Where to Start with AI (When You Don’t Know Where to Start)
Most AI advice starts with the tools. The right starting point is an honest look at where your best people are spending their time, and what’s actually consuming it.
There is plenty of advice about what AI could do for your industry. The harder question is where it can help your business, given your team, budget, and priorities.
Meanwhile, you’re sitting in the chair where the decision actually has to be made. Your team probably bought ChatGPT licenses at some point, a few people use it regularly, most don’t, and nothing about how the company actually operates has changed. So the honest question isn’t whether AI matters. It’s where on earth you’re supposed to start when every option seems to require expertise you don’t have and a budget you can’t justify yet.
A practical answer starts with the work, before choosing the tools.
Start with a Strategy Audit
The instinct when you decide to "do AI" is to buy something: a platform, a license, a chatbot for the website. That instinct is understandable, and it’s also where most companies end up with a stack of subscriptions nobody uses and a budget line that’s hard to defend at the next planning meeting.
The right first move is a strategy audit of your people’s time. Not your technology stack. You’re trying to answer one question: where is senior capacity being consumed by repeatable work that doesn’t actually require senior judgment?
Every knowledge-work company has this pattern. A senior team member spends three hours formatting a proposal that follows the same structure every single time. A director re-reads the same five industry sources for every competitive analysis. An account manager drafts client update emails that are nearly identical week to week. This isn’t the edge of the workday. It’s the middle of it, and it’s happening at your most expensive hourly rates.
McKinsey estimates knowledge workers spend 9.3 hours a week just searching for and gathering information. Whether your revenue model is billable hours, project fees, or subscriptions, that time is margin you’re paying for but not capturing. The strategy audit finds those hours, ranks them by how repeatable and low-risk they are, and surfaces the ones where AI augmentation gets you the most value with the least exposure.
That’s where you start: the task your best people do too often, that follows a recognizable pattern, and that an AI can get meaningfully right on the first pass. The exact accuracy varies by task and by how the system is configured; some workflows land at 70% on day one, others at 90%. The point is that starting from something is a fundamentally different experience than starting from nothing.
Start with Experienced Reviewers
A useful starting point is this: start with people who understand the work well enough to assess the result.
Your best people already know what good output looks like, so when an AI gives them a strong first draft, they can identify what needs fixing in a few minutes. They know the context, the quality bar, and the standards. They review, correct, and approve faster than anyone else on the team, which means the efficiency multiplier is highest with them.
A senior team member who used to spend four hours on a competitive analysis might now spend 45 minutes: some time setting up the request, a few minutes waiting on the draft, and the rest reviewing, sharpening, and adding the judgment only they can provide. The exact time savings vary by task, by person, and by how the system is configured, so the improvement should be measured against the current process.
The goal is to give experienced people more time for judgment, client work, and decisions that need their expertise.
Give the same tool to someone without the judgment to evaluate the output, though, and you get a very different result. They accept the draft as final, the errors stay in, and the quality control burden ends up landing on the exact people you were trying to free up.
Start with the people who know the work best. Let them validate the approach, establish what good AI-assisted output looks like at your company, and set the standard everyone else will work toward. Then expand from there.
Why Starting from Something Changes Everything
The math shift most leaders miss is this: an AI that gets you most of the way to a finished deliverable isn’t just "faster." It’s a fundamentally different starting point.
Think about what happens when someone on your team sits down to write a market analysis from scratch. They open a blank document, search for data sources, read industry reports, structure their argument, draft something, decide it’s wrong, and start over. The blank page is the hardest part, not because the person lacks expertise, but because getting from zero to something coherent is largely undifferentiated effort.
Now consider the alternative: the AI generates a structured first draft, formatted in your template, drawing from your company’s knowledge base. It’s not perfect: the conclusions need sharpening, one data point is off, and a section wants reordering. But the person isn’t starting from zero. They’re starting well into the process, and that changes the whole economics of the task.
The first stretch of any knowledge work task is where most of the time goes: the research, the structuring, the formatting, the assembly of raw material into something coherent. That’s the undifferentiated part. The final stretch is where the value lives: the judgment call, the strategic recommendation, the insight that demonstrates you understand the situation better than anyone else in the room. How far the AI gets you on the first pass depends on the task, the person, and how the system is configured, and we benchmark that number for every workflow before making promises. But the principle is consistent: eliminate the blank page, and you eliminate the part of the work that was never worth paying senior rates for.
Keeping an experienced reviewer involved can make practical and financial sense for work that requires judgment. The human does the work that justifies the price, the AI handles what was eating the hours, and the output quality stays the same while the margin improves.
What Most People Get Wrong About AI
Before making any decisions about where to deploy AI in your company, it’s worth addressing three things most executives get wrong about how it actually works. Understanding these limits helps leaders choose tools, set expectations, and avoid unnecessary frustration.
