# Scope the human work first Source: https://revductive.com/blog/scope-the-human-work-first Summary: Most AI projects turn out to be process problems. How to scope what a person actually does, which chunk to automate, and what to leave alone. Every page, and the topic files: https://revductive.com/llms.txt [Blog](https://revductive.com/blog) AI AI6 min readAug 2026 Most AI projects turn out to be process problems. How to scope what a person actually does, which chunk to automate, and what to leave alone. Kevin Stout Founder, Revductive [Share on LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Frevductive.com%2Fblog%2Fscope-the-human-work-first) Contents 1. Why everybody skips it 2. What happens when you skip it 3. What scoping actually looks like 4. What I won’t automate 5. A real example 6. One more thing, before any of this 7. The gut check ![Four pairs of hands around a table laying out black and paper cards, with a translucent orange band across the middle carrying a single routed line between four marked points.](https://revductive.com/_astro/scope-the-human-work-first.DVQVmPdH_Z2gIgxn.webp) Most of the AI work people bring me ends up being a process problem. Somebody calls because they want to automate their sales handoff, or onboarding, or support triage. Then a few minutes in it’s pretty clear nobody there can even describe how the process works today. That’s most of what’s going wrong right now. People want to drop AI into a role, or into a process if they’re being smarter about it, but nobody ever wrote down what actually happens in that process. So there’s nothing to automate yet. Half the time it’s [four people doing the same job four different ways](https://revductive.com/blog/ai-gave-you-50-silos-instead-of-5). > “You have to understand the actual work that the human does in the first place, which I feel like is the thing people tried to skip.” ## Why everybody skips it I get why people skip it. Scoping the human work is boring and it’s slow, and it means sitting on a call with a rep for an hour asking the same annoying question six different ways and making them pull up real examples on their screen, and none of that demos well or goes in a deck for anybody. The tools also get sold like you don’t have to do any of it. Connect your data and it just works, right? So people connect their data and it doesn’t really work, and then they figure the model isn’t smart enough yet. The model is usually fine. Nobody ever sat down and scoped the work, and I think that part gets skipped because it doesn’t feel like you’re getting anything done while you’re doing it. ## What happens when you skip it The AI SDR stuff is probably the clearest example I’ve got. SDR just means the person doing cold outreach and booking meetings for sales, if that’s not a term you use every day. > “All these AI SDR tools came out and, like, being built by people that have never been SDRs and have never done the SDR work itself. And they’re just like, oh, but it’s an AI SDR, and it’s just doing all this stuff, and realistically not doing the valuable part of the work at all.” When you’ve never done the job, all you see is the part that’s visible. An SDR sends emails, so an AI SDR sends emails, and now you’ve automated the typing. But typing was never what made a good SDR good. The good ones read a company, figure out what’s broken over there this quarter, and then pick which of the forty accounts on their list is worth a real shot. Those tools copied the emails and skipped all of that, and then everybody acted surprised when reply rates dropped. That’s a scoping problem. Nobody building those tools ever did the job, so they guessed at it. ## What scoping actually looks like This is roughly how I do it: **1. Interview one person who does the job.** Not their manager or reviewing a documented process sitting in a wiki, but the person who does it every day. Ask them how it works, then keep asking for examples as you go. Have them share their screen and walk you through the last few times they did it, because what people tell you they do and what they actually do are usually not the same. You’re looking for the stuff nobody wrote down, like the copy paste between two tabs, or the spreadsheet they keep on the side because the real system won’t do the one thing they need. **2. Find the repetitive twenty percent.** Pretty much every role has some chunk of the week that’s the same action over and over with no judgment in it at all, and that chunk is what you’re aiming at rather than the whole role. **3. Build for that chunk and leave the rest of it alone.** If you take the admin work off an SDR, they get more hours to go fill a calendar, which is what you hired them to do in the first place. **4. Then be honest about whether AI should be touching it at all.** I feel like this is where people are weakest. > “Making the right call as to whether it’s ready to handle a particular thing or not is not something I feel like many people are doing particularly well. They just want to do it all.” Knowing what to leave to a person is most of the skill right now, and I don’t think a lot of people are being honest with themselves about it. ## What I won’t automate SDR, onboarding and customer success. Those are the first real human touches somebody gets with your company, and I’m not handing that off to a bot to save a headcount, even if the tooling could technically pull it off. Strip the admin out of all three of them though. Logging the call, updating the record, drafting the follow up, pulling account history before a meeting, that stuff is fair game. I just want a person on the actual conversation. ## A real example I [ran an audit](https://revductive.com/ai/audit) for a company’s CS