Letting AI touch live client sites

Bob Ojo-Ami's framing is the uncomfortable starting point: the barrier to entry for agency-quality output has collapsed, and a client can now produce comparable work for less than 10% of the cost of hiring you. His example is vivid, a UI/UX designer who once handed off to developers can now use AI to build the whole site themselves, making the one-person agency increasingly viable and eating into the traditional model.

Speed and polish, once premium services, are now simply the baseline clients expect, and any billing model tied to hours is structurally exposed because the hours are collapsing. You cannot decide how to use AI on live sites until you have accepted that it is reshaping what clients will pay for at all.

Thomas Amos is emphatic that you cannot plug AI into chaos, because it accelerates the chaos rather than fixing it. Map the end-to-end customer journey first, because the journey defines which SOPs you actually need, then apply a version of Musk's five-step framework: question every requirement, delete the unnecessary steps, simplify what remains, accelerate, and only then automate.

Building SOPs before you understand the journey produces documentation nobody uses. Amos also insists on measuring a baseline first, his agency knew their sales cycle was seven hours before they touched it, so they could prove whether AI was helping, and without that number improvement is unmeasurable. Groundwork is not the boring part before the AI, it is the part that determines whether the AI helps or harms.

Rather than sprinkling AI across the whole business at once, Amos says identify the single most painful problem and solve that first. For Designbox that was sales-cycle time, and fixing one bottleneck well beat a dozen half-finished integrations that impress nobody.

He is clear that AI works best as an amplifier alongside your team, not a replacement for headcount, giving every person effective assistance rather than trying to remove them. Crucially, the differentiator in AI-assisted output is training the system on your specific business, general tools like ChatGPT or Claude do not know what your agency delivers, its pricing, or its processes, so a system grounded in your specifics produces work that is actually usable without heavy correction. Pick the one place a fix pays for itself, prove it there, and expand from evidence.

Aaron Edwards frames the whole opportunity as service as software: clients do not want AI, they want reduced support load, higher conversions, and automated workflows, and AI is only the means. His proof points are real, 76% automatic resolution of support queries, 89% escalation prevention, 97% satisfaction, which is what lets you sell an outcome rather than a novelty.

He is candid that AI is not yet eliminating most work, but that when organisations try to implement it, it exposes pre-existing gaps like bad data and broken process, which is itself a service you can sell. He lists concrete new lines available today: AI audits, automation retainers, support and sales agents, AI employee packages, internal knowledge bots, bespoke integrations, and team education. Package the result, price it against the client's pain, and leave the word AI off the invoice.

Edwards is reassuring on the point that paralyses many owners: you are not behind, because nobody has fully figured AI out, and being one week ahead of your client is enough to be the expert. The tension for agencies is that they are supposed to sell certainty, yet the foundational technology shifts weekly, so it can feel like guesswork, but the fix is not to become an AI developer.

You do not have to build the technology yourself, partnering with AI platform providers lets you deploy sophisticated solutions in days or weeks without running an internal R&D function. The agency's irreplaceable asset is the client relationship and the domain context, not the model, so let someone else maintain the engine while you own the outcome on the client's site.

Bowe Frankema's thesis is that AI now delivers roughly 70% of what any agency does, and that 70% is free to everyone, so your survival lives in the last 30%: your aesthetic judgment, client history, scoping standards, voice, and delivery process. Because the model itself is available to all, it cannot be your advantage; everything you bring to the model can be.

He illustrates it with onboarding a brilliant new hire who knows nothing about how you work, that is exactly what an ungrounded AI is. Context engineering is the practice of capturing that agency brain and exposing it to your agents, and the simplest start is a single markdown file built by having an AI interview you in voice mode until it understands your processes, brand, ideal client, and culture. Without this, agents on a live site produce generic work that sounds like everyone else's.

