Why AI Gains Disappear Before They Reach the P&L

Someone in operations gets access to an AI tool on a Tuesday. By Friday they have built an agent that reconciles vendor invoices in twenty minutes instead of two days. They demo it. People clap. Somebody says "imagine this across the whole company."

Six months later the invoices still take two days.

The agent works fine. It just hands its output to an approval queue that meets on Thursdays, pulls from a spreadsheet finance doesn't trust, and runs on a personal account that security found out about last week.

I hear some version of that story almost every time I talk with leaders about AI. Different industry, same plot.

Here's the problem. The AI isn't failing. It's being bolted onto a company that was built for something else.

The numbers agree

This isn't a vibe. Earlier this year, nearly 6,000 executives in the US, UK, Germany and Australia were surveyed through central bank business panels. Nine in ten said AI had no impact on productivity or employment at their firm over the last three years.

McKinsey's latest State of AI survey comes at it from another angle. About 80% of respondents say AI made them personally more productive. Only 37% see any impact on EBIT, and that number didn't move from last year.

Everyone feels faster. The P&L doesn't.

You've probably seen the headline that 95% of AI pilots fail. That figure comes from interviews and conference surveys, and I wouldn't bet a budget on it. The stronger data tells the same story anyway.

And the technology works. When a Fortune 500 software company rolled an AI assistant out to 5,172 customer support agents, issues resolved per hour went up 15%.

The tool works for the person. It just never shows up for the company.

We've seen this before

When factories first got electric motors in the 1890s, most of them did the obvious thing. They put a big motor where the steam engine used to be and kept everything else. Same overhead line shafts. Same leather belts. Same multi-story buildings designed around one power source in the basement.

It was a better engine running the same old factory. As late as 1899, electric motors were under 5% of factory drive power. They didn't pass half until the early 1920s, four decades after the first central power stations.

The payoff came when factories were rebuilt around the motor instead of the shaft. One motor per machine. No overhead shafts. Single-story plants laid out around how materials actually moved. That redesign is when manufacturing productivity finally took off.

Paul David made this comparison in 1990, and one of his points has aged well: a company's information systems are its factory floor, except they never wear out. Nothing forces you to rebuild them.

Most companies today have bolted a motor onto the line shaft.

Four places the gains get lost

1. Legacy process

AI speeds up one step. The work still flows through the same handoffs, approvals and queues. The time saved doesn't disappear. It just piles up as waiting at the next desk.

McKinsey's data shows who gets past this. Among the small group seeing real EBIT from AI, nearly three-quarters have fundamentally redesigned workflows. Among everyone else, about a quarter have.

The winners didn't buy better AI. They changed the work.

2. Legacy incentives

A Danish study linked payroll records to surveys of 25,000 workers at 7,000 workplaces. AI chatbots saved them about 3% of their work hours. Earnings and hours worked didn't measurably change.

The saved time didn't vanish. It got poured into other tasks. More email. More meetings. More polish on things that didn't need it.

Nobody decided where the forty minutes should go, so they went where they always go.

3. Legacy data

AI is only as good as what it can reach. In most companies that means data scattered across systems, defined three different ways and owned by nobody in particular.

Gartner found 63% of organizations either lack or aren't sure they have the right data management practices for AI.

A great model on bad data is a fast way to be confidently wrong.

4. Legacy operating model: heroes instead of teams

This is the one I see most.

Someone has a great idea. They vibe code a tool over a weekend, or build something another team should own. The data team, IT, security, the people who run the process. None of them were asked. It works on a laptop, on a copy of the customer list, under a personal account.

Multiply that by a thousand people. That isn't an AI strategy. It's a thousand AI strategies.

It also leaks. Nearly half of employees (48%) say they've uploaded company information like financial, sales or customer data into public AI tools. That number rises to 67% at organizations that banned generative AI outright.

Bans don't stop it. They just hide it.

IBM's latest breach research found shadow AI, meaning employees using unapproved AI, was involved in 43% of the security incidents studied, up from 20% a year earlier. Those breaches averaged $5.39 million. And 68% of breached organizations had no AI governance in place.

