The Big Problem with AI Agents

It's been roughly half a year since the 'agent era' was officially announced—a title so confidently futuristic it might as well come with its own hype soundtrack. But while everyone's busy chasing cost savings and labour replacements, something bigger is quietly happening in the wings.

Of course, we're still in the awkward infancy stage. Like a toddler confidently sprinting into a coffee table, AI agents are stumbling; though it's not really their fault. The real problem is how narrowly we've been thinking about them.

Even the terminology is getting fuzzy. Things we casually called "automation" a year ago have suddenly been upgraded to "agents." It's like rebranding leftovers as "chef-inspired cuisine"; catchy, but questionable. Yet behind these buzzwords lies an important difference:

• Assisted AI helps you do your job.

• Agent AI does your job while you have a coffee.

One enhances your actions; the other fundamentally reshapes what you do—or rather, what you don’t have to do.

Beyond Cost-Cutting: Unlocking Real Innovation

We might be missing the point if we only talk about AI agents in terms of cutting costs. That’s like describing smartphones purely as "phones without wires"—accurate but entirely missing the point.

Salesforce’s Agentforce is a good example. CEO Marc Benioff insists 2025 will be the year agents truly arrive. But most conversation remains stubbornly about replacing workers rather than creating genuinely new possibilities. It's like using a rocket just to commute to Tesco—technically impressive, but hardly visionary.

Facing Up to Reality

Despite the enthusiastic promotion, Salesforce's CFO Amy Weaver recently delivered a decidedly less excitable forecast, predicting only "modest" sales growth for Agentforce. Of the 5,000 deals proudly announced, around 2,000 are still languishing in the "free-trial abyss," while many paying customers might not even realise they have it—bundled like a free pen you didn't know came with your bank account.

This isn't a unique Salesforce problem; it's the classic tech industry challenge, a brilliant solution eagerly searching for a problem. The tech’s impressive, but the market is still scratching its head, figuring out where exactly these agents belong.

Pricing Problems: A Multi-Faceted Mess

Adoption isn't just a technical question—it's an economic puzzle too, and one we haven't solved yet:

1. The Usage Paradox

Salesforce’s pricing—about $2 per conversation—ironically discourages heavy use. It's akin to an "all-you-can-eat" buffet that charges extra every time you reach for dessert.

2. AI vs Global Labour

If an agent deals with ten customer interactions per hour, that’s $20 an hour—often pricier than outsourcing. Saving just a few pennies per interaction hardly justifies the futuristic fanfare.

Interestingly, Manus AI, a Chinese competitor, pays about the same per interaction but isn’t charging yet, opting instead for viral popularity. It’s like throwing a lavish party with an open bar;

you’ll be popular, but someone's going to have to pick up the tab eventually. Manus might have a secret weapon though—specialisation.

The Power of Specialisation: Do It Like a Visionary

Vertical, industry-specific agents are more promising than those broad, do-it-all AI solutions. It’s reminiscent of Steve Jobs' famous belief: focus on simplicity over complexity and specialisation over generalisation.

Here’s why vertical agents work better:

• Defined Risk Zones: Like bowling with bumpers, it’s harder to fail disastrously.

• Easy Adoption: They fit neatly into existing company workflows, like finding the perfect USB port on the first try.

• Simple Success Metrics: Clear measurements speed up feedback loops; it's easy to tell when you’re winning.

Yet too many companies ignore these benefits, aiming for the flashy horizontal solution; like designing a Swiss Army knife with 87 functions when all you really need is a decent corkscrew.

UX Matters: Waiting for an "iPhone Moment"

Manus AI succeeds not because it's cheap, but because it's enjoyable. Steve Jobs famously said people don’t always know what they want until you show it to them; a principle clearly lost on many current solutions, which often feel more like "mandatory digital colleagues" rather than tools you choose to use.

We haven’t yet had our agent "iPhone moment"—right now, we’re more in the Nokia 3310 phase. Sure, it's robust and nostalgic, but it's mostly memorable for playing Snake.

The Market Isn’t Quite Ready Yet

Most businesses are still just dipping their toes in with simpler assistants like ChatGPT. Jumping straight to fully autonomous AI agents feels intimidating; partly because the risk of AI hallucinations is real. It's manageable if your chatbot misquotes Shakespeare, less so if your autonomous agent accidentally refunds thousands of customers.

