Your Agentic AI Isn’t Failing. Your Math Is

MIT’s NANDA initiative found in 2025 that 95 percent of enterprise generative AI pilots delivered no measurable P&L return, against an estimated 30 to 40 billion dollars invested. Gartner’s forecast for what happens next is worse: more than 40 percent of agentic AI projects will be canceled by the end of 2027, driven in part by escalating costs and unclear business value. The pilots that don’t quietly die become the projects that get killed later.

It would be easy to mistake this data as proof the technology isn’t ready.

It’s not the technology. It’s your business case.

I’ve managed through enough disruptive transformations to recognize this pattern immediately. I watched leaders obsess over MIPS when computing capacity was the scarce resource. I watched them obsess over uptime and query volume during the cloud transition. Now it’s tokens processed and API calls. The instrument changes. The mistake doesn’t: leaders measure the technology. They rarely measure the capability it creates.

Agentic AI didn’t create that dynamic. It just made it faster and more expensive.

This isn’t a tool you adopt. It’s how you transform your business

I’ve led through two of these technology disruptions as an operator, print to digital and digital to platform, and I’m managing through the third right now: platform to machine, agentic systems that don’t just deliver information, they act on it. The lesson repeats every cycle: the shift looks survivable to the leaders defending the old form, and inevitable to the ones who see the function underneath and build for it early. The leaders who move first aren’t smarter. They just stop asking whether the shift is real and start asking what to build on it.

This one is not a fad, and it is not optional. It will touch every function in business sales, finance, operations, product.

To be clear, because it’s where most companies lose the thread: agentic AI itself was never the point, and the technology wasn’t in the three shifts before it either. The point is what you build with it, and how differently you run the business because it exists.

Confuse the tool with the outcome and you’ll spend two years and a real budget proving nothing at all.

Efficiency is where good intentions go to die

Most agentic AI initiatives get funded on a word that sounds like a standard and isn’t: efficiency. Efficiency is whatever the person reporting it needs it to be that quarter. Hours saved. Tickets closed. Adoption rate. None of it is wrong to track, and none of it survives a P&L review, which is exactly the gap Gartner is describing when it names unclear business value as the top reason these projects get killed.

The only measurement that matters is whether a technology improves your business. Period.

Here’s what that looks like in practice:

Agentic AI success is defined by direct commercial impact: a measurable, attributable improvement in revenue, margin, or cash conversion.

Not adjacent to those numbers. Not a leading indicator of those numbers, someday. Those numbers, named in advance, with a baseline, isolated from everything else moving in the business at the same time.

I learned this discipline decades before a large language model existed. At Pipeline360, EBITDA margin didn’t move because we picked a new technology. It moved because we named the number before we started, built the plan around hitting it, and checked it on a calendar we’d committed to in advance. Agentic AI doesn’t change that discipline. It just raises the cost, and risks, of skipping it.

Think big. Start small. Measure what you can actually prove.

Two instructions, in this order, and most companies only manage one of them.

Think big. The ambition has to match what’s actually happening. Agentic AI will change how your business runs and grows, not which vendor logo sits in your tech stack. If your plan tops out at a chatbot on the website, you’re thinking too small to justify the investment.

Start small. One workflow. One named owner. One proof point, defined before you spend a dollar. The companies pulling ahead right now aren’t running the most pilots. They’re the ones willing to kill nine to get one into production with a number attached to it.

The gap between those two instructions, big ambition and narrow execution, is where most companies fall apart. They either think too small, a tool instead of a transformation, or start too big, a company-wide rollout with no single owner and no defined proof point. What’s missing is a way to know, project by project, exactly where you stand and what has to be true before the next dollar goes in.

The ladder: Assisted, Delegated, Accountable, Compounding

I built a framework for exactly this, a four-stage ladder called ADAC. It doesn’t score your company. It scores the initiative in front of you, and it works the same way whether you’re the one funding it, running it, or diligencing it before you invest.

