I Introduced Claude to ChatGPT. Then They Built My Prospect List
The assignment sounded simple: find companies we might be able to help, then find the people inside them who could use that help.
Then I watched Claude and ChatGPT talk it through.
They went back and forth over multiple turns, divided the work between Apollo (a B2B sales intelligence and prospecting platform) and LinkedIn Sales Navigator, and kept going without asking me to carry messages between them. They needed me once. When an action required Apollo credits, I was brought back into the conversation to approve the spend. I could have given them blanket permission, but I wanted to keep a human in the loop whenever money was involved. Once I said yes, the two of them went back to work.
That is the short version. The longer version starts with a confession that will sound familiar to a lot of agency owners. You know who you are.
I have spent most of my career helping clients improve their digital experiences and business systems while neglecting our own prospecting efforts. It was the agency version of “the shoemaker’s kids have no shoes.” After all these years of being very good at helping other companies, I am still reminded how difficult it can be to turn that same perspective inward. You are simply too close to your own business. An agency owner may be the worst client an agency could ask for.
For more than 25 years, word of mouth and referrals did the prospecting for me. Clients and projects seemed to appear almost magically whenever we needed them.
That is no longer a sustainable way to run an agency, so I finally gave our own prospecting the same deliberate attention we bring to client work and asked ChatGPT to help me approach the process more objectively.
We were using Apollo to identify manufacturers that might be a good fit for Impartium. We were not trying to generate the largest list possible. We wanted a focused group of companies that we genuinely believed could benefit from the ways we might help them.
Turning that sense of fit into searchable criteria was harder than it sounded. Company statistics and database filters helped narrow the field, but we cared more about whether a future relationship made sense for both Impartium and the prospect than whether a company checked every box. I supplied the business context while ChatGPT searched, evaluated the results, and helped refine the criteria. There was plenty of trial and error. Some companies fit the broad profile but made little sense once we looked closer.
If I’ve reached out to you recently on LinkedIn, now you know how you ended up on the list. With a little AI help, I personally reviewed your company and decided there was a strong chance we could be a useful resource for you.
Congratulations. You made the list. Go ahead and accept that invite.
At this stage, we were still deciding which companies belonged on the list. The people would come later.
At the same time, I was working with Claude through the Claude for Chrome extension, which allows it to interact with websites and perform actions inside the browser. Claude was using LinkedIn Sales Navigator to independently identify companies that might be a good fit.
I now had two AIs doing roughly the same work with no knowledge of the companies the other had found. That left me with two lists and a choice. I could pull the data out of both systems and deduplicate it myself, or I could see whether the two AIs could resolve the overlap together.
I chose the latter.
I added the ChatGPT tab to the Claude for Chrome extension, which gave Claude access to my ChatGPT conversation. I introduced them, defined their roles, and let them work through the handoff. They knew exactly who they were working with and referred to each other by name. At one point, ChatGPT told Claude, “Claude, I can pull that list for you as soon as Paul approves the credit spend.”
Once I approved it, ChatGPT provided the company list. Finding the right people inside those companies was more complicated. Impartium’s work spans both The Agency and The Bureau, so the work we might help with does not belong to one predictable department or job title. Depending on the company, responsibility might sit with digital experience, ecommerce, marketing technology, product data, business systems, AI, automation, operations, or somewhere between them. Claude used Sales Navigator to evaluate those possibilities, identify the people who appeared closest to the work, and add them as prospects for me.
I watched.
I could have kept the two efforts separate. I also could have started over and asked Claude to work directly with Apollo. But by the time I recognized the gap between Apollo and LinkedIn, I was already eyeballs deep in the process with ChatGPT. Connecting the two made more sense than throwing away the work we had already done.
It was also a lot more fun.
The handoff carried more than a list
A spreadsheet export could have given me the company names, but Sales Navigator does not offer a bulk import. I would have been left with the spreadsheet on one monitor and Sales Navigator on the other, manually searching for each company, finding the appropriate people, and adding them to prospect lists one at a time.
Connecting ChatGPT and Claude removed that manual bridge. ChatGPT could tell Claude which companies had made the cut. Using its browser access, Claude could then find those companies in Sales Navigator, evaluate contacts whose responsibilities aligned with the work of The Agency, The Bureau, or both, and add the strongest matches to the appropriate prospect lists.
The shared conversation gave Claude more than a set of names. It could see the business goal, the criteria we had refined, the false starts, and the judgment involved in deciding whether a company belonged. That context helped it understand the assignment, while its ability to act inside the browser turned the conversation into completed work. I did not have to reconstruct earlier decisions, translate between the tools, or work through the list manually one company at a time.
The technical term is multi-agent orchestration. But keep in mind that this was not a planned super feat of technology, nor are we claiming this is how AI should be developed. This was a fun experiment of having two AI models freely chat with each other in an unstructured interaction between the browser based LLM chat interfaces. This was not multiple well trained and tested agents designed to execute a complex workflow. I like to think of it as introducing two new coworkers who immediately became friends. They both wanted to help me, neither cared who controlled which territory, and within a few minutes they were working together as though that had been the plan all along.
If you can see several disconnected AI tasks inside one business process, we can help define where the handoffs and approval points belong.
Map an AI workflow →
I stayed responsible without steering every turn
Keeping a human in the loop did not require me to supervise every message. Most of my responsibility sat at the boundaries of the work. I helped define what mutual fit meant before the search began, decided that spending required approval, watched the exchange, and retained the final decision about what would happen with the resulting contacts.
I kept the experiment deliberately narrow. Claude helped with research and prospect list-building while I watched the work and reviewed the resulting contacts. Any invitations or messages that followed were handled manually by me.
Nothing was sent during the experiment, and no outreach sequence was launched. It produced research and a prospect list, while the consequential action remained with me.
That division of responsibility felt practical. The tools handled more of the searching, transferring, and organizing. My attention shifted toward the decisions only I should own: whether a company was truly a fit, whether to spend money, and whether another person should eventually hear from us.
An informal experiment can reveal the workflow
This was improvised. I had two AI tools, two business platforms, a live task, and the strange pleasure of watching my new coworkers figure each other out in real time.
A formal version would need clearer operating rules. The criteria would need to be documented, permissions defined, and important actions recorded. Paid activity, changes to business data, and outbound communication would each need an explicit approval policy.
Those requirements are part of the distance between a promising experiment and a dependable production system, a distinction we explored in A Great AI Demo Can Still Make a Terrible Production System.
The informal version showed me what a formal workflow would need before anyone started designing one. I could see which responsibilities naturally belonged together, what context had to move between tools, and where I wanted the work to stop for review.
That is often the right place to begin. A real process exposes useful boundaries more clearly than a diagram built around hypothetical AI capabilities.
Cooperation was more useful than comparison
AI conversations often turn into model comparisons. Which one reasons better? Which one browses better? Which one should a company standardize on?
My prospecting experiment suggested a more useful question: What can each tool contribute to the work? ChatGPT had accumulated the research and reasoning behind the company list. Claude could operate inside the browser and continue the process in Sales Navigator. Their value came from the combination of context, access, and a shared objective.
The same principle applies beyond prospecting. One AI tool might extract information from an email, another might evaluate it against business rules, and a third might prepare an update inside the system where employees already work. A person can remain responsible for approvals without becoming the courier who manually carries every piece of information between them.
The prospect list was the immediate result. The experiment also gave me a process I can now examine, improve, and potentially formalize around the decisions I still want to own.
If work in your business already moves among several people, systems, and AI tools, we can help shape those pieces into a practical workflow with clear human control.
Explore a multi-agent workflow →
