A Fortune article hit the front page of Hacker News this week with a provocative headline: 'Thousands of CEOs Just Admitted AI Had No Impact on Employment or Productivity.'

My first reaction was to roll my eyes. My second reaction was to actually think about it.

Because the reality is - I've been running AI in production for over a year. I fine-tuned a 20B model on Apple Silicon. I have local LLM inference running on my Mac Studio M3 Ultra at 80 tokens per second. I've integrated AI into Clearline workflows, built AI-assisted tooling at a fintech firm, and watched colleagues either swear by it or swear at it daily.

And the CEOs? They're not wrong. But they're measuring the wrong thing.

🧠 What the Productivity Paradox Actually Says

The article leans on something called the Solow Productivity Paradox. Robert Solow famously said in 1987: 'You can see the computer age everywhere except in the productivity statistics.' Same story.

The point isn't that computers were useless. It's that transformational technology takes time to show up in macro-level economic data. You have to retool processes, retrain workers, and rethink workflows - not just hand everyone a new toy. The payoff for the PC revolution showed up in the mid-90s, about 15 years in.

The AI productivity paradox is the same movie. Except this time everyone can see the frame skipping in real time and is tweeting about it.

⚙️ Where I've Seen AI Actually Make a Measurable Difference

To be concrete:

At CTM, I use AI to compress what used to be 3-hour close looks into 20-minute synthesis sessions. When a client needs a technology assessment - say, 'Should we move our LIMS to the cloud?' - I used to need days. Now I generate a structured brief in an hour, validate with a call, and have a recommendation in the client's hands same day.

That time I saved didn't show up on a productivity dashboard anywhere. My revenue didn't spike. But I served two more clients that week instead of one.

At a fintech firm, the integration engineering work is repetitive in specific ways - parsing API specs, drafting field mapping docs, writing boilerplate transform logic. An LLM handles those in minutes. I still review everything. But the dumb parts of my job got a lot shorter.

None of this shows up in a CEO survey.

🏗️ The Infrastructure Problem Nobody's Talking About

My actual take on why the CEO data looks flat: most enterprise AI deployments are still in the 'hand everyone a ChatGPT subscription' phase. That's like giving every office worker a computer in 1983 and expecting it to transform productivity - when they're using it to play Solitaire and type documents that still get printed and filed.

The gains come when AI gets embedded in workflows, not just available beside them.

I learned this firsthand when I integrated AI into a lab workflow at a SENAITE LIMS client. The before state: technicians manually filled sample forms, ran a calculation spreadsheet, then copied results into a report. Three separate tools, two copy-paste steps, maybe 20 minutes per sample batch.

The after state: one Python script, triggered from LIMS, calls a local LLM to parse free-text sample notes into structured data, validates against known analyte ranges, and generates a draft report. Time per batch: under 4 minutes.

That's not 'AI at the side of the workflow.' That's AI in the workflow.

# Simplified version of the integration
import anthropic
import json

def parse_lab_notes(raw_notes: str, analyte: str) -> dict:
    client = anthropic.Anthropic()
    
    message = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=512,
        messages=[{
            "role": "user",
            "content": f"""Parse these lab technician notes for analyte '{analyte}'.
Return JSON with: {{result_value, units, flags, confidence}}.
Notes: {raw_notes}"""
        }]
    )
    return json.loads(message.content[0].text)

# Wraps into SENAITE's REST API for automated ingestion
def submit_result(sample_id, parsed_result):
    # ... SENAITE API call here
    pass

That 80% time reduction? A CEO survey won't capture it. The lab doesn't have a 'workflow efficiency' metric they track. The technician just goes home earlier on Tuesdays now.

🔬 The Measurement Problem Is The Real Problem

This comes up often, especially wearing the fractional CTO hat. When clients ask me whether AI is 'worth it,' I have to push back on the question itself.

Traditional productivity measurement is designed for output-per-hour work. Factory widgets. Support tickets closed. Code commits merged. It struggles with the kind of value AI generates, which tends to be:

  • Cognitive overhead reduction (I thought about this problem 20% less than I would have)
  • Decision speed improvement (I made this call in 2 hours instead of 3 days)
  • Error rate reduction (the LLM caught the field mapping mismatch before I pushed it)
  • Recaptured async time (I didn't have to schedule a 90-minute meeting; I summarized the docs instead)

None of that shows up in the metrics CEOs are watching. So of course the data looks flat.

📡 What Actually Needs to Change

On the practical side: If you're running a team or a company and you want AI to actually move productivity numbers, you need three things:

First: stop treating AI as a standalone tool. It needs to live inside the workflow. API-first. Embedded. Not just 'available in a browser tab next to your actual work.'

Second: measure the right things. Add telemetry to your workflows. Time-to-first-draft. Decision cycle time. Rework rate. If you don't measure it before AI, you can't measure improvement after.

Third: be honest about the transition cost. AI integration isn't free in engineer-hours. It takes time to build, validate, and trust. The Solow Paradox is real - expect 12-24 months before macro productivity shifts are visible even in well-run deployments.

I'm running local inference on a Mac Studio M3 Ultra for exactly this reason. I don't want to depend on a cloud provider's rate limits or pricing model for infrastructure that's embedded in client workflows. Low latency, zero per-query cost, full control. That's what makes embedding viable at scale.

# Ollama serving locally, called from workflow scripts
curl http://localhost:11434/api/generate \
  -d '{
    "model": "qwen3-coder-20b",
    "prompt": "Parse this field mapping spec and output JSON: ...",
    "stream": false
  }' | jq .response

🧠 Lessons Learned

  • The AI productivity paradox is real and historically consistent. Don't panic, don't dismiss - contextualize.
  • Enterprise AI is still in the 'computer next to the typewriter' phase for most companies. The gains are coming.
  • Embedding AI in workflows beats making it available beside them. Every time.
  • Traditional productivity metrics are blind to cognitive overhead savings. You need new measurement frameworks.
  • Local inference on capable hardware (hello, Mac Studio) is the enabling infrastructure for embedded AI that actually sticks.
  • Time to value is 12-24 months. Anyone promising immediate economy-wide productivity transformation is selling something.
  • The CEOs aren't wrong - but they're looking at the wrong data. The question isn't 'did AI improve productivity this quarter?' It's 'did AI improve the right things, measured the right way, over the right timeframe?'

I don't know if AI is going to be the PC revolution or the flying car. But I know that the people making the most of it right now aren't the ones handing out subscriptions. They're the ones rebuilding processes around it.

That's slower work. But it's the only work that actually compounds.