AI tools may retain selected details through memory features, but that is different from maintaining a complete, current understanding of your business. Critical instructions and company information still need to be supplied and checked. ChatGPT’s memory documentation describes how saved information and past chats can inform responses.
Every model has limits on how much information it can consider in a request. A larger context window does not ensure that every relevant detail will be used correctly. A business workflow needs a way to select the right records and instructions for the task, then check the result.
AI model providers update their models regularly, and when the underlying model changes, your outputs change too. This is called model drift, and it happens without warning. A workflow that produced excellent output last month might produce noticeably different results after a quiet model update. Without a system that monitors output quality and flags when performance shifts, drift goes undetected until someone notices the output is wrong, and that someone is usually a client.
These three considerations matter when moving from individual use to a shared business workflow. Access to the tool is a starting point; reliable delivery also needs clear context, review, and ownership.
Sustained value depends on providing relevant context, checking output quality, and improving the workflow as needs change.
The ROI Conversation That Actually Matters
Executives push back on AI investment for a predictable set of reasons, and most of them are genuinely reasonable: the cost is uncertain, the timeline to value isn’t clear, and the risk of a visible failure is real. Here’s how the math actually looks when you start the way we’ve described.
Consider a simplified example: a 40-person company where 15 senior people are each spending roughly 8 hours a week on repeatable tasks: proposal drafting, competitive research, client reporting, internal documentation. That’s 120 hours of senior capacity every week consumed by work that doesn’t actually require their best judgment.
A human-in-the-loop system that reduces the time spent on those tasks by even 50% frees up 60 hours a week of senior capacity. The exact number depends on the tasks, the people, and how well the system is configured, and we benchmark every workflow before projecting a return, and results vary meaningfully by company. But the objective is to recover senior time for work that benefits the business.
The business case should compare implementation, review, and operating costs with the value of time recovered. Starting with a focused workflow and human review helps make that comparison concrete before expanding the investment.
Companies that struggle to justify AI investment are almost always trying to justify the wrong thing. They’re trying to calculate the ROI of replacing a person, which is a high-risk, politically charged argument that rarely gets approved and rarely should. The easier argument, and the correct one, is the ROI of giving your existing people a meaningful multiplier on the repeatable work that’s been consuming them.
Expand from the Work You Know
Once the first workflows are running and your experienced reviewers have validated the approach, the obvious question is what to do next. Our answer is always the same: keep expanding the augmentation. Don’t pivot to replacement.
- High upfront cost, long implementation timelines
- Significant business risk if the AI makes errors
- Requires the AI to handle edge cases and judgment calls
- Internal resistance from the team ("they’re replacing us")
- Months before you know if it works
- Start with a focused workflow and measure time saved
- Human review helps manage errors before output is used
- AI handles the repeatable portion; your people handle the judgment
- Involve the team in testing and refining the workflow
- Compare results with a baseline as the pilot runs
After the initial workflows are running, you audit the next set of repeatable tasks: onboarding documents, status reporting, research synthesis, pricing analysis. Each one gets the same treatment: the AI generates a structured first draft, a human reviews and refines it, and the output goes to whoever needs it.
As the augmentation spreads across more tasks and more people, you start accumulating something valuable: data on where the AI performs well and where it falls short. That data is what lets you tune the system: adjust the instructions, refine what context it receives, improve the templates it’s working from. Each optimization cycle moves the starting line forward, and the human review gets faster because there’s less to fix.
The work continues after launch. Review where the AI performs well, where people correct it, and what those corrections reveal. Use that evidence to improve the instructions, information, and checks rather than assuming accuracy will rise on its own.
Apply that review process to existing workflows as you add new ones. Compare the AI’s drafts with the corrections people make, then use the findings to improve the next version.
A useful first workflow gives you something to measure, improve, and build on.
At Olytic, we build around the way your team works and use its feedback to improve the system. Reviewing corrections and checking results remain part of the work after launch.
Start with a task your team can review and a result it can measure. That gives you evidence for deciding whether to expand, adjust the approach, or stop.
Smaller Than You Think
You can begin with one recurring task and the people who know it well. Agree on what a useful result would look like, try the approach, and assess the work before committing further.
But the starting point is smaller than most people think. You don’t need a company-wide AI strategy, a Head of AI, or a board presentation. A focused pilot can often use existing systems, with training tailored to the people doing the work.
You need to identify the three to five repeatable tasks consuming your best people’s time, put an AI system in front of those tasks with a human reviewing every output, and measure what happens. That’s the strategy audit, and that’s where you start. Everything else builds from there.
The companies that get this right won’t win because they bought the best tool. They’ll win because they started in the right place, expanded methodically, and built something that gets better every week it runs.