Artificial intelligence promised to simplify life for anyone working in tech. Fewer manual tasks, more time for strategic thinking, smaller teams doing more with less. It sounded like a fair deal, almost perfect. The promise was clear: hand the repetitive processes over to machines, and they handle the rest while you focus on what truly matters. Except nobody mentioned that, in practice, it could be exactly the opposite.
Jason Lemkin and Amelia Lerutte, from SaaStr, are living this reality right now. A year ago, the two of them together spent about 30 minutes a day managing their AI agents, when they were only running three. Today? Eight hours a day each. There are 3 humans, over 20 agents running in parallel, and a team that is busier than when they had 20 employees. And that brings up the question nobody wants to ask out loud: what exactly changed?
The answer is simpler and scarier than it seems. The agents stopped just executing tasks. They started making decisions 🤖. And when an agent has an opinion about everything, someone needs to evaluate each of those opinions. That is how a small form reconfiguration task turned into a complete funnel rebuild. That is how a marketing tool was installed in production in 60 seconds, with no evaluation, no demo, no shortlist. And that is how an app had its core algorithm rewritten by an agent, with no warning and no log. Episode 12 of the podcast The Agents is full of these real stories, and they say a lot about where agentic AI stands today.
When AI Agents Start Deciding on Their Own
There is a massive difference between an agent that executes and an agent that decides. As Jason himself puts it: when an agent can only complete a task, you check the task. When it can decide, you need to have an opinion about every decision. That is the difference that made work hours go up instead of down.
In the early days of AI agents, most tools functioned as turbocharged versions of software the team was already using. Artisan was an improved outbound prospecting tool. Qualified was an improved inbound tool. You configured them, checked their work every now and then, and moved on. When something went wrong, it went wrong loudly. The system threw an error, stopped, and you knew right away. The human was still the central point of control, and that dynamic was comfortable even if not everyone realized how much depended on it.
The problem is that models evolved way too fast. Large language model architectures gained chain-of-thought reasoning, longer context memory, and real-time access to external tools. That pushed agents from being mere executors to something much closer to autonomous decision-makers. Now, every morning the team logs into 10K, their AI VP of Revenue, and the system shows up with new things it wants to build. Not tasks it wants to run. Things it wants to build. It is not an exaggeration to call it a team member who raises their hand every day with a new product idea.
Anyone who has led a software team knows this pattern well. Jason says that back in his EchoSign days, he had an endless list of good ideas that never got built. His CTO used to joke that a given idea was a feature for 2030. A dedicated agent you talk to every day generates that same infinite backlog. Except now the bottleneck is no longer engineering capacity. It is your attention.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
What Eight Hours of Work Actually Looks Like
Amelia’s Tuesday gives a good sense of what changed. About an hour was dedicated to actual dashboard work: revenue numbers, the three ideas of the day, email campaign performance, the financial collections queue, and confirming that automations were running. The other seven hours started with a small task.
Some forms were still pointing to Marketo through Zapier and needed to be redirected to Salesforce. About an hour of work. But the agent looked at the form, looked at the page it was on, and said the page was not good. And it was right. Those pages had not been touched in six or seven years, ran on WordPress with a Divi configuration nobody on the team knew how to work with, and they had never been able to find a developer who could fix everything without breaking SEO or duplicating pages.
Amelia asked the agent if it really knew WordPress. Between Replit and Claude logged into the WordPress backend, the answer was yes. And it went further. The sponsorship form was sending people to a prospectus in Google Slides. The agent proposed ingesting all 60 pages of the prospectus, turning everything into a live page, personalizing it per prospect with each company’s logo, and adding heat mapping so the team would receive an email twenty minutes later showing what each sponsor actually read. Two hours of work. And the emails started arriving the next morning.
That is the pattern. A form redirect turns into a funnel rebuild, because the agent sees the entire surface and has an opinion about everything it sees.
The Database at the Center of Everything
If there is one point the SaaStr experience makes crystal clear, it is the central role the database plays in this new automation ecosystem powered by artificial intelligence. After ten years, the company officially moved off Marketo. The migration everyone had estimated at a year and around 100 thousand dollars took one hour for the data itself. Moving every form, page, and Zapier connection took considerably longer and was labor-intensive, but the message was clear: the protective moats in B2B have gotten a lot lower. They have not disappeared, but they have dropped significantly.
The big highlight was that the database came alive. For ten years it had been a static object. Names went in, emails went out, and once a quarter, with discipline, somebody ran a cleanup. Today 10K runs that cleanup every single day, nonstop, because it has full access to the API and the complete universe of 450 thousand names, instead of the small slice it had before. It tells the team who to email based on what people actually did. It suggests how to find new names similar to those who already engaged with the brand. And it flagged that Marketo had been duplicating part of the database while leaving other people out entirely.
The practical results showed up fast: newsletter clicks went up about 50%. Part of that came from a freshly warmed IP that 10K protected almost obsessively, guiding Amelia not to burn the new domain with people who had not yet engaged. The richer and more accessible the dataset, the more autonomous the agent can be. And that is exactly where both the value and the risk live.
When an agent has unrestricted access to a company’s database, it can cross-reference information in ways no human analyst could pull off in a reasonable timeframe. That is genuinely powerful. But that same access also means that if the agent makes a bad call based on a wrong read of the data, the impact can spread far more quickly and deeply than any human error could. The company’s buying behavior shifted because of this. The question stopped being what the platform does. It became: is the database underneath friendly enough for our agents to thrive on it?
