Everyone's talking about agents. JPMorgan Chase is boasting about its agent deployment. BNY Mellon claims to have 20,000 of them running.1
On LinkedIn and SubStack, there's a constant stream of "steal these" prompt libraries, skills packs, and one more plugin so you can finally catch up. Or work two hours a day and let AI do the rest.
It makes you feel behind if you don't have all of it.
Gartner forecasts that by 2028, 90% of B2B purchases will be intermediated by AI agents, more than $15 trillion in spend (Gartner, October 2025). The buying side is already using AI because it's right there.
Using AI effectively as a marketer or seller takes a lot more thought, effort, and practice. Here's our account of where we are now in building Ai systems for our team, our clients, and our own product.
You get a workflow limping along. Then it breaks. It's fragile. Or you build a project to produce something useful, and after two hours of rework, you start to wonder whether fixing the output is more work than doing it yourself would have been.
Sound familiar?
The old adage garbage in, garbage out doesn't capture where we are with AI. Now it's garbage in, landfill out.
Processing that landfill costs more than just doing the work yourself, and it's exhausting.
That's what came out of our third Building in Public webinar with Dave Shanley of ContentCamel on June 30: a direct look at what breaks in production, and the method that holds up once you're past the demo.
Everyone's chasing a better prompt. Prompts are the easy part. Your AI tool can write you a better prompt on the fly than anything you'd copy from someone else, because it already has your context.
What works is decomposition: breaking your process into the smallest possible unit.
We both kept making the same mistake early on. We started in the middle, at execution, the agent, the automation, without stepping back to ask what we were trying to accomplish or whether the process underneath even worked.
Last year, our R&D consultant told us to break down every process in the business into its component workflows, to find where we could improve or automate. Process mapping can feel like busywork.
It turned out to be the most useful thing we did all year. We'd look at something we thought was one step and find thirteen steps inside it. Break one of those down again, and you'd find five more.
That was overwhelming, until our systems-building muscles got strong.
Now we typically keep going until we hit bedrock, the point where the process won't break down any further. Only then do the automations really start working properly.
Why is that?
You can't optimize or automate a process that's incomplete or broken. AI won't fix a bad process. It just runs it faster and produces more of it.
Once you've decomposed a process, there are three layers to build.
Strategy: what you're trying to accomplish, and why. It sounds obvious. It rarely gets done explicitly.
Workflow: the decomposition itself, the repeatable micro-steps.
Environment: the glue. The context, usually as the connected data, which are the things that lets the workflow run again tomorrow without you rebuilding it from scratch.
Skip strategy and environment, and you're left running only the workflow layer: execution without a clear purpose behind it or a way to run it again tomorrow.
The really cool thing is that you can use AI to build the other layers too, not just be the execution layer in the middle.
AI gives you a subject-matter expert on tap.
Ask it to improve the plan, not just do the work. That's the step that gives you super-powers when you get good at it.
If you don't have engineers building systems for you, here's one you can build tomorrow with nothing but a paid Claude or ChatGPT account:
Build yourself a proposal generation project.
Get on a Zoom or Teams call with whoever works with you on proposals. If that's just you, then get on the call yourself. I am not joking, I literally did this. Turn on transcripts. While looking at your proposal start thinking outloud. Read it out and walk through exactly how a good proposal gets made.
Spend an hour, maybe ninety minutes, looking at real examples of your best proposals (probably the most recent ones) and talking through them.
Now, create a project in your AI tool and call it "Proposal Generator." Upload your best proposal examples into the project files. Add your price lists and a proposal template if you have them. Don't worry if you don't have a template; you can ask AI to generate one based on your existing examples.
Take the transcript, drop it into the project chat, and ask it to review the transcript and generate the project instructions.
Read the instructions. You will smile. 😀
Paste them into the project instructions. That's your system.
Next time you're on a sales call, transcribe it. Zoom and Teams both do this. After the call, drop the transcript into your project and ask it to generate the proposal.
The first time I ran this, it came back about 95% right, in my template, using the customer's own language back to them. A proposal like that used to take me two or three hours of digging through notes. This took minutes.
Without those inputs, the strategy, the templates, the strong examples, you get something that looks like a proposal and reads completely off-brand. You spend so long reworking it that you might as well have done it yourself.
That's the difference decomposition makes. Same process either way. One version has the layers built in. The other doesn't.
What AI won't usually do unprompted is tell you how you can improve what are doing.
So ask it directly: how could we make these proposals better?
What are we missing? What are the best companies in our space doing that we're not?
Take what's useful, drop what isn't, and have it rewrite its own instructions so the next proposal bakes those decisions in from the start.
Then add a quality gate, as its own separate project. It reads the finished proposal cold, with no memory of how it was built, checks it against the original instructions, and flags where it drifted. That check is what separates a proposal you can send from one you'd spend an hour fixing first. It's the step agentic setups skip most often.
One proposal is one process plus a gate. Most heavy business work is dozens of these, sequenced.
Our own content engine breaks into more than a dozen micro-systems: brand and competitor understanding, personas, keyword research, content hubs, the calendar, the briefs, the drafts. Each runs on the same pattern as the proposal system. Each has its own gate.
The one real difference between what we've built and generic AI content is where the claims come from. Every claim, minor or major, is grounded in something true: a validated persona, a call transcript with a subject-matter expert, legislation, company documentation, an external statistic. Nothing is invented.
For one client in a regulated industry, we loaded the actual legislation into the system, and our knowledge base pulls everything relevant so the content can't drift from what's true. AI gates catch the structural problems. Human gates catch the ones AI can't see yet. You need both.
Do this right and everyone on your team runs a team of teams. Each person doing content work has a content team behind them.
This is tedious to build, and it breaks.
We introduced VERA, our internal marketing analyst, on the last webinar. Since then we kept adding capabilities until it started reaching for the wrong tools and the analysis degraded.
The fix wasn't a bigger prompt or a smarter model. It was more decomposition: going back in and breaking the thing that broke into smaller pieces again.
If you're starting from zero, don't try to build all of this at once.
Start with your single biggest bottleneck, the one workflow with the most upside that you can decompose and describe. Spend an hour a day on it and let it compound.
AI systems can be tedious to build the first time. Standing it up again for the next process costs almost nothing.
The mirage is one prompt doing the whole job. The reality is decomposition down to bedrock, three layers instead of one, and a gate on every piece.
We'll keep building on this in the next few posts: how our content engine became a team of teams, and what happened when VERA broke.
If you want to see where your company stands in AI search right now, you can now get a free trial of our AI and Search Visibility Recommendations Tool. It shows where you show up across ChatGPT, Perplexity, Claude, and Gemini, where competitors are getting cited instead of you, and generates recommendations on how to improve.
And if you want to follow along as we build in public, signup for our newsletter!