team a while back. CS couldn’t do any real prep before a churn save call, because everything sales learned during the deal stayed with sales. The fix that had been sitting in their queue for months was to [add a bunch of required fields in Salesforce](https://revductive.com/blog/stop-adding-required-fields-automate-the-transcript) so reps would document the handoff. That request was going to sit there forever, and it probably should. Look at what it’s asking for. The rep has to remember, the rep has to type, ops has to build the fields and then keep them alive, somebody has to police whether they’re getting filled out, and after all that CS gets whatever the rep typed into the box. So we backed up and looked at what the rep actually does. They sit through a fifty minute call where the customer tells them everything, and that call is already recorded and transcribed. The information was already there. It was just sitting in a tool nobody in CS ever opens. So we pulled the transcripts, ran them through an extraction step, and wrote a summary onto the account record. The rep doesn’t have to do anything, ops doesn’t have to maintain anything, and CS gets something better than a typed note anyway. That’s usually what happens when you scope the work first. You end up building something pretty different from what you walked in planning to build, and it’s usually cheaper and it ships sooner. ## One more thing, before any of this There’s a step before all of this that comes up constantly. > “We need to get people to use just Claude a little bit better and stop sending around AI garbage to everybody all day before we start automating anything.” A lot of companies are trying to build agents on top of [a team that hasn’t really figured out the chatbot yet](https://revductive.com/blog/theres-no-pattern-to-who-picks-up-ai). You handed out seats, nobody got trained on anything, and now people are pasting bad output into real client emails. Automating on top of that just moves the mess around faster. None of that is hard to fix either, it’s mostly [training and a few internal examples of what good output looks like](https://revductive.com/ai), but it has to happen before you start automating anything on top of it. ## The gut check Pick whatever process you most want to automate right now. Can you write down every step somebody takes, in order, including the parts that live in Slack and in that spreadsheet on the side? Not the org chart version of it, the real one. If you can’t get that written down, that’s your project. And I’d say about half the time, once it actually is written down, you find out the fix doesn’t even need AI. I’m always available to brainstorm on this process with anyone that needs it. [Book a call with me](https://revductive.com/contact) and let’s figure it out. ## Want a look at your setup? 30 minutes, free. Leave with a recommended tool stack, no matter what. [Book a 30-min call →](https://revductive.com/contact) ## Common questions Why do AI projects fail? Most of the time nobody wrote down what actually happens in the process first, so there is nothing concrete to automate. The model is usually fine. What is missing is a description of the real work, including the parts that live in Slack and in somebody's side spreadsheet rather than in the documented process. How do you scope a process before automating it? Interview one person who actually does the job, not their manager and not the wiki. Have them share their screen and walk through the last few times they did it. Find the repetitive chunk of the week that has no judgment in it. Build for that chunk and leave the rest alone. Then be honest about whether AI should touch it at all. What should you not automate with AI? SDR, onboarding and customer success conversations. Those are the first real human touches somebody has with your company. Strip the admin out of all three, so logging the call, updating the record, drafting the follow up and pulling account history before a meeting, but keep a person on the actual conversation. Why did AI SDR tools disappoint? They automated the visible part of the job, which was sending emails, and skipped the part that made a good SDR good. That is reading a company, working out what is broken there this quarter, and picking which accounts are worth a real shot. Copying the output and skipping the thinking is a scoping problem, not a technology problem. What has to happen before automating anything? The team needs a working floor with the basic tools first. A lot of companies are building agents on top of people who have not been trained on the chatbot yet, and automating on top of that just moves the mess around faster. It is mostly training and a few internal examples of what good output looks like. ## Related AI ### [Using AI as the execution layer, not an advisor](https://revductive.com/blog/using-ai-as-the-execution-layer) Most teams use AI to ask questions and then do the work by hand. Hand it the work instead, and write the spec before you do. [Read →](https://revductive.com/blog/using-ai-as-the-execution-layer) AI ### [llms.txt is bullshit. Do it anyway.](https://revductive.com/blog/llms-txt-is-bullshit-do-it-anyway) 97% of llms.txt files never get requested and Google ignores them. I built one anyway, found AI was reading a third of my site, and rebuilt it so it isn't. [Read →](https://revductive.com/blog/llms-txt-is-bullshit-do-it-anyway) RevOps ### [You probably don't need to upgrade HubSpot yet](https://revductive.com/blog/avoid-hubspot-costs-early-stage) Most upgrades happen because somebody needed one feature. What HubSpot locks, what to use instead, and when to just go buy it. [Read →](https://revductive.com/blog/avoid-hubspot-costs-early-stage)