Here’s What Actually Happened

There’s a lot of hype right now about AI “agents” that can just go off and do real work for you, log into your accounts, complete tasks, come back when it’s done.

We decided to actually test that claim properly, using our own sample WordPress site as a real-world task: log in, and draft a couple of blog posts.

It did not go the way the marketing suggests.

What Happened With Grok Bot

We gave Grok Bot access to a mailbox and asked it to log into our test WordPress site and start drafting blog content.

The moment it hit our site’s two-factor authentication, things fell apart. 2FA requires a live, real-time response, a code sent to an inbox or device that has to be entered within a short window, and an AI agent working through email access simply isn’t built to do that smoothly.

The process stalled, timed out, and then asked us to pay for more usage credits, without having produced any actual blog drafts.

To be fair to Grok Bot specifically, this isn’t really a “Grok is bad” problem. It’s a structural problem with how most agent platforms currently handle authentication, and it’s worth understanding why.

Why 2FA Breaks Agentic AI (It’s Not Just Us)

Two-factor authentication is built entirely around the assumption that a human is sitting there to respond to it, tap “approve” on a phone notification, read a text message, scan a QR code.

AI agents break every one of those assumptions. There’s no person available to respond to a real-time challenge at 2am, and there’s no clean way for most current agent platforms to hand that moment back to a human and then resume exactly where they left off.

This isn’t a fringe complaint either. Broader research backs up what we saw first-hand:

On the WebArena benchmark, which tests AI agents on realistic website tasks, even the best current agents complete only around 60% of tasks successfully, compared to a 78% human success rate
Industry analysts at Gartner predict that over 40% of agentic AI projects will be abandoned by 2027, largely due to governance and reliability gaps, not model quality
Security researchers have flagged authentication and identity management as one of the biggest unresolved problems in agentic AI for 2026, since most agent platforms still can’t cleanly hand off “this needs a human” moments

In other words: we didn’t hit an edge case. We hit exactly the kind of wall that’s showing up across the industry right now.

Lindy.ai: A Different Kind of Limitation

While looking at alternatives, we found Lindy.ai currently requires a Slack account to actually operate, Microsoft Teams support is listed as “coming soon” but isn’t available yet.

For a lot of small Australian businesses that live in Outlook and Teams rather than Slack, that’s an immediate practical blocker, not a technical limitation of the AI itself, just a platform integration gap that rules it out for now.

What This Means in Practice

Between our own test and what the wider research shows, here’s the honest state of play with agentic AI in September 2026:

Authentication is still a fundamental blocker. Anything behind 2FA, and most business logins are, remains genuinely difficult for an autonomous agent to get through cleanly.
“More work needed” often means “more credits needed.” We hit a paywall before we hit a finished task, which is a pattern worth watching for.
Platform lock-in is real. Which chat or work tools you already use (Slack vs Teams, Gmail vs Outlook) can determine whether a given agent platform is even usable for you yet, regardless of how capable the underlying AI is.
The gap between “demo” and “your actual business” is still wide. Agents perform well on curated demos and much less reliably on a real website with real security measures in place.

What We’re Testing Next

We’re not writing off agentic AI, we’re just being realistic about where it’s actually at. Next on our list is putting Make.com and Claude’s own Agent interface through the same real-world test: log in, draft content, report back. We’ll share how that goes.

In the meantime, this is exactly why our own automation setup uses a deterministic, draft-and-approve model rather than full autonomy, a human checks and approves every step before anything actually happens. If a system built by a company as well-resourced as OpenAI can get caught out by its own agents (as we covered in a recent post), it’s reasonable that a small business shouldn’t hand over full unsupervised control just yet either.

You can read more about how we think through AI for marketing, sales and customer service on our AI in Marketing, Sales and Customer Service page.

This article reflects our own hands-on testing as at September 2026. Agent platforms are being updated rapidly, results may differ by the time you read this.