Enterprise adoption of agentic software — AI systems that can independently plan and execute multi-step tasks — is accelerating. Vendors including Microsoft, Salesforce and a wave of startups are embedding agentic capabilities into productivity suites, customer service platforms and supply chain tools. The promise is compelling: reduced human labour, faster decision-making and round-the-clock operations.
Yet a less discussed reality is emerging from early enterprise deployments. Agentic software introduces a new category of operational cost that traditional automation budgets do not account for. These costs stem from the inherent unpredictability of large language model (LLM)-driven agents: they can fail, hallucinate, get stuck in loops, or produce outputs that require human review. Without a dedicated risk reserve, finance leaders risk budget overruns, stalled rollouts and eroded trust in AI investments.
The Nature of Agentic Costs
Traditional automation — robotic process automation (RPA), fixed workflow tools, rules-based scripts — has predictable cost profiles. Licensing, infrastructure, implementation and maintenance are well understood. Agentic software is different. Each task an agent attempts may consume variable compute resources, incur API call fees, and require human intervention when the agent fails or produces an ambiguous result.
Early adopters report that agentic systems can require multiple retries per task. A single customer query handled by an agent might involve several LLM calls, database lookups, and fallback logic before reaching a resolution. Each retry consumes tokens and compute time. In a production environment handling thousands of requests, these costs compound rapidly and are difficult to forecast.
A second cost layer is oversight. Agentic systems require monitoring, logging and human review of edge cases. Organisations deploying agents in customer-facing or regulated contexts must invest in guardrails, audit trails and escalation workflows. This is not a one-time setup cost; it is an ongoing operational expense that scales with agent activity.
Why Traditional Budgeting Falls Short
Most enterprise automation budgets are structured around fixed costs: software licences, implementation fees, and a contingency line item of 10-20 per cent. Agentic software does not fit this model. Its cost drivers are variable, non-linear and tied to the complexity of tasks rather than the number of agents deployed.
A finance team budgeting for 100 agentic workflows might assume each workflow costs the same. In practice, a simple data retrieval task may cost pennies, while a multi-step negotiation or triage task could cost orders of magnitude more due to retries and LLM token consumption. Without historical data, these variances are nearly impossible to predict.
Furthermore, agentic systems can exhibit emergent behaviours that increase costs. An agent designed to optimise a supply chain might, for example, generate hundreds of alternative scenarios before selecting one. Each scenario consumes compute and API resources. The cost of such exploration is not visible in a traditional budget line.
Commercial Impact: Who Bears the Risk?
The financial risk of agentic software is distributed unevenly across the enterprise. Procurement teams may negotiate fixed-price contracts with vendors, but the variable costs of LLM usage, cloud compute and human oversight fall on the operating unit deploying the agent. This creates a misalignment: the business unit that champions the automation bears the cost overruns, while the vendor captures the upside of increased usage.
For startups building agentic products, the cost structure is equally challenging. Many offer usage-based pricing tied to tokens or API calls. As customer adoption grows, so does the vendor's infrastructure cost. If the vendor has not priced in the risk of high-retry scenarios, margins can erode quickly. Some vendors are responding by introducing tiered pricing or caps on retries, but this shifts risk back to the customer.
Enterprise buyers should therefore scrutinise pricing models carefully. A fixed per-agent monthly fee may appear attractive but could mask significant usage variability. Conversely, a pure consumption model exposes the buyer to unpredictable spikes. Hybrid models — a base fee plus a consumption cap — are emerging as a compromise, but they are not yet standard.
Why It Matters
Agentic software represents a genuine productivity opportunity, but its financial characteristics are fundamentally different from prior automation technologies. Finance leaders who treat agentic budgets as extensions of existing automation budgets risk underestimating total cost of ownership by a wide margin. This can lead to stalled projects, unexpected capital calls, or a retreat from AI investment altogether.
Moreover, the unpredictability of agentic costs complicates ROI calculations. If a finance team cannot forecast the cost of a given automation initiative, it cannot reliably calculate payback periods or compare it against alternative investments. This uncertainty may slow enterprise adoption at a time when competitive pressure to deploy AI is high.
Risks / Unknowns
Several unknowns remain. First, the pace of LLM cost decline is uncertain. If token prices continue to fall, the variable cost risk may diminish. However, if agentic systems become more sophisticated and consume more tokens per task, costs could rise even as unit prices fall.
Second, the insurance and risk transfer market for agentic failures is immature. It is not yet clear whether standard cyber or professional indemnity policies cover losses caused by autonomous agent errors. Enterprises may face uninsured liabilities.
Third, regulatory developments could impose additional costs. The EU AI Act and similar frameworks may require human oversight, explainability and audit trails for agentic systems, all of which add operational expense.
FY Outlook
Over the next 12-18 months, we expect three developments. First, a new category of financial software will emerge to track and forecast agentic costs in real time, similar to cloud cost management tools. Second, enterprise procurement teams will begin demanding cost caps, retry limits and transparency from agentic vendors. Third, a standard methodology for calculating total cost of ownership for agentic systems will develop, likely led by consulting firms and industry bodies.
Finance leaders should begin building risk reserves now. A prudent approach is to allocate 30-50 per cent above the estimated baseline cost for any agentic deployment, with a quarterly review cycle to adjust as usage data accumulates. This is not a sign of pessimism; it is a recognition that new technologies carry unknown cost profiles until they are operated at scale.
Conclusion
Agentic software is not a cost substitute for human labour; it is a new category of operational expense with its own risk profile. Enterprises that treat it as such — budgeting for variability, investing in oversight, and negotiating transparent pricing — will be better positioned to capture its benefits without financial surprises. Those that do not may find that the hidden cost of autonomy outweighs its promise.
For further reading on AI cost management, see our analysis of LLM pricing trends at /category/ai-economy and our guide to AI vendor procurement at /category/business-tools.
