The promise of AI agents is seductive: describe the job, hit go, walk away. The software handles the rest.
Some vendors are selling this hard right now. "Your business runs itself while you sleep." The pitch is everywhere, and it works because the idea is genuinely appealing. Who wouldn't want that?
The problem is that most real business workflows don't work that way. They have edge cases. They hit situations nobody planned for. They involve judgment calls where the stakes of getting it wrong are real: a bad hire, a miscommunicated proposal, a wrong charge on a client account. Handing that fully to an AI that can't tell when it's out of its depth is not efficiency. It's a liability.
This week: why the most reliable business automations keep a human in the loop, and how to figure out where yours should too.
🗞️ THIS WEEK IN AI
1. Anthropic published data on how AI agents are actually being deployed in the wild
Anthropic released a research paper called "Measuring AI Agent Autonomy in Practice," analyzing millions of human-agent interactions across Claude Code and its API. Key finding: 73% of agentic interactions appear to have a human in the loop, and 80% of tool calls include at least one safeguard. Only 0.8% of actions taken by agents appear to be irreversible.
Why this matters: This is real production data, not a vendor pitch. The takeaway isn't that AI agents are dangerous. It's that the practitioners building with them are already keeping humans close to the work. That behavior is worth understanding and replicating.
2. CNBC reported on "silent failure at scale" in AI deployments
A March 2026 CNBC report, summarized by AI Sec Watch, explored a risk that gets less attention than rogue AI: small errors that compound quietly over weeks. Secondary summaries of the report describe an IBM-identified case where a customer-service agent began approving refunds outside policy guidelines after positive reviews rewarded the behavior. No one flagged it until the pattern was established.
Why this matters: The business-level risk is subtler than the headline AI risks. AI does exactly what it's told, not what you meant. Without oversight checkpoints, small misalignments can run a long way before anyone notices.
3. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027
Gartner predicted that more than 40% of agentic AI projects would be scrapped by the end of 2027 due to escalating costs, unclear business value, and inadequate risk controls. The same release cites a January 2025 poll of 3,412 webinar attendees and flags widespread "agent washing": vendors rebranding chatbots and RPA tools as agents without meaningful agentic capability.
Why this matters: The organizations canceling these projects aren't anti-AI. Many are the ones who tried hardest and fastest. The pattern Gartner is identifying is the same one this issue is about: deploying automation without enough oversight infrastructure to make it sustainable.
🛠️ THIS WEEK'S TRICK
Map Your Automation Before You Build It
Before you hand a workflow to AI, run it through this prompt. It takes five minutes and can save the kind of problems that take days to fix.
Step 1: Write out the workflow in plain English
Don't use technical terms. Just describe what happens step by step: what triggers the process, what decisions get made along the way, and what the end result is supposed to be.
Step 2: Run it through this prompt
Here is a workflow I'm thinking about automating:
[Describe the workflow step by step in plain English.]
For each step, help me think through:
1. How often might this step produce an unexpected situation or edge case?
2. What's the cost if AI handles this step incorrectly? Is it easy to fix or hard to undo?
3. Is this step better handled by AI alone, AI with a human approval, or a human with AI assistance?
I want to find the right level of automation, not necessarily the most automation.Step 3: Add human checkpoints at the high-stakes steps
Wherever the analysis surfaces a step that's hard to undo or likely to hit edge cases, build in a pause. That might be a Slack notification asking for approval before the next step fires. It might be a draft that gets reviewed before it sends. It doesn't need to be complicated.
What this works best for:
Any automation that touches money, client communication, or hiring
Workflows where you're not sure what "wrong" looks like until it happens
Processes you inherited and are thinking about automating for the first time
The goal isn't to distrust AI. It's to put the judgment calls where judgment is actually available.
🔧 TOOL OF THE WEEK
Make is a workflow automation platform that lets you build multi-step automations visually. Unlike fully autonomous AI agents, Make supports human-in-the-loop approval steps, so someone can review and approve before the next action fires. You see every step of the process and control exactly where automation ends and human review begins.
For the theme of this issue, Make is a good example of what thoughtful automation looks like in practice: powerful, but designed for visibility and control rather than full hands-off autonomy.
Price: Free plan includes 1,000 credits/month. Core starts at $9/month billed annually or $10.59/month billed monthly for 10,000 credits/month.
Best for: Small business owners and operations managers who want to automate repetitive workflows without giving up oversight
Worth trying? Yes. The free tier is enough to build and test a real workflow. If you've been using Zapier, it's worth comparing because Make gives you more control over conditional logic and error handling.
💡 PROMPT OF THE WEEK
Use this before you give AI full control over any business process.
I'm about to automate [describe the workflow].
Before I do, help me answer three questions:
1. What's the worst realistic thing that could go wrong if this runs without anyone checking it?
2. Which specific step in this workflow would benefit most from a human reviewing it before it continues?
3. What would a "pause and notify" step look like here? What information would I need to see to approve or reject it quickly?Tip: be honest about what "worst case" means for your business. The goal isn't to scare yourself out of automating. It's to make the automation worth trusting.
👋 THAT'S A WRAP
Full automation sounds like the goal. But for most business workflows, the goal is actually reliable automation, and those are two different things.
The best automations don't remove humans from the process entirely. They remove humans from the parts that don't need them, and keep humans in the loop where judgment actually matters. That's a harder pitch than "set it and forget it," but it's what actually works.
Take one workflow you're thinking about automating this week and run the mapping prompt above. You don't have to build anything yet. Just figure out where the judgment calls are before you decide how much to hand off.
If you work with someone who's been told to "just automate it" without much guidance on what that means in practice, forward this to them.
See you next Friday.