How We're Building Effective AI Agent Pilots in 30 Days
One of our partners forwarded me an email recently from a business owner in his tech network group. A business owner named Kyle was looking for help with AI. His message was the refreshingly honest summary of what so many companies are facing right now. Paraphrasing, he said:
"I'll be upfront that my knowledge in this space is limited and I don't always know the right terminology, but we have very real use cases and problems we are trying to solve with generative AI. We are not looking for hype or experiments. We want to understand what is actually possible today and what makes sense to build in a practical, scalable way."
This is so common. I've had some version of this conversation a dozen times in the past few months. Curious people leading companies into whatever comes next, skeptical of the hype, but worried they're falling behind. Many have heard the promises of what these tools are capable of. They've also heard about the failures. They just want someone to be straight with them.
So, let's do that.
The growing gap
There's a 6x productivity gap between power users and everyone else according to OpenAI's recent enterprise report. The people who have figured out how to use these tools are working faster but, as importantly, also doing things that were previously impossible for them.
But it goes beyond individual performance. It's very easy to tie adoption back to whether organizations have done the work to make AI useful to their teams. That includes training on off-the-shelf tools like ChatGPT Enterprise (or Claude, Copilot, Gemini—whichever makes sense for your organization), but also extends to custom pilots to leverage the technology beyond those accessible front doors.
MIT's research found that 95% of enterprise AI pilots fail to deliver measurable returns. BCG says 74% of companies can't move beyond proof-of-concept. The industry is calling this pilot purgatory. Companies run experiments. The experiments don't scale. Leadership loses confidence. Everyone goes back to waiting for the technology to mature.
The real problem is that they give up.
Failure is the signal
I think most companies have this backwards.
The ones winning at AI aren't the ones avoiding failure. They're the ones failing more, cataloging what went wrong, and building a library of test cases for when the next model drops.
Every speaker I've listened to who's built practical AI solutions at scale had a long list of failures to share. That's the signal I've come to find most valuable. If you're not finding where AI falls short, you're not pushing the models far enough.
This matters because the failures tend to get fixed with new models. The latest batch of models—GPT 5.2, Opus 4.5, and Gemini 3 Pro—can do things the past set of models couldn't. They're better at following instructions, some hallucinate at much lower rates, and when using the models through coding harnesses like Codex and Claude Code, they can build features and full-scale products that weren't possible before.
So the companies that know exactly what to retest when a new model releases hit the ground running. Everyone else is still experimenting.
Waiting for AI to be "ready" means falling behind companies that are learning through failure right now.
What actually works today
All of that said, I'll admit that the hype is ahead of the reality.
You may have heard about AI in the org chart, filling full positions. Or companies that have completely automated their growth teams. Or promises of swarms of agents handling complex work flows end-to-end. Directionally, yes, that's where we're headed. But right now, what works (and can be proven to work) is focused agents doing specific tasks, matched to your actual workflows.
That's not insignificant. A well-designed automation can save hours every week, sometimes days. Multiply that across a team and the impact compounds fast. But we have to be practical about what's possible.
The off-the-shelf tools like ChatGPT Enterprise, Copilot, and Claude hit walls. They help individuals become more productive, especially the power users who spend time learning the full set of features available, like Custom GPTs, Deep Research, and Claude Cowork. But they're limited in what data they can access, what IT will allow, and how many steps they can string together. They're great for going to the tool to do something. They're not great at taking action on your behalf.
Model quality and specialization is often overlooked. Claude is currently better than ChatGPT for ~90% of day-to-day tasks because it feels more like a human who understands how work gets done. But ChatGPT's 5.2 Pro is the most intelligent model for heavy lifting data analysis, coding, and math.
The problem is that you're often locked into one solution. You either get Claude's more human feel or ChatGPT's robotic intelligence.
Custom solutions can fix this problem. We can connect to your data, automate multi-step processes, and build agents that leverage the right models at the right time to best complete the task. But it takes work to get there. It takes understanding your workflows first.
What the hell is an agent?
Another buzzword? Yes, but let's go ahead and define it so we're speaking the same language.
When we say AI agent, we mean a system that can plan and act independently to achieve a goal while operating within guardrails set by humans. For the capabilities of AI agents today, we limit the number of decisions they need to make about what to do next, but that expanded definition of agents is on the near-term horizon as models get better at holding context across sessions.
Here's an example of a lightweight agent we recently built, an AI-powered interview tool for content creation.
The problem: a marketing team needs content from domain experts inside the company. But the marketing team doesn't have the expertise to ask the right questions. The resident expert doesn't want to take the time to do the marketing team's annoying interview. They don't speak the same language as the engineers or the operations people or whoever has the knowledge they need.
So we built an agent that conducts the interview. It studies the topic beforehand (a mix of utilizing its training data and web search). It asks questions like a peer would. The domain expert talks to it like they're talking to another expert, using text or voice, not dumbing things down for marketing.
The output captured the golden nuggets from the interview. This is the stuff that actually matters in content creation. The stuff buyers actually want to know. Then, the marketing team can take these drafts and edit them into final content, fine-tuned to the brand voice.
Then, when the piece is finalized, the AI agent turned the content into many different use cases, including an article, social media posts, newsletter content, and even internal training content.
The result is higher quality and faster. Not because AI wrote the content, although it does help. AI-generated content isn't inherently bad, slop writing is bad (which comes from both humans and robots). What makes this work is capturing the unique perspective of someone who deeply understands the subject.
What it takes to make this work
Most companies that struggle with AI are trying to skip steps.
They want to jump straight to automation without understanding their workflows. They want AI to fix processes that are already broken. They buy tools and expect them to work out of the box.
What actually works is more methodical:
- Map the workflows first
- Identify where time is being wasted, where errors happen, where the bottlenecks are
- Brainstorm what tools could help
- Build rapid prototypes to test whether those ideas actually work in practice
That's what we do through our Implementation Launchpad. It's a 30-day sprint that takes companies from curious about AI to production-ready tools to test.
The process combines a few things that usually require separate partners. Habitat brings strategic planning and organizational design—the ability to understand how teams actually work and where the friction is. Mostly Serious brings technical skills to build the solutions. And we've spent the past four years going deep on the AI ecosystem, learning what's possible and what's hype. We know, because we're building these same solutions for ourselves.
If your data isn't ready, we can help with that too. The answer isn't to wait until everything is perfect. It's to start with what you have and improve the infrastructure alongside the pilots.
The pilot program offer
We are launching the Implementation Launchpad to help companies dive into this space. The first 5 companies will get a fixed-fee engagement. $15,000 to $40,000 depending on complexity of the agents we build. For that, you'll get:
- A single-day intensive discovery sprint with your leadership team
- Three working AI pilots tailored to your workflows
- A 90-day integration roadmap
- Hands-on training so your team can maintain and extend what we build
Four weeks from kickoff to production. No scope creep, no surprises.
If you're like Kyle—ready to move but not sure where to start—we would love to help. We'll figure out together whether this is the right fit, or whether something simpler like team training makes more sense as a first step.
The companies that figure this out aren't waiting for the technology to mature. They're learning now, failing now, and building the muscle to move fast when the next wave hits.