Skip to content
Life With Data
Book a call

Systematize Before You Delegate to AI

  • #artificial-intelligence
  • ·#productivity
11 min
A small business owner reviewing research notes on a laptop before a sales call
Pre-call research: one annoying workflow, done well

TL;DR

Most AI efforts fizzle because people try to change everything at once. Even at big companies, 95% of AI experiments never move the bottom line. What works instead: pick one task that annoys you every week and improve it in stages. This post follows one example the whole way — researching a prospect before a sales call. The five stages are: do it by hand, ask AI ad hoc, give AI a repeatable recipe, keep score on quality, then hand it off to a you-plus-AI workflow. The payoff comes from getting the process right and knowing it works before you hand it to AI. The written-down, checked process is the valuable part. The chatbot is just the tool.

Introduction

You have a call at 10am tomorrow with a company you've never heard of. So tonight — or more likely at 9:40 tomorrow — you'll open eight browser tabs and start piecing together who they are, what they sell, and what the person you're meeting actually cares about.

That's fifteen or twenty minutes, every call. The more calls you book, the more time it eats. And it's the first thing to get skipped when the day gets away from you.

Most owners handle this one of two ways. Either they never improve it, and it stays a tax on every deal. Or they decide to "roll out AI" across the whole business, which becomes a project that never quite happens.

The second failure is better documented than you'd think. Even at big companies with real budgets, 95% of AI experiments never move the bottom line, and only about 5% drive real revenue (MIT, The GenAI Divide: State of AI in Business 2025). Companies are quietly giving up, too: 42% abandoned most of their AI projects in 2025, up from 17% the year before, and on average nearly half of AI experiments get scrapped before anyone uses them day to day (S&P Global). Gartner expects more than 40% of "AI agent" projects to be canceled by the end of 2027, mostly over cost, unclear value, or weak controls (Gartner, June 2025).

Those are enterprise numbers, but they fail for the same reason your ChatGPT experiment fizzled: there was no process behind it. Look at why these projects die and the causes rhyme — the work never connected to anything, nobody checked the output, and nothing improved after the first demo. That's a process problem, not an intelligence problem, which means it's fixable without a budget.

The alternative is unglamorous. Pick one task and walk it up a few clear stages. Pre-call research is a good first pick: it happens constantly, it's contained, nothing breaks if it's wrong, the payoff shows up immediately, and you can tell good work from bad at a glance.

The five stages (overview)

There are five stages. Each one takes a little upfront effort, then pays off every time you run the task.

The first jump — from doing it by hand to asking a chatbot — makes you faster once. It feels great, and it doesn't compound. The later stages are where the return lives, because the process turns into something you can reuse, check, and hand to someone (or something) else.

Notice who's doing the work. Through Stage 4, it's still you. What changes is that you write the process down in Stage 3, then start checking how well it's going in Stage 4. Only at Stage 5 do you redesign the task as a split between you and AI. That single, deliberate handoff is the point of the whole exercise.

The goal was never a faster chat session. It's a reliable process you can measure, improve, and hand off.

Stage 1 — Do it by hand

Right now it's a dozen browser tabs: LinkedIn, their website, Google News, maybe a look at who else works there. You copy the useful bits into a doc, then start from scratch on the next one.

There's something to be said for this. You have total control, and you come out knowing the account.

But it's slow, it's different every time, it lives entirely in your head, and the quality depends on how much time — and coffee — you had that morning.

You've outgrown it when you've done the same kind of lookup a dozen times and it still takes just as long.

Stage 2 — Ask AI ad hoc

Instead of tab-hopping, you open a chatbot and ask:

"I have a call with someone from Acme Manufacturing tomorrow. What do they do, and what should I know going in?"

This is much faster, and it's genuinely useful for poking around a company you know nothing about.

The problem is that it's different every time. You ask different questions depending on your mood, the format wanders, and unless you specifically demand sources, the AI will hand you a confident answer that's wrong — which you'll then repeat on the call.

You've outgrown it when two similar prospects produce two very different write-ups, and you couldn't hand your approach to anyone else if you tried.

Stage 3 — Give AI a repeatable recipe

This is the one that matters. You stop winging it and write the recipe once: a saved prompt that asks for the same things, in the same format, every time. This is the "systematize" in the title.

"You are my pre-call research assistant. Given a company and the person I'm meeting, produce a one-page brief with these sections, every time:

  1. Company snapshot (offering, size, recent developments)
  2. The individual (role, background, likely priorities)
  3. Relevance to what we offer
  4. Three qualified discovery questions
  5. Sources (cite every non-obvious claim; mark 'unverified' where you cannot) Cap at 400 words. Flag anything low-confidence."

