Five Things a Year of Putting AI Into My Own Business Taught Me

Everyone in marketing is being told to put AI into their business, and most of the advice skips the part where it goes wrong. Before I get into that, here is why I am the one telling you: I am Igor, I have spent about fifteen years buying paid media, I run Doctor Ads, and I have audited hundreds of advertising accounts. For the past year I have been testing AI systems across marketing and sales on every one of my own projects. Almost everything below is what happened to me rather than what I read, and I will flag the one place it is not.

The headline result first, because it is the reason I kept going. My team now produces the volume of work I would previously have needed twenty to twenty-five people for. Tasks that used to take twenty to thirty hours take about three.

The caveat second, and this is the borrowed part. MIT's State of AI in Business report for 2025 found that 95 percent of organisations running enterprise generative AI initiatives saw no measurable return, with the value concentrating in a small minority that integrated properly. I cannot tell you whether my own projects would clear their bar, because what I measured was hours and output volume and what they measured was profit and loss. What I can tell you is which five things made the difference on my side, and none of them were about which tool I bought.

1. Decide what the AI is for before you decide what to buy

The way I explain the job to my own team is with perfume. You have a mountain of rose petals, and what you need is a few drops of rose oil. That is the entire function: take an enormous, messy pile of input and distil it into something small enough to act on.

The projects I have watched stall, my own included, do the opposite. They generate more variants, more drafts, more reports, and now a human has to read all of it. If your AI initiative increases the amount of material somebody has to process, you have bought a petal machine.

Write down, in one sentence, what smaller thing you want out. If you cannot, you are not ready to choose a vendor.

2. Ask five or six models, never one

This is the habit that changed my output most, and the arithmetic is easy to see.

When I take a strategic question to a single model, it offers me about five approaches. Two or three of them do not survive scrutiny, and I am left with two. When I put the same question to five or six models separately, I get around thirty ideas. Twenty of them do not survive, and I am still left with ten.

Five times the surviving options, from the same question. The disagreement between the models is the useful part, because it shows me which part of my question was ambiguous and which claim only one model was willing to make.

3. Your own expertise is the gate, not a nice-to-have

Every time I follow an AI recommendation in an advertising account, I am correcting it as I go. If I were not an expert and simply added what it asked me to add, the campaign would either not work at all, or it would work but not in the way I wanted. What I actually do is give the model a bit of a head start and then fix the gross errors that would have led somewhere painful.

That is worth saying plainly because the pitch being made to business owners is the opposite one: that AI removes the need for the expertise. In my experience it does the reverse. It raises the value of knowing when the confident answer in front of you is wrong.

4. It fails hardest exactly where your data is biggest

I connected a Google Ads account to an agent-based system and watched it get confused by its own data because there was too much of it. This is the pattern I now expect: when models analyse large datasets they drift past the edge of what they actually know and start inserting things that do not exist.

I have a related rule I apply to anything numerical. I do not trust these systems when they analyse data, because I do not know which methods they are using, and there are a lot of possible methods. Until I can control how the mathematics is being done inside, I stay careful with numbers that come out of a model.

That is not scepticism for its own sake. It is that a hallucinated figure looks exactly like a real one in a slide.

5. Do not give it freedom, give it a corridor

The single highest-leverage skill I have picked up is not writing better prompts. It is finding, in advance, all the directions in which a model could reason wrongly, and closing each of them off explicitly.

Rather than letting the system think however it likes and hoping, I map where it can go astray and use the prompt to cut those branches, leaving a narrow corridor toward the result I need. It is more work up front than a one-line request and it is the reason my outputs are reproducible instead of lucky.

Final Thoughts on Putting AI Into a Business

The thread running through all five is the same. These systems are excellent at producing volume and poor at telling you which volume matters, and every failure I have listed comes from handing them the second job.

If you take one thing from this, take the sentence I now make everyone write before an AI project starts: which number in this business do we expect to move, and by when. MIT did not test that question, so I am not claiming it predicts anything. I am claiming that every project of mine that could not answer it turned out later to have had nothing to measure.