Behind the Scenes: The AI Workflow We Use to Run an Agency (and What You Can Borrow)

By Leon Harris, Founder & AI Creative Director · Updated June 2026

You asked ChatGPT for a supplier price comparison and it invented numbers that looked real, costing you a quote. You want AI that does the grunt work without making you fact-check every sentence. This article gives you five daily habits from an AI-native team so your assistant becomes reliable, not reckless.

Here is the honest starting point. A 2025 study from the Philippine Institute for Development Studies (PIDS) found that only about 14.9% of Philippine firms use AI tools, and only around one in five firms are even aware of these technologies. That is not an access problem. The same study notes 90.8% of establishments own computers and 81% have internet. The missing piece is knowing how to use AI well. Learn these habits and you are already ahead of most of your competition in Metro Manila.

Habit 1: Feed AI the current facts, do not trust its memory

AI models make things up. The polite word is "hallucinate," but the result is the same: a confident, wrong answer about a price, a policy, or a date. The fix is simple and costs nothing. Before you ask AI anything fact-based, paste in the real source. Instead of asking "What is our refund policy?" and hoping it remembers, paste your actual policy text and say "Rewrite this clearly for a customer." For prices, room availability, or stock, give it today's numbers every time. The principle is everything here: give the AI the source, do not ask it to recall. A draft built on facts you supplied is safe to use; a draft built on the AI's memory is a guess.

Habit 2: Have a second AI check the first one's work

This is the single highest-value habit we use, and it is free. After one AI writes something, we open a fresh chat in a different app and give it a job: "You are a strict editor. Here is a caption written by another AI. List every factual error, weak claim, and anything off-brand." A fresh AI with no attachment to the first draft catches mistakes the original would defend. You can do this in two browser tabs, or draft in ChatGPT and review in Claude. Treat the critique as a checklist to work through, not gospel. It is the same fresh-eyes review good teams use on any draft, and it works just as well on a Facebook post as on a client proposal.

Habit 3: Let AI do the tedious research

The "deep research" mode in ChatGPT, Claude, and Perplexity reads dozens of sources and returns a summary with citations. Use it like a junior analyst. A useful prompt: "Research the five top competitors to my coffee shop in Makati, list their pricing and main offer, and cite your sources." What used to take an afternoon takes about half an hour. The catch, and it is a real one: you must verify. Open two or three of the cited links yourself before you trust anything. The AI drafts the research; you still make the call.

Habit 4: A browser-connected AI can research behind logins (optional, paid, risky)

Tools like Claude for Chrome work inside your already-logged-in browser, so the AI can read from sites you are signed into without copy-pasting. It is genuinely useful, but be careful. In 2025 and 2026, security researchers publicly documented serious flaws in this kind of browser AI, including ways a malicious website could quietly feed the AI hidden instructions (a "prompt injection") and take it over. Anthropic has patched the specific issues that were reported, but the category stays risky, and the extension requires a paid Claude plan (around US$20 a month, roughly PHP 1,150; the Chrome extension is not on the free tier). If you use it, point it only at sites you trust, keep your eyes on the screen, and never let it move money, post publicly, or change account settings without your own click.

Habit 5: Always keep a human approving (the non-negotiable)

This is where most of the value and all of the safety lives. A 2026 Connext Global survey of working AI users in the US found that only about 17% trust AI to run on its own, while roughly 70% say it is reliable only when paired with light review or dedicated human oversight. So set a hard rule: nothing AI-made gets posted, sent to a client, or published until a human reads it end to end and approves it. Keep approvals fast, under two minutes for low-risk items, so the workflow stays useful. AI does not decide your strategy, your tone, or what is accurate. You do. The AI is a workflow layer for the boring parts, not a replacement for judgment.

How to start without spending anything

Pick one repeatable, low-risk task first. Do not try to automate everything. Good first picks: drafting social captions, summarizing a client report, or writing FAQ replies. Then run the loop: paste the real source, let AI draft, have a second AI review, then you approve. Four of these five habits run entirely on the free tiers of ChatGPT, Claude, or Perplexity. Only the browser tool in Habit 4 is paid, and it is optional. You can even run the loop from your phone between client visits; the Claude and ChatGPT mobile apps are free to start, and Claude's voice mode became free for everyone in 2026. Review weekly: if you are rejecting more than half of what the AI produces, your inputs or your chosen task are wrong, so tighten the prompt or pick a different job. The real cost here is discipline, not pesos.

Frequently asked questions

What does it mean that AI "hallucinates," and why should a business owner care?

Hallucinating means the AI confidently states something false: a wrong price, a made-up policy, a fake statistic. It matters because people act on it, and a wrong number sent to a client is your problem, not the AI's. The fix is simple: feed the AI the correct current facts instead of asking it to remember, and have a human or a second AI check before anything ships.

Do I need to be technical or spend money to copy these habits?

No. Four of the five habits run on the free tiers of ChatGPT, Claude, or Perplexity plus your own copy-paste. The only paid habit, a browser-connected AI, is optional and needs a roughly US$20-a-month plan (about PHP 1,150). The real investment is discipline: give correct inputs, double-check with a second AI, and approve before publishing.

How exactly do I have a second AI check the first one?

Open a fresh chat in a different app or tab and give it a role: "You are a strict editor. Here is something another AI wrote. List every factual error, weak claim, and anything off-brand." A fresh AI with no attachment to the first draft catches mistakes the original would defend. You can do it for free in two browser tabs, or draft in one app and review in another.

Is it safe to let an AI use my logged-in browser?

Be careful. Browser-connected AI tools like Claude for Chrome are useful but have had real, publicly documented security flaws, including ways a malicious website could feed the AI hidden instructions and hijack it. The reported issues have been patched, but the category stays risky. If you use one, point it only at sites you trust, keep your eyes on the screen, and never let it send money, post publicly, or change settings without your explicit click.

If AI does so much, do I still need people?

Yes, and that is the point. A 2026 Connext Global survey of US workers who use AI found only about 17% trust it to run on its own, while roughly 70% say it is reliable only with light review or human oversight. The best small teams use AI for repetitive work, drafting, summarizing, and researching, while humans own strategy, tone, accuracy, and final approval. AI does not decide what is true or what ships. You do.

Where should a Philippine small business start, given most peers have not adopted AI yet?

Start with one repeatable, low-risk task: email drafts, client-report summaries, or social captions. With only about 14.9% of Philippine firms using AI tools (PIDS), even adopting these basic habits puts you ahead of most competitors. Run the loop: paste the real source, let AI draft, have a second AI review, then you approve. Add a second workflow only once the first is reliable.

Can I do all this from my phone?

Yes. The Claude and ChatGPT mobile apps for iOS and Android are free to start, and Claude's voice mode became free for everyone in 2026. You can paste a source, ask for a draft, switch apps for a second-AI review, and approve, all from your phone between client visits.

Want this done for you? We build the AI content system; you approve. Free audit + custom quote within 24 hours.

Get a free audit →