Governance isn't the brake here. It's how one person's good idea gets safely to a thousand people instead of staying on one laptop.

People, process, and technology

We've said "people, process and technology" in this industry for decades. With AI, most companies are running technology and process, executed by one person.

They left out the people.

The factories didn't get faster because someone bought a motor. They got faster because the people who designed, ran and worked the floor rebuilt it together.

None of this means stopping your builders. The enthusiasm is the most valuable thing you have. Point it somewhere useful.

The person with the great idea should be the one who calls the meeting, not the one who ships it alone.

What to do Monday

  1. Start with people. Before anyone builds, get the owners in a room. The team that owns the process. The data owner. Security. The people who will actually use it.
  2. Pick one workflow and redesign it together, end to end. Not one person automating their slice. The whole flow, including the approval queue that meets on Thursdays.
  3. Lay the foundation. Fix the data behind that one workflow. Give people approved tools with guardrails, because the ban is already failing.
  4. Decide where saved time goes. Then measure the business result, not the hours saved. Hours saved is how you get 80% feeling productive and 37% seeing it in EBIT.
Stop buying better motors. Rebuild the floor.

Sources

The views here are my own and don't represent my employer.

This blog is written by AI & myself. Concepts completely human.

Comments

Post a Comment

Popular posts from this blog

I Took My Own Advice in an Interview. Pure Storage Didn't Flinch.

Image
I'm joining Pure Storage on March 2nd to be part of the technical marketing team, and I'm genuinely excited about it. During the interview process, I took my own advice from a blog post I wrote a few weeks ago and asked them the question that makes most candidates nervous: "What concerns do you have about me for this role?" They didn't flinch. They gave me real feedback. We had an actual conversation about fit, not just a polished interview performance. That's when I knew this was the right place. What Got Me Excited The conversations with Pure focused on enablement. Not campaigns. Not metrics. Enablement. How do we help the people in front of customers have better conversations? How do we give them what they actually need to succeed? Those questions resonated with me because I've done that work before. I know what it's like when you have the right materials at the right time. When customer stories are authentic and metrics are real. When enablem...
Read more

The Infrastructure Reckoning, Part 2: The Virtualization Reckoning

Image
The invoice landed and nobody wanted to open it. Not because the number was surprising. Because it confirmed what everyone in the room had suspected for months: the platform that enterprise IT had built its entire operational model around had just repriced itself out of the relationship. No negotiation. No grandfathering. New terms, new math, and a renewal window that didn't leave much room for alternatives. That moment — repeated across thousands of enterprise environments between 2023 and 2024 — was not a licensing dispute. It was a stress test. And a lot of stacks failed it. 59% of VMware customers saw prices rise by more than 25% after the Broadcom acquisition, with the median increase landing between 25% and 49%. Extreme cases hit 350% to 700%. Gartner's Peer Community survey found 74% of IT leaders actively exploring VMware alternatives. And yet most of them haven't moved. 86% of organizations are actively reducing their VMware footprints, but just over half plan phas...
Read more

If I do the homework, you owe me a phone call. The death of decency in hiring.

Image
Honestly, I almost didn't post this. You know how it is. You don't want to look salty. You don't want to look difficult. It’s easier to just roll your eyes and move on to the next application. But I can’t be the only one dealing with this nonsense, so I’m saying it. Here’s the situation: I’m interviewing for a senior role. Things are going great. The hiring manager is cool, we’re clicking, the vision matches. Then comes the homework. And not a 30-minute writing sample. They asked for the full package: A 12-month strategic plan for Analyst Relations. How to position them for MQs, Waves, the works. I sat down and did the work. I audited their competitors. I built the roadmap. I gave them a strategy I would usually charge a consulting fee for. I sent it off, ready to talk shop. The result? Ghosted by the human, rejected by the robot. No feedback. No "hey, thanks for the effort." Just a generic, no-reply template email: “We have decided not to move forward.” Since w...
Read more