Emerging markets prioritise top-line growth with agentic AI

Escaping ‘pilot purgatory’

The demand for top-line value in emerging markets reflects the general frustration with generative AI. Despite heavy investments in underlying infrastructure, many enterprises only achieve modest AI gains. According to a BCG study, finance market leaders reported AI returns of about 10%, which is barely above the cost of capital.

Srini Tallapragada, president and chief engineering and customer success officer at Salesforce, who has spent recent weeks interacting with global CIOs and board members, noted that many organisations are also increasingly tired of running AI pilots.

“I call it ‘pilot purgatory’,” said Tallapragada. “Everybody is cool to do quick demos, but if you ask a C-level executive or the board, ‘Did it have an impact on your P&L?’, it’s not there,” he added, referring to profit-and-loss statements that reflect an organisation’s financial performance.

The crux of the ROI issue lies in what Tallapragada refers to as the last-mile gap: a deficit in trusted data context, workflow integration, and guardrails required to make AI agents safe and effective in delivering business outcomes.

To bridge the gap, Salesforce is urging the industry to rethink how it measures AI consumption. Currently, the industry relies on tokens, the basic units of data processed by a large language model, which Tallapragada argued is a flawed metric for business value.

“People are measuring tokens, which are like CPUs – important but not showing value,” he said. Instead, Salesforce is tracking agentic work units, a metric that measures the actual tasks completed by an agent. According to Tallapragada, Salesforce customers have already executed over 2.4 billion agentic work units, which increased by 57% in the fourth quarter alone.

As more AI projects move from pilots to production, the software industry is also grappling with how to charge for autonomous agents. Software-as-a-service (SaaS) suppliers are currently caught between offering consumption-based models and traditional, predictable licensing.

“The pendulum has swung quite dramatically back and forth,” said Parameswaran. “Consumption-based pricing is highly non-predictable. If you remember the early days of the cloud where people had runaway bills, we’re starting to see elements of that in the AI world. Customers want the best of both worlds: they want consumption-based, but they also want it to be predictive.”

To accommodate this, Salesforce is currently offering a mix of pay-as-you-go consumption models and an agentic enterprise licence agreement (AELA) for customers that want unlimited use of consumption-based services, such as Agentforce, Slack and MuleSoft, for a fixed fee over two or three years.

Lisa Singer, principal analyst and vice-president of Forrester, a market research firm, noted that licensing models like AELA show that the value of AI agents is not correlated to usage volume, but to the economic outcomes they enable.

“Accepting short‑term unprofitability only makes sense when a vendor believes that agents materially reshape enterprise cost structures, productivity or growth,” she wrote in a blog post. “The pricing signals confidence that agent-driven value is durable and monetisable over time.”

The real bottleneck

As multi-agent ecosystems mature, enterprises will eventually need a so-called agent fabric or control plane to govern interactions between AI agents built by different suppliers, such as a Salesforce customer service agent needing to query an SAP supply chain agent via the agent-to-agent protocol.

“Everybody is great on day one. I always say the real work starts on day two, day three, day 100,” said Tallapragada. “These agents are running, but who is ensuring they’re running properly? Just like humans, they need to be managed. They can drift. You need traceability, auditability and observability.”

Crucially, as these governance and economic models take shape, the barriers to agentic AI adoption are changing as well. While compute capacity once hindered adoption, the challenge today lies in the enterprises themselves.

“If you asked me a year back, the biggest issue was getting GPUs [graphics processing units]. That is gone, frankly,” said Tallapragada. “The bigger constraint is customers trying to identify use cases where they generate value instead of demos. It’s the internal change management and business process reengineering – that’s the bottleneck more than infrastructure.”