Is Your AI Agent Worth It?

Why Is Spending Rising If AI Keeps Getting Cheaper?

Three factors are behind this paradox:

• Companies are scaling their AI initiatives, which multiplies total usage volume.
• Model providers shifted from subscription pricing to consumption-based pricing, a change that rewards longer answers: more tokens generated also means more tokens billed.
• Many simple tasks are still being handled by expensive models, when cheaper ones could do the job just as well.

As a sense of scale, AI spending is expected to reach close to 25% of companies' technology budgets in the years ahead.

Six Reasons Agentic Spending Keeps Climbing

In agentic processes, token consumption isn't explained by model pricing alone. These are the drivers pushing operating costs upward:

• Context that drags along. Models don't retain memory between calls, so agents resend the accumulated context at every single step. This can make an agentic task consume up to 1,000 times more tokens than a simple chat or reasoning task.
• The cost is in the review. About 60% of an agentic task's spend goes to verifying, correcting, and re-validating results, more than to the initial response itself.
• Autonomy creates variability. The same process can be solved through different paths, with different tool calls and a variable number of retries. In coding tasks, this can create cost differences of up to 30x for the exact same task.
• Extended reasoning on simple tasks. A model's extended thinking pays off on complex tasks, but it's an unnecessary cost on simple ones. Model routing — using the right model at each step — helps avoid it.
• Orchestration adds weight too. How an agent coordinates and distributes work across tools and other agents can change the final cost without changing the business outcome at all.
• The shape of the information matters. Prompt design, context length, language, and data format all affect token consumption. The same exchange in a language other than English can fragment into more tokens and end up costing more.

Lessons for Managing AI at Scale

Once AI consumption reaches a volume of several trillion tokens a month, a few concrete practices emerge for managing it:

• Consumption is concentrated: 10% of users generate close to 65% of total token consumption.
• Pooling consumption spend across the whole organization, instead of individual licenses, supports the heaviest users without leaving idle capacity.
• Older, cheaper models are enough for most tasks. Smart routing between models generates meaningful savings.
• Giving users visibility into the cost of each prompt, together with simple guidelines like asking for more concise answers, can cut token consumption by 30% to 40% without hurting output quality.
• A growing share of AI spend comes from the layers that surround the model: security gateways, orchestration platforms, and monitoring tools, which also bill based on token consumption.

What Changes in the Competitive Dynamic

This redefines competition in four ways:

• Execution advantage becomes harder to sustain. For years, competitive advantage came from executing a process better: faster onboarding, more efficient customer service. Agentic systems let any competitor with access to the same technology replicate that capability at scale.
• Data and context gain value relative to the model. As models converge on similar capabilities, the difference comes down to each organization's own information: decision history, customer signals, operational context.
• AI governance becomes a competitive advantage. The cost variation between teams solving the same task with AI can be enormous, as shown above. Those who manage to measure, allocate, and improve their agents' work will get more value from the same investment.
• The line between what's worth building in-house and what's worth outsourcing shifts. Each business's own context becomes a strategic asset worth keeping in-house, while more standardized tasks can be handled through external services.

The Agenda for Building an Agentic Operating Model

Building an agentic operating model means treating AI economics with the same discipline applied to capital or operations:

1. Allocate intelligence the way you allocate capital: not every process needs the most advanced model. Many tasks can be solved with smaller models, open source, or even traditional automation.
2. Define business strategy before technology strategy, to concentrate investment where AI can genuinely change the economics of the business.
3. Build the agentic operation as a management discipline, with budget, autonomy, and defined stop rules for every autonomous system.
4. Assign clear owners, with differentiated roles: the CIO connects systems and data so agents have the right context; the CTO manages architecture, model routing, and results evaluation; the CFO turns AI into a measurable spending category, with metrics like cost per outcome.
5. Measure agents' work with business metrics, not only technical ones like token volume or model calls.
6. Review what's worth outsourcing and what's worth keeping in-house, activity by activity, based on where the value actually sits.

What This Means for Your Organization

The conclusion is clear: managing agents' work with the same discipline applied to human work, capital, and operations will be what separates companies that scale AI with measurable results from those that just accumulate spend they can't explain.