Salesforce’s aggressive bundling tactics; like threatening to hike prices on other services if Agentforce isn't adopted, betray exactly this kind of hesitation. After all, if the product truly sold itself, there wouldn’t be any arm-twisting required. These are the moves of an organisation trying to force adoption of something the market isn't naturally embracing.

Another problem is data access. An agent restricted to a single platform is like owning a Ferrari but only allowed to drive it around your driveway. Agents need open roads; in other words, flexible data integration—to truly flourish.

Turning Things Around

Three factors could change the landscape:

• AI Model Price Wars: Falling costs among top AI providers could finally tip the economics in favour of wider adoption.

• High-Value, Specific Use Cases: Solving clearly defined tasks, like lead qualification, will prove tangible ROI.

• Cross-Platform Data Integration: Agents that easily pull data from various sources will dominate the market.

The challenge is that agent companies have a strong incentive to always use the most state-of-the-art models; precisely the ones least affected by price drops. The breakthrough will come when "good enough" models become dramatically cheaper while still performing adequately for agent tasks.

Salesforce's current push might simply be strategic positioning: losing a little ground today to dominate tomorrow.

This land grab, early experience, and deeper understanding of where agents actually deliver value could prove strategic even if they're 6-18 months ahead of market readiness.

One Model Isn't Enough: The Multi-Model Future

Different jobs demand different skills. Similarly, successful AI deployments will involve multiple specialised models, not a single "mega-model" trying to do everything. It’s the difference between a toolbox and a Swiss Army knife; both useful, but one is clearly more practical for serious work.

This architectural approach creates significant advantages in both performance and economics, but requires fundamentally rethinking how we design agent systems. It's the AI equivalent of Jobs' insistence on controlling the full technology stack rather than relying on general-purpose components.

The companies that master this architectural approach won't just have cheaper agents; they'll have fundamentally more capable ones that create entirely new possibilities.

Organic Beats Top-Down

Truly transformative innovation rarely comes from corporate directives. It's teams tackling real problems who’ll naturally integrate and adopt agents; an "accidentally agentic" approach.

Platforms like Cursor or Replit have thrived precisely because they focused on solving real user problems, not grandiose tech promises.

Cursor and Replit are AI-driven developer agents. Their founders didn't start by trying to replace developers; they focused on making specific coding tasks more efficient. As users experienced the value, the scope naturally expanded. This bottom-up adoption path feels much more aligned with how revolutionary technologies actually take hold.

The Top-Down approaches we see in AI Agent frameworks like CrewAI, LangChain and SmythOS made the mistake of grandiose anticipation. Some even tried to establish themselves as the authority on the framework. So whilst powerful tools and platforms, they do not meet the needs of today but tomorrow, which will change, and the user base is frustrated.

The visionaries from Cursor and Replit, they understood and anticipated user needs, but incrementally built the products to meet the user where they were; all whilst creating more demand with next step innovations. Users adopting the tooling then created natural demand for more; ergo anticipation of their needs became more accurately predicted.

Thinking Bigger: Agents as Organisational Change

Like smartphones and the internet, AI agents aren't merely tools; they’re a whole new way of operating. Companies will need roles like "agent supervisors," incremental trust-building methods, and clear guidance on human-machine collaboration.

Practical Steps: Six Key Principles

1. Pick One Vertical: Master a single area first.

2. Prioritise Great UX: Make agents intuitive and indispensable.

3. Use Multi-Model Solutions: Match AI models precisely to tasks.

4. Allow Organic Adoption: Let practical successes drive agent usage.

5. Prepare for Change: Anticipate and plan for shifts in roles and responsibilities.

6. Measure Real Impact: Focus metrics on opportunities created, not pennies saved.

Every new technology passes through an awkward teenage phase—remember early mobile phones? The winners aren't cautious bystanders; they're those brave enough to experiment.

A Vision Worth Pursuing

Soon, agent ROI will become undeniable in select cases, spreading quickly from there. The true breakthrough happens when we stop viewing agents as budget replacements and start seeing them as innovation partners.

That’s how we'll move beyond incremental improvements and start defining entirely new ways to work; and that’s when things get genuinely exciting.