I have been building with AI for about two years now. Long enough to stop believing either headline.

You know the two. "Company X lays off thousands, replaced by AI." And then a few months later, "Companies regret AI layoffs, start rehiring." Both keep showing up. And here is the strange part. Both are true. Forrester found that 55% of employers now regret cutting people for AI, and roughly a third have already rehired into the roles they eliminated (CNBC, on Forrester's 2026 Future of Work report).

That contradiction is honestly the most accurate description of where we are. I don't have a clean answer for you. But I work inside this every day, so I can at least describe the weather.

How we got here

Early AI was a question-and-answer machine. It could talk, plan, suggest. But it couldn't actually do anything. The real shift came when we put tools in its hands. Give it your email and it composes and sends. Give it design tools and it produces creatives. Things that used to sit scattered and separate started coming together. That is when AI went from interesting to useful.

Here is a small story about the pace of it. A couple of years back we built an AI harness for generating marketing creatives, promo images and ad multimedia. One problem killed it. The models couldn't render text. Seven times out of ten the words on the image came out mangled (imagine a "GRAND SALE" banner that says "GRNAD SLAE"). We had three options. Build our own text layer on top, like an editable slide over an AI background. Build elaborate prompt safeguards to reduce the errors. Or just wait for the models to get better. We budgeted six months of waiting. It took three. The next generation of image models fixed it across the board.

So that is the thing that actually worries me. Not that AI is good.

AI fixes its weaknesses faster than people fix theirs. Every limitation on the list below is a snapshot. And the snapshot keeps expiring.

Two things about the cost, before the comparison

First, please ignore the "build an AI agent in 6 hours" videos. The base harness, sure, half a day. But making it do your company's actual work, the niche tasks, the extra data, the tuning, takes weeks of real effort. Anyone telling you otherwise is selling a video, not a system.

Second, the way you use AI depends on what you pay for. Cheaper models need strict workflows. Do X, then Y, and if X fails do Z. Rules everywhere. The top models are the opposite. Give them enough context, stop micromanaging, and let them work. That distinction matters more than any feature list. And it means that for a lot of roles today, especially here in India, an experienced person is still cheaper than a properly built AI system. That gap is closing. But it's real, and I am not going to pretend it isn't.

Seven rounds: humans vs AI

1. Memory. Both forget. Your employee forgets specifics unless they are unusually organised. AI forgets instructions you gave it an hour ago. And with AI you hit a strange trade-off. Give it too much context and costs rise while it struggles to prioritise. Give it too little and the output is vague. Humans mostly get priorities right by instinct. But memory is being solved on the AI side, and fast. Round goes to humans, narrowly. And temporarily.

2. Emotions. When a manager works with an employee, there are two emotional beings in the room. With AI there is one. That cuts both ways. AI never takes feedback personally, never sulks, never brings a bad morning to work. But a good employee reads the room. They know which part of an outburst to set aside and which part is the actual instruction. AI doesn't. And no, getting angry at it does not help (I have tried). It absorbs the swearing and returns the same answer, politely. Round depends entirely on the quality of the human.

3. Tools. No contest here. AI composes the email, reads the spreadsheet off one high-level sentence, works across a dozen tools at once, at any hour. And the next step is already visible. Agents with faces and voices that sit in your meetings. Round goes to AI, and it's accelerating.

4. Cost and the 80/20 problem. AI gets you roughly 80% of the way. Someone experienced has to own the last 20%. So the shape that is emerging is this. Four execution roles collapse into one AI system plus one experienced person who checks everything. Cheaper, yes. But notice what just happened. That one person is now the single point of failure for everything that ships. We removed headcount and concentrated the risk. IBM's own HR agent shows the split cleanly. It handled about 94% of routine requests and needed a human for the roughly 6% that involved real judgement (Forbes). That 6% is the whole ballgame. Round is a split decision.

5. Learning. Humans adapt on their own. Nobody has to configure them. AI learns a bit like the Matrix. Hand it a book and it absorbs the thing instantly, but someone has to set up the download first. Round goes to humans for adaptability, AI for raw speed.

6. Consistency. Set it up right and AI does the repetitive work every single time, without boredom, at 2 AM. Very few humans manage that reliably. But AI hallucinates. It confidently makes things up, which is its own category of failure a bored human never quite produces. Round goes to AI, with an asterisk.

7. Creativity. In two years, AI has never once handed me a truly original idea. It analyses, it remixes, it shows me the best version of what already exists. Because that is exactly what it is trained on. The chaos inside a human head, the connection nobody saw coming, that is still ours. Round goes to humans, clearly. For now.

Here is the whole scorecard in one place.

So what do you actually do with this?

If you are pure execution, following instructions with no judgement and no reading the room, I can't sugarcoat it. That work is going to AI, and faster than the comfortable timelines suggest.

If you are an employee, move toward the 20%. The judgement, the context, the catch-before-it-ships. That is the part companies are rehiring for right now.

And if you are a founder, the question was never "AI or people". It's which 80% to hand to the machines, and who owns the 20% that is left. That 20% is a whole subject on its own. I wrote about why it is where the value now lives, here: The New Value Equation: Why the Final 20% Is Everything. Getting that split right is the actual work. It takes weeks of building, not a six-hour tutorial.

I don't know where all of this is heading. To be honest, nobody really does. But I would rather work inside the change and report back than watch it from the sidelines.