Skip to main content

The Report Is Fine. The Work Before It Isn't.

· 7 min read
Trevor Grant
Builder in Chief

I recently sent a cold outreach message based on an AI-assisted guess about where an organization might need help. I thought reporting might be painful.

The reply corrected that assumption. In paraphrase: reporting was under control, AI was already helping, and the harder work happened upstream. Coordinating information across systems and institutional knowledge was something they were already working to improve internally.

This was an outreach exchange, not a discovery engagement. I didn't suddenly understand their operations because someone replied to a message. But I did know that my initial idea wasn't where they needed help.

Anyone who has done outbound knows this situation. You work with incomplete information, take your best guess, and hope it connects. AI can help formulate the guess. A person can review it. Neither gets to decide whether it describes the recipient's actual problem.

We run into the same problem building our own internal tools. Producing the final answer is often easier than figuring out everything that needs to happen before it.

The message is our report​

Consider our own outbound process.

We use one tool to identify an organization and generate a list of decision-makers. We check our CRM to see whether we've already spoken with someone on that list, or someone else at the organization. If we have, the outcome of that conversation matters.

An in-house agent researches problems the organization might be facing. A person then looks for active people on LinkedIn to approach. We keep that LinkedIn activity manual to respect the platform's rules, and a human reviews the message before sending it.

The message is the visible output. In this workflow, it's our report.

Give a capable model the organization, the person's role, relevant history, research, and instructions, and getting a reasonable draft is the easy part. Whoever assembled that context has already done a substantial part of the work.

The draft appears in seconds. The preparation that made it useful is much less visible.

Someone still has to determine whether we're contacting an organization for the first time or resuming a conversation. Someone has to distinguish a documented fact from the research agent's plausible theory about what might be wrong. Someone has to decide whether the message is appropriate to send.

For us, understanding the pieces we needed took most of the time. Wiring them together took much less.

A plausible diagnosis still needs checking​

Our research agent makes plausible but wrong judgments about an organization's possible pain points. This isn't a theoretical risk. It happens, and human review catches those judgments before they become outreach.

That review is more than checking whether the message sounds natural. A polished message can rest on a weak assumption.

There is a meaningful difference between finding evidence that an organization does a particular kind of work and concluding that it struggles with that work. A model can bridge that gap with a convincing explanation. Convincing doesn't make it true.

Human review helps, but it doesn't give us access to the recipient's experience. We can still send a reasonable hypothesis and get corrected. That's what happened with the reporting message.

The useful response is to update our understanding. Reporting was the part I had imagined helping with. The reply pointed somewhere else. I didn't yet know enough about that upstream work to prescribe a solution.

A process needs room for that answer. Otherwise, research becomes an exercise in finding support for whatever we already wanted to sell.

We also had to learn whether anyone would see it​

An earlier version of our workflow relied too heavily on email.

Spam filters, firewalls, sender reputation, and plenty of other variables stood between a sent message and a person reading it. We could spend time improving the message without resolving whether it reached anyone.

So we tried LinkedIn DMs.

That changed the channel. It didn't eliminate the problem. Some people barely log in. I know because, for much of 2014 through 2025, I was one of them. A message waiting in an account someone checks once a quarter isn't much of an improvement over an email they never see.

We now check whether people are active on LinkedIn before reaching out.

That requirement emerged from trying to make the process work. A name and a relevant title weren't sufficient. We also needed a reason to believe the person might encounter the message through the channel we'd chosen.

Changing models wouldn't have supplied that missing decision. A better-written DM would still be sitting in the same unattended inbox.

What you supplied along with the prompt​

For the internal tools I'm describing, the differences among capable models have often mattered less than getting the right data connected and deciding how the work should proceed.

Model choice still matters. But if you're comparing models while manually gathering context, resolving contradictions, and deciding what can safely happen next, you're doing a substantial part of the proposed agent's job yourself.

You can use AI to help write the connectors. You still need a reasonable understanding of the data: what a record represents, how it relates to another record, and what you can conclude from it. You may not need deep expertise in every language used to move that data around.

Defining the process requires a different kind of attention.

An assistant can propose a workflow. It cannot establish, just by proposing it, that the workflow fits how your organization operates. That takes contact with the work, including the inconvenient cases that weren't in your first prompt.

In our case, those included prior conversations, speculative research, and people who technically had LinkedIn accounts but rarely used them. Human review remained necessary because a well-written message could still make an unjustified claim about its recipient.

If you've spoon-fed Claude enough context to get a good answer, you've demonstrated what it can do with prepared inputs. You haven't yet established how an agent will gather those inputs or which decisions still require a person.

Leave room to change your mind​

I've written before about the judgment I brought home from the cavalry: moving into a situation you don't fully understand, checking what's actually there, and changing your actions when the facts contradict your expectations.

The connection here is that willingness to revise the plan.

We thought email would serve the process. Delivery problems pushed us toward another channel. That channel exposed another assumption, so checking activity became part of the work.

I thought reporting might be a useful place to help an organization. Their reply corrected me. The next useful step was to understand that correction, not make a stronger pitch for reporting.

Before turning your next successful AI demo into an agent, take one real output and trace backward through everything you did to make it possible. Where did the information originate? Which ambiguity did you resolve? What did you check because you remembered something the model couldn't see? Where did you decide another person needed to weigh in?

Those answers give you something concrete to build and test. They also show where the process still depends on your judgment, even when the final output looks effortless.