The hard yards of enterprise AI: What it really takes to make it work.

The noise around AI shows no sign of abating, and I understand why. The technology is genuinely mind-blowing. But if I’m honest, I also see a lot of organizations struggling to turn the promise of consumer AI into real, repeatable value in the workplace.

An analyst said something to me at Dreamforce 2025 that stuck with me. “Everyone’s sung the national anthem,” they said, “but no one’s pitched the first ball.” That neatly sums up the state of enterprise AI today. We’re inspired, we’re excited, but many are still experimenting and piloting and haven’t yet walked the hard yards to make AI useful in the narrow, everyday jobs that actually run a business.

At Bionic, we help more than 200,000 small businesses across the U.K. compare, switch, and manage essentials like energy, insurance, loans, and broadband. Agentforce already handles a meaningful share of routine requests, enabling customers to get help without the hassle and freeing our teams to focus on the moments that matter.

From what I’ve learned firsthand through deploying Agentforce in live operations, making AI work in the enterprise isn’t about chasing the most impressive demo. It requires you to escape the experimentation trap, stay disciplined in the face of sometimes underwhelming minimum viable products, and maintain an unrelenting focus on the outcomes you set out to achieve. Here’s what I’ve learned so far.

There’s a lot of debate about where companies should start with AI — sales, marketing, finance, or back office. In practice, most organizations start in service, and for good reason.

Service work tends to consist of routine and repeatable tasks. For us, service is an opportunity to build trust. We want small business owners to focus on running and growing their own businesses, not spend their time waiting on hold to get basic answers about essential services.

So, we looked closely at our customer contact drivers and asked a straightforward question: Which of these genuinely need hands-on, urgent, and human support, and which don’t? That analysis led us to a clear north star: Around 50% of customer needs could be met agentically and, of those, 60% shouldn’t need a human in the loop. Conversely, this created the opportunity to focus reps on those cases that require empathy, contextual awareness, and hands-on support.

Our goal wasn’t to replace people, but to elevate human support and free up time for the conversations where our staff can shine.

One of the most sobering surveys from 2025 was the MIT report, famously headlining that 95% of enterprise AI fails. If you delve beneath the headline, you uncover an important root cause: Many companies implement AI, but don’t integrate it with the way work already gets done.

If AI sits outside your core workflows and CRM, you risk getting stuck in the experimentation trap, with AI forever remaining a pilot that can’t be scaled and won’t deliver meaningful value.

We integrated AI with Salesforce from the outset so it wasn’t a pilot or bolt-on, but rather an integral part of our CRM solution that helped us scale quickly.

AI success isn’t a vibe — it’s a set of outcomes.

In our sales and service operations, we look at performance through the lenses of efficiency, productivity, availability, and conversion. How many calls does a sales rep make per day? How much time are reps actually spending talking to customers? How often are we connecting the right customer to the right rep at the right time?

When we started using generative AI and agents from Salesforce for intelligent routing and progressive dialing prioritized by lead scoring, we saw improvements across the board. Sure, AI was applying some intelligence, but we saw that automation was relentlessly removing waste from the system, meaning fewer missed calls, less manual work, and more focus on helping small business owners sort their business essentials.

If you can’t tie AI initiatives directly to operational metrics, you’re probably still stuck in experimentation mode.

With heightened expectations, the pressure to attain a quick “speed to value” can be enormous, but the two speeds are often not appreciated.

You can deliver AI quickly - often in days or weeks, but value doesn’t arrive on a fixed schedule. Spoiler alert: AI is not going to transform your business overnight!

Many organizations assume AI agents will instantly improve as they’re exposed to internal data, but LLMs don’t automatically learn from your systems or customer interactions. They’re pre-trained, not self-updating.

To deliver reliable, high-quality responses, AI agents need an intentional cycle of continuous review and refinement. That means business teams regularly reviewing transcripts, updating instructions, tuning responses, and enriching the knowledge base over time.

This isn’t a sign that the technology is falling short. It is the work.

One of the most exciting things we’re building with Agentforce is what we call the Wingman: an AI agent that sits alongside reps, offering real-time guidance, sentiment analysis, and quality assurance.

This isn’t about automation for its own sake. It’s about empowerment, giving all our reps a new superpower that enables them to have more powerful and higher converting conversations.

In the case of call quality, the impact is game-changing. Our Wingman shifts compliance from an opaque back-office activity and puts reps fully in control of their own call quality.

Enterprise AI requires a leap of faith.

During the last year, the AI conversation has shifted. The agentic AI metaphor was that it’s just like hiring 100 new interns, the only limiting factor being your own imagination. Now that we’ve experienced digital labor in the workplace, it turns out that some of these interns aren’t well placed to tackle narrow jobs, aren’t totally reliable, and can be frustratingly inconsistent.

Yet we’ve also seen the promise and examples where AI can be just as mind-blowing as consumer AI.

There’s so much focus on creating bigger and better LLMs, but I wonder whether we’re waiting for small LMs to work alongside LLMs and bring that company-specific intelligence?

Whether the answer becomes bigger or smaller models, the organizations that will win aren’t the ones chasing the flashiest use cases. They’re the ones doing the hard yards: integrating AI deeply, refining it continuously, and staying focused on real business outcomes.