In the 80s, the economist Robert Solow quipped: "You can see the computer age everywhere but in the productivity statistics." It became known as the Solow Paradox, and forty years later, we have a new version of the same puzzle. ChatGPT reached 100 million users faster than any product in history. And yet the majority of enterprise AI initiatives are still stuck in pilot. The technology is everywhere. The impact… isn't.

That gap is the defining tension in product right now. On one side, there’s enormous pressure - from boards, investors, and the relentless drumbeat of model announcements - to move fast on agentic AI. On the other side, there’s the product mindset that most of us were trained on: build things that solve real customer problems, ship incrementally, and don’t mistake excitement for value. In B2B, that tension is sharper still. The workflows we’re trying to automate are the ones that move billions of dollars, carry legal liability, and land on the front page if they go wrong.

This isn’t an argument to slow down. It’s an argument to get it right. Because the product teams that are actually delivering on AI aren’t the ones chasing the latest model release - they’re the ones who refuse to let panic override good product thinking.

The reality on the ground

Most AI conversations are happening from the wrong perspective. Roadmaps are being shaped by model announcements and market capabilities - what the latest GPT release can do, which agent framework is winning, how quickly a competitor has shipped a Copilot feature. Pilots are spun up to figure out how the latest technology can be applied to problems, rather than focussing obsessively on the problems that actually need solving.

It’s technology looking for a use case, not the other way around.

And a lot of that is because the problems themselves are harder than they look. Most of the workflows that B2B customers actually care about - the ones driving their biggest costs and risks - are messy, fragmented, and poorly defined. Processes spread across spreadsheets, email threads, and legacy systems. Institutional knowledge locked in people’s heads. No single source of truth, no consistent data, no standardised workflow. Point an AI agent at that reality and you don’t get automation… you get garbage. Garbage in, garbage out.

“The question isn’t what the AI can do. It’s what it should do.”

Paul Mumford, Product @ Omnea

On top of that, there’s a widespread misconception that agentic AI is a switch you flip. The jump from “we have a Copilot feature” to “AI handles end-to-end workflows autonomously” is enormous, and trying to make that leap in one go is a reliable path to failed pilots, eroded customer trust, and awkward board conversations.

The product teams getting this right aren’t moving faster. They’re thinking more carefully about three things.

Three things that separate success from failure

1. Turn chaos into order first

You can’t automate chaos. It sounds obvious, but it’s the single most common reason enterprise AI initiatives fail. Teams try to layer AI on top of fragmented, inconsistent processes and wonder why it doesn’t deliver.

The organisations that succeed treat this as a sequencing problem. Before they automate, they organise. They consolidate fragmented tools into unified workflows. They standardise processes so there’s consistency in how work moves through the organisation. They clean up their data so AI has something reliable to work with.

This isn’t the exciting part - nobody writes a LinkedIn post about cleaning up a procurement intake process - but it’s the foundation everything else depends on. And for product teams, it’s actually a significant opportunity: building AI-ready workflows often means helping customers get genuinely organised in ways that deliver value on their own, before AI is even in the picture.

In procurement, where we work closely with Fortune 500 companies at Omnea, we’ve seen teams transform their operations simply by consolidating scattered intake processes into a single, structured workflow. The AI capability comes later - and when it does, it works dramatically better for it.

2. Crawl, walk, run

The product teams delivering real value aren’t trying to leap straight to fully autonomous agents. They’re taking a deliberate, phased approach - and crucially, they’re being honest about what level of AI autonomy is actually appropriate for each use case.

In practice, we think of AI autonomy mapping to four levels:

• Level 1: AI generates content or suggestions - a draft clause, a flagged risk, a spend summary - but a human always executes the action.
• Level 2: AI performs multi-step workflows but requires human approval before execution.
• Level 3: AI executes end-to-end with post-execution human auditability.
• Level 4: Tasks run autonomously with humans only stepping in for exceptions.

Use cases determine the right level of AI autonomy - the same organisation will operate at different levels across different workflows.

The key insight - and the one that’s easy to miss when you’re excited about what the technology can do - is that the right level is determined by the use case, not the technology. Different workflows within the same product will sit at different levels, and that’s by design.

Take procurement: I’d be comfortable with an AI agent autonomously handling a £10,000 software renewal. The stakes are low, we’re not negotiating it at the moment, and the cost of a mistake is recoverable. But that same agent handling a £2 billion property deal? Absolutely not - regardless of what the technology is capable of. The question isn’t what the AI can do. It’s what it should do, given the stakes, the data quality, and the customer’s readiness to trust it.

This is where genuine product thinking is invaluable. Good product practice means asking not “what can our AI do?” but “what should it do in this context?” That’s a harder question, and it’s the one that determines whether a product creates real value or just creates a very sophisticated liability.

3. Start with the use case, not the technology

This is the one that sounds obvious but gets violated most. Roadmaps built around shipping an AI agent feature by Q2 almost always produce worse outcomes than those built around a specific problem worth solving.

Starting with the use case changes everything. It forces a clear definition of success before anything gets built. It keeps the team grounded in actual constraints - compliance requirements, approval chains, data availability. And it means the level of AI autonomy is chosen to fit the problem, not to demonstrate capability.

The practical approach: identify the workflows that cost your customers real time, real money, or real risk. Find the right level of AI involvement for each. Push as far up the autonomy levels as the technology and customer readiness genuinely support. Then revisit regularly - a use case that warranted Level 1 six months ago may be ready for Level 2 today. New models, better data, and growing customer trust all shift the calculus. Build the habit of reassessment, not the expectation of solving it all at once.

The unglamorous truth - and what to do with it

None of this is complicated. The difference between AI initiatives that deliver and the ones stuck in pilot purgatory isn’t access to better models. It’s discipline: resisting the pull of the hype cycle, ignoring the noise about what competitors are supposedly shipping, and staying focused on the fundamentals.

Before you build anything, these are the four questions worth asking:

1.  Is the underlying process clean enough to be worth automating?

If you can’t define the workflow step-by-step, and get access to high-quality data across it, it’s not ready for AI. Fix the process first. This is often where the biggest near-term value lives anyway.

2.  What’s the cost of AI getting it wrong?

This determines your autonomy level. Low stakes, well-defined process: higher autonomy is fine - shoot for Level 4. High stakes, ambiguous context: keep a human close. Be explicit with customers about where that line sits - and your roadmap to climb the levels - they’ll trust you more for it, not less. Customers burned by an overambitious Level 4 pilot won’t give you a second chance.

3.  What does the customer actually need?

Not what they’re asking for, and not what the technology enables. What specific problem costs them real money or real risk - and are you solving that, or building something that’s interesting to you?

4.  Do you have the data to make this your advantage?

AI is only as good as the data it has access to. LLMs are powerful general reasoners, but they rarely have the specialist, cross-functional context needed to perform reliably in areas like Legal, Risk, and Finance. Identify what structured, high-quality data your AI needs - and be honest about whether you have it.

Conclusion

The companies actually embedding AI into real workflows - and the B2B products helping them get there - aren’t doing anything flashy. They’re doing the basics exceptionally well. In a market where everyone is chasing the same model announcements, that discipline is becoming the real competitive advantage.