Assisted: a human still owns the task and the result

This is where almost every company starts, and where most never leave. Someone uses an AI copilot for research, drafting, summarizing. Genuinely useful. Not yet a workflow, and not yet anyone’s job to make it one. The sign that you’re stuck here: your only evidence is a usage log. “The team uses it daily” is a habit report, not a business case. Name an owner this quarter, and give them one workflow to rebuild, not ten to experiment with.

Delegated: AI owns the task, a human still owns the result

This is where most of Gartner’s cancellations happen. AI executes a defined piece of work inside a workflow, drafting the follow-up, qualifying the lead, extracting the contract terms, while a human reviews everything before it moves. Real progress, and also where companies get addicted to running pilots instead of finishing one. The tell: a portfolio of pilots, none with an exit date. Freeze new ones. Take the pilot closest to done, name its financial proof point, and get it into production or kill it.

Accountable: AI owns the task and the result, a human owns the exceptions

This is where my sentence stops being aspirational and starts being enforced. The workflow runs end to end. A human steps in only when something breaks, and you should expect an attributable line item on the P&L by now. If you can’t name it, you’re not actually here, you’re still at Delegated wearing this stage’s name tag. The sign you’re stuck: the workflow performs in every test, but leadership still won’t remove the human checkpoint. That’s a governance problem, not a technology one, because nobody has agreed who’s accountable when the agent is wrong on a live customer or on something with a real dollar amount. Name that owner, with defined exposure and a plan to reverse it, before you remove the checkpoint. Not after.

Compounding: AI owns the result and keeps improving how it gets there

The system doesn’t just run the workflow, it improves it, without anyone filing a change request. This is where the mechanism I described in “Loop Engineering” stops being a coding practice and starts being an operating one: judgment encoded well enough that the system holds your standard without you restating it every time. It’s rare, and it’s the only stage where the return compounds instead of holding flat. The sign that you’re stuck: improvement stops the day your one champion gets pulled onto something else, which means you never built the capability, you borrowed one person’s judgment for a while. The fix is institutional: the improvement loop needs a standing role, not a hero.

Score the project, not the company

One warning before you use this stages framework. Don’t score your company. Score the initiative. A business can be Accountable in sales operations and stuck at Assisted in finance, and that’s not a contradiction, that’s normal.

A founder uses this to decide what to fund next. An operator uses it to know where the org actually stands, project by project. An investor uses it in diligence to tell a real deployment from a demo.

* What’s the metric? Revenue, margin, or cash conversion.

* What’s the baseline? Where this number stood before the initiative touched it.

* What’s the required delta? The number this has to move to justify the spend.

* What’s the date? When you’ll check it.

If you can’t answer all four before you write the check, you’re not funding a workflow. You’re funding a hope.

The Math Isn’t Complicated. The Discipline Is

None of this requires you to become a data scientist, and none of it requires you to wait until the technology settles down before you move. It will not settle down. What it requires is the discipline that has separated companies that make money from companies that make noise since long before agentic AI existed: name the number before you start, know exactly which stage you’re standing in, and refuse to fund the next one until the last one paid for itself.

Confused, frustrated, behind: pick whichever fits this week. None of it disqualifies you from starting. A 95 percent failure rate isn’t evidence you’re late. It’s evidence almost everyone is still measuring the technology instead of the capability it creates. Fix your math first, and you move ahead, not catch up.

Think big. Start small. And never confuse activity with proof.

The views expressed in Uphoff on Media are entirely my own. They don’t represent the opinions of any company I’ve led, any board I’ve sat on, or any investor who’s had the pleasure of debating strategy with me over the years. If something I write here sounds brilliant, I’ll take full credit. If it turns out to be wrong, I was clearly misquoted by myself.

”Uphoff on Media” is published by Tony Uphoff, Founder and Managing Partner of Uphoff Advisory, a strategic advisory practice for founders, CEOs, and investors in B2B information, marketing, and technology. The businesses that drive business.