Now it's consistent and fast, you could hand it to someone else, and it's more trustworthy — because it always demands sources and flags what it couldn't confirm.

You still run it yourself, and it's still worth a look before you trust it. But you own something now that you didn't before: a process rather than a habit.

This is the step most people skip, and everything after it depends on it. If you do nothing else from this post, get here.

Stage 4 — Keep score on quality

Now you run the recipe regularly and start keeping score — noting when the write-up was good and when it missed. You're still the one doing the work. No handing off to AI yet.

That order matters. You can't safely hand something to AI until you can tell whether it's being done well. Keeping score while you still run it gives you a known-good version to measure against later.

Keep this simple; no dashboards. A quick log of what each write-up got wrong or missed, and whether the call went well. A short checklist it has to pass: sources included, claims checked, questions actually useful. And one "something's off" trigger, like two write-ups in a row with a wrong fact that reached the call. That's the whole system.

The AI can do some of the checking for you:

"Review this brief. Verify claims 2 and 4 against the cited sources, flag anything that reads as an assumption, and list what you couldn't confirm so I can log it."

What you get is quality you can see instead of a gut feel. The notes you collect here become the instructions and limits you'll hand to AI in Stage 5.

It costs a few extra minutes per run, which is only worth it if you actually intend to hand off.

One thing to know: this scorekeeping is permanent. It doesn't stop at Stage 4. In Stage 5, it just watches the AI instead of you.

Stage 5 — Hand off to a you-plus-AI workflow

This isn't flipping a switch from "you" to "AI." It's a redesign.

You go through your recipe step by step and decide what each step needs: AI on its own, AI drafts and you approve, or you keep it entirely. What you learned keeping score in Stage 4 tells you where those lines go — by now you know which parts the AI gets right and which parts you always end up fixing.

It's not a clean swap, because AI fails differently than people do. A person rarely invents a source. AI will cite a page that doesn't exist, state a three-year-old fact as current, and sound completely confident doing it. The checklist you built watching yourself won't catch those on its own, so part of the redesign is adding checks for AI's particular mistakes — and deciding which steps you aren't willing to automate at all.

Keeping yourself involved is a choice, not a failure. A good split deliberately keeps you on the steps that need judgment, carry real stakes, or involve a relationship. That's what makes the output worth trusting.

A mix doesn't mean a small payoff, either. When it's working, a well-designed you-plus-AI split is a real step up from how you work today: faster, more consistent, and covering every call instead of only the ones you had time for. You were never chasing zero humans. You were chasing a better process.

Treat the AI like a new hire

There's a simple test for any step you want to hand off. If a new hire were doing this work, could they succeed with what you gave them?

Three questions make that concrete.

Does it have everything it needs to do the job? That's your recipe from Stage 3, plus context, examples, and access — onboarding docs, basically.

When and how should it stop and check with you? Those are your escalation rules, and they have to be explicit.

How will it know the job is done, and done well? That's your Stage 4 checklist, written down as a definition of "good."

If you can't answer those for a step, it isn't ready to hand off. That sends you back to Stages 3 and 4, which is where the work belongs anyway.

Improving it over time

You won't always have a tidy way to grade the AI, and that's fine. You improve by feedback: run it, see where it falls short of your checklist, fix the instructions, run it again. Each round tightens it up.

The tricky part is the specifics. Which steps to trust to AI. How to write the escalation rules. What "done well" means for your business. Those details separate a handoff that keeps paying off from one that quietly goes downhill.

That's also the part worth getting right the first time. A short working session to map the task, decide the you/AI split, and set the checks and escalation rules will save you from a half-automated mess. Book a discovery call if that's where you are.

Handing off is the reward for writing it down and checking it. It isn't a shortcut around either.

Takeaway

You didn't transform your business. You took one weekly annoyance and turned it into a mostly-automated process that beats how you did it before. That's the whole win, and it's a real one.

The same five stages work for chasing invoices, meeting recaps, first drafts of proposals, and weekly reports — anything you do over and over.

The rule of thumb is short: once a prompt works well twice, write it down, then start checking it. That's the line between a neat trick and something you can rely on.

So start this week. Pick the task that annoys you most and get it to Stage 3 — write the process down and run it the same way every time. When you're ready to turn it into a you-plus-AI workflow, book a discovery call and we'll map the task, set the split, and make sure it's built to last.

References

👋 Good read?

Subscribe for more insights on AI, data, and software.

One email when there's something worth reading. Unsubscribe any time.

Systematize Before You Delegate to AI