I Gave an AI Agent My Credit Card and Went to Sleep
When Clawbot launched, I felt the pull immediately. A proactive AI agent that runs in the background, finds things to do, and does them. No prompting. No scheduling. It watches your codebase, your tools, your workflows — and acts on its own.
The pitch was magnetic. I gave it a simple test task: I discussed movies I like, gave ratings, explained why I liked them, and set up a schedule for it to recommend new movies. Not a complex engineering problem. A straightforward task with a clear output.
I went to sleep. By morning, it had burned through $80 in API credits. This was not GPT-4 or Claude Opus pricing — this was a cheap model, DeepSeek-level pricing. Fractions of a cent per token. And it still blew through $80 doing nothing I could use.
The agent spent the night browsing websites about movies and adjacent topics — but it never saved any information, never produced a recommendation list, never reported back with results. It explored endlessly without delivering anything. Eighty dollars of compute spent on what amounted to aimless browsing.
Maybe this was a skill issue. Maybe with more tuning — clearer instructions, tighter constraints, output templates — it would have worked. I am honest about that. But that thought is exactly what led me to the real lesson: the tuning it needed — defined tasks, output rules, cost limits, approval gates — is the same work you do when building proper automation. The question became: why spend that effort on something unpredictable when I could spend it on something I control?
I shut it down the same morning.
The Slot Machine Problem
Here is what I expected from a "proactive" agent: that it would notice things I had not gotten to yet, pick the right moment, and handle them. A kind of background assistant that acts when it is convenient for me.
Here is what I got: an expensive slot machine.
The agent had no definition of "done." It had no way to know whether a task was worth doing at 3 AM or if it should wait for my input. It had no approval gates. It had no cost ceiling. It had no process — only permission to act and an API key with a balance.
The failure was not unusual. Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Real-world costs have been dramatic: one production multi-agent system reportedly ran an infinite conversation loop for 11 days straight, generating $47,000 in API charges before anyone noticed.
Autonomous AI agents cost more than scheduled workflows because they lack natural stopping points. Without defined tasks, approval checkpoints, and exit conditions, agents fill time with activity that looks productive but delivers nothing. A workflow running once daily at $0.03 will always beat an agent running overnight for $80 with nothing to show. As of Q2 2026.
This is the core problem the industry does not talk about honestly. "Autonomous" sounds like progress. It sounds like the AI is smarter, more capable, more advanced. But autonomy without guardrails is not intelligence. It is improvisation. And improvisation at API rates is expensive.
What Boring Automation Looks Like
After the Clawbot experiment, I put the same energy into something I could control. Not more autonomy — more structure.
My current setup runs on tools like Claude Cowork and n8n workflows connected to external systems — multiple Gmail accounts, HubSpot, other tools I use daily. Every morning, a scheduled workflow pulls unanswered emails across three accounts, drafts responses, and waits for me to approve before sending. Another workflow checks HubSpot for new contacts, enriches them, and asks me which ones to follow up with. The LLM does not decide what to do. It does what I told it to do, on the schedule I set, and asks before it touches anything external.
The architecture is the opposite of autonomous:
Scheduled. Each workflow has a defined trigger. Once a day, once a week, on a specific event. The system does not decide when to run. I decide.
Approval-gated. Before sending an email, before updating a record, before any action that affects something external — the system asks for approval. This takes seconds. The cost of that pause is negligible. The cost of the agent doing the wrong thing without asking is not.
Reports back. When the workflow finishes, I get a summary. What it did, what it found, what it skipped. No log-diving. No guessing.
This is boring. It is also reliable, cheap, and useful every single day. The effort I could have spent babysitting an autonomous agent — tuning its behavior, monitoring its spending, worrying about what it does while I sleep — goes into improving workflows I already trust.
Structured AI automation has three properties: scheduled triggers the human sets, approval gates before any external action, and a summary report when it finishes. Every morning I get draft email responses across three accounts and a HubSpot enrichment list — both waiting for approval, neither having touched anything yet. The LLM does work; the human controls what happens with it.
The AI reception systems we deploy for clinics follow the same pattern. They handle 70+ patient interactions monthly, operate 24/7, and respond in under 30 seconds. But they are not autonomous in the way Clawbot promised to be. Each follows a defined process: greet the patient, collect the reason for contact, check the schedule, confirm the booking, notify the clinic. Every step is mapped. Every edge case has a fallback. The AI handles the interface — the conversation with the patient — but the process is ours.
The LLM Is the Interface, Not the Decision-Maker
The difference between my $80 Clawbot night and my daily automation runs is not the model quality. It is not even the cost per token. It is the role the LLM plays.
An autonomous agent generates its own tasks and evaluates its own output — every decision weighted by training data and context that degrades over long sessions. A workflow automation has human-defined steps; the LLM handles execution within them but does not decide what happens next. The value of automation is predictability, not autonomy.
We learned this the hard way with our cold-email pipeline too. Four versions, and the fix was never smarter AI — it was better process. Observability, guardrails, cost tracking. The answer to "how do I get more from AI automation" is almost never "give the AI more autonomy." It is "give the AI more structure."
And this matters more every month. As I tracked in my analysis of AI pricing subsidies, the flat-rate era for AI compute is ending. According to The Register's coverage of Anthropic's April 2026 enterprise restructuring, usage-based billing is replacing subscriptions across the industry — and GitHub Copilot follows on June 1, 2026. In a usage-based world, every token matters. An autonomous agent with no cost controls is not a feature — it is an open tab at a bar where the drinks get more expensive every quarter.
Flat-rate AI subscriptions masked the true cost of unconstrained agents. As usage-based billing replaces subscriptions across the industry, every token an agent burns without a defined stopping point becomes a direct cost. A structured workflow has a predictable token budget; an autonomous agent does not. The risk compounds as metering tightens. As of Q2 2026.
The Test for Any Automation
Before you invest time in any AI automation — autonomous or otherwise — run it through three questions:
Can you predict what it will do? If you cannot describe the output before the system runs, you do not have automation. You have an experiment. Experiments are fine — label them as experiments, set a budget, and expect to throw away the output.
Can you set when it runs? If the system decides its own schedule, you lose control over resource usage, timing, and priority. The system should run when it fits your workflow, not when the model decides something looks interesting.
Can you approve before it acts? For any action that touches external systems — sending messages, updating records, spending money — a human approval step is not overhead. It is the difference between an assistant and a liability.
Test any AI automation against three questions: Can you predict its output? Can you set when it runs? Can you approve before it acts? If any answer is no, you have a prototype, not production. The gap is process design, not AI capability. Start with one task, one schedule, one approval gate. As of Q2 2026.
The Clawbot experiment cost $80 and gave back nothing. The lesson was not "agents are bad" — it was that autonomy without process is just improvisation at scale. Every dollar wasted on unconstrained browsing was a dollar that could have run a scheduled workflow with a defined output and a cost I could predict.
Proactive does not mean useful. Autonomous does not mean convenient. The best automation is the one that waits for you, runs when you tell it to, asks before it acts, and reports what it did. That is not a limitation of current AI. That is good system design.
Mind Momentum builds AI automation systems for healthcare clinics and service businesses — the boring, structured kind that runs every day without burning your budget overnight. If you want to talk through what controlled automation looks like for your workflow, get in touch.