There’s a ton of space between software that helps build something and software that actually acts on your behalf. The differences are innumerable and worthy of attention.
The world of web hosting has spent the last few years focused on the first, using innovative software that was great at helping to build websites. But now there is a rapid movement underway toward allowing AI to actually run websites, built on promises of a website or “store” that runs itself.
Sounds compelling, after all, for a small business owner, that kind of assistance might feel like relief. Most SMBs are saddled with running their websites in between everything else on their plates. The ability to automate customer interactions, updates to the online store, even the routine work that goes into managing a site, is genuinely attractive.
I’ve been in the web hosting game for two decades now and I completely understand the appeal.
But there’s a fundamental change that takes place the moment software stops suggesting and starts doing. Think of it this way. A tool that drafts a product description you have to edit and approve cannot embarrass you. An agent that skips those processes and publishes one certainly can. An agent that answers a customer question can commit you to a price, while one that updates a policy page can completely change what you are promising people.
None of that is meant to build an argument against using AI agents in this way. It’s simply making the case to decide the rules before they’re switched on, rather than after things go awry.
To that end, here are five critical checks worth making before letting an AI agent run your site.
Five checks before letting AI run your site
1) Decide which actions require human oversight: Draw a line between what the agent is permitted to do freely and what it must ask about before doing. If it’s drafting, sorting, flagging and preparing, let it do its thing. If it’s publishing to the live site, sending anything to a customer, changing prices or making policy adjustments, those should wait for approval before the agent completes them. Figuring which approach to choose comes down to one simple test: If the action is visible to the customer or difficult to step back, a human should see it first and provide approval. Most agent tools enable this approach, but very few businesses actually utilize it.
2) Know exactly what it can reach: AI agents tend to be given more access than is needed because creating those narrowed permissions is fiddly and simply granting access is quick. Before allowing an agent to go live, make sure you know what it’s able to read, what it’s able to change, what it can publish and what it can send on your behalf. If these can’t be answered without checking, the agent already has been given more authority than anyone actually granted it. Keep it scoped to the job it’s actually there to perform.
3) Have a rollback path, and test it: Before handing over the keys to the site, find out how an unwanted change would be reversed, how long it would take and who can actually make the reversal. Then try it while nothing is wrong to test the process; a recovery plan that never gets tested hinges on hope, not an actual plan. Every business knows the importance of having backups, but far fewer have ever tried to restore one, leaving a gap at the best of times, and considerably worse when the software has carte blanche to make changes faster than a person ever would.
4) Guard the facts that cost you money: The fluency and confidence that come from language models is exactly what makes them risky, especially around specifics. Things like pricing, delivery time, return and refund terms and product availability – anything that needs to be regulated – are where a plausible-sounding incorrect answer turns into real liability. Maintain a source of truth, keeping your most important business facts in one approved place the agent is required to read from each time it completes a task. Then test the agent on those facts, because those are the answers that can cost you money if it gets them wrong.
5) Define the moment a human takes over: Every automated system should have an exit, meaning a way to move particular scenarios to a human being expeditiously. An angry customer or a confused one is a good example of a scenario to build the agent’s understanding around, knowing when to hand it over to us rather than to keep trying. A loop that doesn’t escalate is the worst version of this. When the agent fails to do so, the customer cannot be helped. Decide on the triggers ahead of time, and make sure there is someone on the other end to take the handoff from the AI agent.
The question underneath all five checks
All of these checks stem from the same underlying question. When AI runs your site, who is accountable for what it does? The answer is not the vendor, not the agent and not the model behind it. It is always you. Customers experience the outcome of AI agents as something your business did, because from where they sit, it was.
Here’s the key takeaway to consider before turning anything on. Automation moves the work, it does not move the responsibility. If it quotes a wrong price the customer won’t be annoyed at the software, they’ll be annoyed with you and you will be the one that will have to make it right.
None of this is meant to suggest that we should be avoiding AI agents when it comes to operating our websites. Quite the contrary. Used diligently, they are highly valuable and small businesses stand to gain the most. Those that will outperform their competitors in leveraging AI are simply those that treated the setup as a decision rather than a default, spending time building rules before letting anything run unsupervised.
That time spent creating rules and guardrails is far cheaper than having to explain to a customer why your website shared something with them that was not true.
Opinions expressed by SmartBrief contributors are their own.
____________________________________
If you liked this article, sign up for the SmartBrief Marketing Pulse, a daily look at the latest in shifts and performance data for marketers.