cheap. Build with the SDK that lets you swap providers in 10 minutes.

Don't let the model generate prices freely. Put pricing in a structured tool call: the bot asks "do you want pricing for X?" then calls a function that returns the current price from your CMS or a static JSON file. The model only paraphrases — it never invents. We use this pattern on every commerce-adjacent bot we ship.

RAG (Retrieval-Augmented Generation) means the bot searches your knowledge base first, then passes relevant chunks to the LLM as context. You need it when your content base is bigger than ~30K tokens (the comfort zone for stuffing everything into a system prompt). For a 50-page FAQ, a 500-row product table, or a year of email transcripts, RAG with a vector DB like Qdrant or Postgres + pgvector is the right call.

Use the official Meta Cloud API via Business Manager — not unofficial wrappers that get your number banned. Verify your business, get a template message approved for outbound, then use freeform replies inside the 24-hour session window. For automation, route inbound through n8n or a custom webhook that decides: send to LLM, send to human, send to template. Approval for the first template takes 1-2 days.

Yes for low-ticket transactions (small average order). Above the impulse-purchase threshold, conversion drops because buyers want at least one human touchpoint before committing. Our pattern: bot qualifies and books a 15-minute call, human closes. For services where the customer already knows what they want (haircut appointment, dental cleaning), bot can close end-to-end — we've shipped these.

Three metrics, not vanity ones. Containment rate: % of conversations the bot resolves without human handoff. Goal conversion rate: % of conversations that hit your defined outcome (booking, lead capture, purchase). CSAT or thumbs-up rate per conversation. Track these weekly. If containment is over 60% and CSAT is over 4/5, the bot is working. Below either, the bot is making your service worse.

Waseem Nasir leads every project. For larger builds a senior partner handles video/visual work. We don't subcontract to junior offshore teams — if you hire SkynetLabs you get the founder on Slack. This caps how many projects we run concurrently (4-6 at a time). When the queue is full we say so and propose a start date.

Bali (UTC+8) currently. Overlap windows: 7am-12pm GMT, 2pm-7pm EST, all day APAC. We schedule sync calls in your window when possible. Async is the default — daily Loom updates, shared Notion, message-when-needed. Most clients prefer this once they realize they don't need to "meet" to get progress.

Yes — happy to sign your standard NDA before scope discussion. We also have our own one-pager mutual NDA we can send if you don't have one ready. We don't sign non-competes that prevent working in your industry — that's how agencies lose the ability to take any client and we'd rather walk away than agree.

We've shipped 180+ workflows and 40+ websites — projects do occasionally go sideways. The protocol: we call it out the same day, propose a fix in writing within 24 hours, and if the fix is our fault we eat the cost. If the issue is scope creep or third-party blocker, we send a change order. We don't ghost. We don't blame the client. We've issued a refund twice in three years; both came with a written postmortem.

Yes, often. We slot in as the automation/AEO specialist while your dev handles the main product. We document everything in your tools (your Linear, your Notion, your Slack) so there's no handover tax later. We've co-shipped with internal teams in 12 of the last 30 projects.

Fine — most clients aren't. We translate every decision into outcomes ("this saves 4 hours of manual data entry per week") and document the build in plain English. After ship, you get a 20-minute walkthrough video and a one-page operations doc. You should be able to explain what your automation does to a friend over coffee. If you can't, we didn't document it well enough.