Frankema is clear that a static context file is a foundation, not a solution, the real leverage comes from dynamic context pulled live from your CRM, client emails, WordPress sites, and other systems. Connect one brain to many tools, via a protocol like MCP, and you stop re-explaining yourself every time you switch between Claude, Cursor, ChatGPT, or a builder, with no manual handovers and no context loss.

Once context is always live, agents can move from reactive to proactive: you can schedule proposals, daily updates, and weekly WooCommerce reports rather than triggering each one by hand. The agency loop the best have mastered is to gather context on the fly, make it available where you work, deliver, remember what you delivered, and repeat. That loop is what makes an agent genuinely useful on live client work rather than a clever demo.

JJ Toothman onboarded an AI agent, Viv, through the exact checklist he uses for every human hire at Lone Rock Point, a WordPress agency that moved NASA off Drupal and builds in mission-critical government environments where accuracy is paramount. He argues you should write a proper job description, title, responsibilities, and access scope, before you touch prompt engineering, because most agencies do not have an AI problem, they have a clarity problem about which roles to delegate.

Real integrations, Google Workspace, Asana, meeting transcripts, are the unglamorous half that gives the agent genuine situational awareness. The feedback loop is five to fifteen minutes, so you learn fast whether your instructions and access were right, but be warned the out-of-the-box memory is weak, retaining context for only 24 to 36 hours, so it needs extending to be durable. Treat the agent as a hire with a name, a role, and rules, and it behaves like one.

Vito Peleg's system shows how context makes on-site AI trustworthy: a project brief generated in one click gives the AI the brand assets, tone of voice, design preferences, value proposition, and project goals it needs to stay in the right lane. His AI review runs specialist agents, Pixel for design, Lexi for content, and others for SEO, accessibility, and UX, orchestrated by a central agent, and it operates at the element level rather than a zoomed-out page view, giving it materially higher fidelity than a general LLM.

A normal QA pass takes 45 to 60 minutes and needs to happen five times per project; the AI review reaches a comparable or better result in minutes. The lesson for letting AI touch a live site is that fidelity and context, not raw model power, are what make its output safe to ship.

Bogdan Condurache demonstrated both a fast flow and a custom flow and was explicit that AI output is a starting point, with 20 to 30% of the work still manual, and he argues the bespoke, human-crafted result carries premium value in an AI-saturated market, just as custom tailoring outsells machine-made. His recommended operating model is an 80/20 or 70/30 split: let AI handle the heavy lifting on structure and content, then invest the final fifth to third in distinctly human creative judgment.

Vito Peleg's Show Me feature reinforces the safety pattern, it renders a suggested change non-destructively on the front end without writing to the database, while Do It executes it, so a human can approve before anything touches the live site. Never let the agent be both author and approver on production work.

Aaron Edwards describes the real frontier as the shift from tools to teammates: AI agents are moving beyond answering questions to taking actions, resolving billing issues, triggering workflows, filling forms, and interacting with external systems autonomously. That is powerful and precisely why it demands guardrails when the target is a live client site rather than a sandbox.

The agencies that win, he stresses, will be the first to implement well, not the ones who understand AI most deeply in the abstract, so bias toward shipping, but scope the agent's authority explicitly. Decide in advance what an agent may do alone, what needs human sign-off, and what it must never touch. An autonomous teammate is only an asset when its boundaries are as clearly defined as a human's.

Alex Frison offers the sober counterweight: AI makes individual developer tasks faster, but communication, alignment, expectation-setting, and decision-making still take exactly as long as before. The build shrinks; the engagement does not, so do not promise clients a compressed timeline just because the code is quicker, or you will manufacture disappointment.

Vito Peleg's own data underlines it, building a five-page site now takes about three hours, yet delivering it with a client still takes four to six weeks, and 67% of email feedback is too vague to action on the first pass. Karim Marucchi's guidance on the dividend is the smart move: reinvest the time AI saves into deeper discovery and more thorough QA rather than simply doing more of the same faster. The agencies that win with AI on live sites spend the saving on quality, not volume.