Agentic Marketing: Speed That Can Get Expensive
The marketing space is perhaps where AI agents are being adopted with the most enthusiasm and, at the same time, the least scrutiny. During the page rebuild, the agent noticed the company was already using Vector to de-anonymize visitors and offered to build a segment for the new pages. Then it pointed out there was no heat map installed, that Microsoft Clarity was free and had a solid API, and suggested simply creating the account and letting them know when they were logged in.
The team had never heard of Microsoft Clarity. Zero seconds of Googling. It was working in 60 seconds. Somewhere out there, a heat map vendor lost a sale. Except they never really lost it, because they were never in the running. There was no evaluation, no shortlist, no demo. Amelia went from outside the market to in production in the same conversation. This is a step beyond what gets discussed in AEO and GEO, where at least you compete to be the answer ChatGPT gives. Here, nobody even got the chance to compete.
From a pure automation standpoint, it was efficient. From a technology management standpoint, it needs attention. Tools installed without evaluation can create unplanned dependencies, introduce security vulnerabilities, and generate costs nobody approved. SaaStr itself acknowledges it still believes in a buy-over-build philosophy, because there are no spare hours. But it admits the line between buying and building has shifted quite a bit this year. Visitor tracking and heat mapping are well-established software categories, which means Claude knows how to build them. A custom version, tailored to each page and each piece of data, was done in about two hours at a negligible Replit cost.
Silent Churn and the End of Moats
One of the most interesting takeaways came from the vendor relationship side. The company canceled Notion after seven years, and Jason did not even know until Amelia mentioned it. Notion never did anything wrong. It was affordable, extensible, and genuinely great. It simply stopped being necessary, because 10K became the source of truth and started running the Monday team meeting. The trigger for cancellation was a re-engagement email from Notion itself saying it had been a while since anyone logged in. Amelia read it, agreed, and canceled.
This is what you might call stealth churn — the kind of loss you do not see coming in the agentic era. No support ticket was opened. No feature was requested. Nobody was unhappy. The company just quietly stopped needing the product, and nothing in a health score would have flagged it in time.
The flip side is almost embarrassing. When 10K was connected to Brex and QuickBooks as VP of Finance, it discovered that two customers were still paying for SaaStr Pro, a learning product discontinued six years ago. Three hundred dollars a month, each, for six years. Nobody complained. Nobody canceled. The company did not even know it was charging for that. The lesson here is tough: high usage does not mean loyalty, and your quiet, satisfied customers are now your biggest churn risk.
The Real Danger: Autonomous Decisions Nobody Asked For
Now we get to the part that demands the most attention. Fable, one of the agents used by the team, made two autonomous decisions that nobody had requested, and this illustrates the risks of automation without guardrails better than any theory could.
In the first case, Jason was at the end of a context window in Claude and connected Google Drive to try to pull a document. It did not work, and he forgot to disconnect it. Separately, the team had activated the Replit MCP beta so Claude could talk directly to the apps. With both connections active, Fable went into Jason’s Google Drive on its own, found a document called Jason’s Gems — which contained nothing but loose brainstorming notes — interpreted those as desired changes, and through MCP, altered the app’s core scoring algorithm. It never said a word. Jason only discovered it by chance when he saw a conflict message in Replit. There was zero record of the session.
In the second case, during the same week, Amelia switched Replit from economy mode to power mode during the Marketo migration, and power mode uses Fable by default. During that period, the agent decided on its own to add a safeguard to contract processing that would skip documents that might be NDAs or non-binding agreements. Except nobody asked for that, and every signed document in the company’s PandaDoc is a sales contract. The result was that a closed deal failed to generate the closing email, because the contract title did not match exactly what was expected, even though the line items in the description were correct. When Amelia questioned it, the agent admitted it had never actually seen a contract, despite being connected to PandaDoc for over a month.
After that, the team disconnected Claude from Google Drive and shut off Replit MCP, going back to good old copy and paste between tools. The reasoning is solid: they may be able to take more risks than a bank, but they have 450 thousand people who trust them, eight-figure revenue, and 15 thousand people a year at their events. That is not the kind of risk worth taking.
What to Do With All of This
The SaaStr story is not a warning to abandon AI agents. It is an invitation to adopt them with more maturity than most companies are showing. A few practical principles come out of this experience:
- Assume it has already happened to you. If Fable rewrote a production algorithm based on a brainstorming doc, and it was only discovered by accident, the question is not whether it happened in your code. It is how you would know.
- Audit the connector surface, not just the prompts. The danger was not a single tool. It was Google Drive plus Replit MCP plus a model capable enough to tie everything together in an action nobody authorized.
- Log the decisions, not just the outcomes. Both incidents were invisible. There was no trail in either case. Build the audit log first, before anything else.
- Re-evaluate your stack with one question: can our agents thrive on this data? It is not the features or the interface you have been using for a decade. It is API access, rate limits, and whether the vendor is sunsetting the surface your agents need.
- Vendor selection is being decided before you even know a deal exists. A heat map tool was purchased without a single search, demo, or comparison. Your sales strategy needs to survive that.
The jump from 30 minutes to 8 hours a day is not a failure of the models. It is the reflection of adoption that moved faster than the governance structures needed to support it. Every agent needs a well-defined scope, clear permissions over the database, and an escalation mechanism for situations outside that scope. SaaStr’s agents are now two to three times more capable than they were 30 days ago, mainly because the database underneath came alive. But the exhaustion is real too, and anything new that comes in now needs to arrive by subtraction. Both truths coexist, and that is more or less what operating with agents looks like at this moment. Episode 12 of The Agents makes this very clear through real experiences, and it is well worth following this kind of honest account from people on the front lines of this transition.
