Overview
Enterprise artificial intelligence adoption is entering a new phase: businesses are no longer focused only on how quickly they can introduce AI, but also on whether they can afford to operate it at scale.
A new study published by the Capgemini Research Institute on 12 November 2026 highlights the financial pressure facing large organisations. In a survey of 500 senior executives, 59% said their organisation's AI-related expenditure over the previous 12 months was significantly higher than expected or budgeted.
Among organisations reporting overruns, expenditure averaged 24% above expectations or budget. The study also found that 53% of executives were concerned about AI costs.
The findings come as enterprises increasingly experiment with AI agents, automated workflows, large language models and AI-assisted software development. These systems can create value, but they can also introduce recurring costs that are difficult to forecast when usage grows across teams and business processes.
What the Capgemini research found
The Capgemini Research Institute surveyed 500 C-suite executives in September 2026.
The respondents represented organisations with annual revenue above $1 billion, across 10 countries and 13 industries.
Several findings stand out:
| Finding | Reported result |
|---|---|
| Executives reporting significantly higher-than-expected AI expenditure | 59% |
| Average overspend among affected organisations | 24% |
| Organisations that exceeded AI budgets by more than 30% | 21% |
| Executives concerned about AI costs | 53% |
| Organisations confident in forecasting future AI costs | 56% |
| Organisations planning to expand open-weight model use | Nearly 3 in 5 |
| Organisations calculating Capgemini's proposed AI Yield metric | 1 in 5 |
These figures come from the company's own research and are survey responses, not an independently audited measurement of the entire enterprise AI market.
Why AI costs are becoming difficult to control
One reason AI expenditure can grow faster than expected is that the price of an individual model request does not represent the full cost of delivering an AI-powered business outcome.
A production AI application may involve:
- Model input and output tokens.
- Retrieval and vector database queries.
- Tool calls and external API usage.
- Agent-to-agent communication.
- Repeated reasoning and retry cycles.
- Cloud compute and GPU infrastructure.
- Human review and approval.
- Monitoring, evaluation and security controls.
- Rework when AI-generated results are incorrect.
For example, a SaaS company might initially estimate the cost of an AI support assistant by calculating the expected number of model requests.
Once the assistant begins using retrieval, multiple tools, agentic workflows and human approval steps, its actual cost per resolved customer issue may be much higher.
Capgemini found that only a third of executives could confidently estimate indirect AI costs, including human oversight, rework and governance. It also reported that half of executives said AI expenditure was approved without adequate visibility into the full cost.
Why companies are considering multi-model AI strategies
A major response to rising AI costs is to reduce dependence on a single model provider.
More than three in five executives expect their organisations to adopt multi-model strategies over the next 12 months. Nearly three in five also plan to expand their use of open-weight models.
A multi-model strategy allows a business to select different models for different workloads rather than sending every request to the most expensive available model.
For example:
| Workload | Possible model strategy |
|---|---|
| Basic classification | Small specialised model |
| Document extraction | Lower-cost model with structured output |
| Customer support summaries | Efficient general-purpose model |
| Code review | Model selected through task-specific evaluations |
| Complex reasoning | Frontier model where additional capability is justified |
| Private or regulated workloads | Approved self-hosted or controlled model deployment |
The objective is not to replace every premium AI model. It is to use the appropriate model for each task, based on quality, cost, latency, security and business requirements.
Open-weight models are becoming more relevant
Open-weight models make their trained model parameters available for download and deployment, subject to their respective licences.
This gives organisations more flexibility over where and how models are operated. Depending on the model and licence, businesses may be able to customise, fine-tune or host them within their own infrastructure.
However, open-weight does not automatically mean fully open-source. Access to model weights does not necessarily include the full training data, training code or unrestricted commercial rights.
Self-hosting also does not make inference free. Organisations must still pay for GPUs, electricity, storage, networking, engineering and operations.
For some workloads, a managed closed-model API may remain more economical than maintaining dedicated inference infrastructure. The decision should be based on total cost and measured task quality, not token prices alone.
What this means for developers and SaaS businesses
For SaaS teams, the findings reinforce the importance of treating AI usage as an operational cost that needs continuous monitoring.
A practical approach is to introduce AI FinOps practices alongside existing cloud cost management.
Useful steps include:
- Track cost per completed task rather than only cost per token.
- Record which model and tools are used for each workflow.
- Set usage budgets by team, project and customer.
- Evaluate smaller models against real production examples.
- Route complex requests to stronger models only when required.
- Monitor retries, failed tool calls and unnecessary agent loops.
- Include human review, governance and rework in cost calculations.
- Compare self-hosted and managed inference using total operating cost.
For companies building AI products, this can also improve pricing decisions. A product that charges customers a fixed monthly subscription while generating unpredictable AI usage bills may become difficult to scale profitably.
Capgemini's proposed AI Yield approach
The report also highlights a measurement problem: organisations may know how much they spend on AI without knowing how much verified business value that expenditure produces.
Capgemini proposes an approach it calls AI Yield, which considers the verified value of an AI-assisted outcome after accounting for the full cost of delivering it safely.
The research says only one in five organisations currently calculates this metric.
This approach shifts attention away from the number of tokens processed, prompts sent or agents deployed, toward whether the AI system actually completed useful work.
For example, a customer support assistant that processes thousands of messages is not necessarily delivering a strong business return if most of its responses require human correction.
Measuring successful resolutions, cost per completed workflow and first-pass success can give business leaders a clearer picture of whether an AI investment is working.
Limitations and what to watch next
The survey is a useful indicator of enterprise sentiment, but it has limitations. It covers 500 senior executives at large organisations, not every company or small SaaS business. Its findings are self-reported and do not establish that multi-model or open-weight adoption will automatically reduce costs.
Capgemini also provides enterprise technology and consulting services, so its recommendations should be evaluated alongside independent operational evidence.
The next important indicators will be actual enterprise deployment data, the cost and quality of open-weight inference, and whether companies can connect AI spending to measurable improvements in productivity, revenue or customer outcomes.
The broader shift is clear: enterprise AI adoption is increasingly becoming an exercise in economic discipline. Businesses will still invest in advanced models, but they will increasingly expect evidence that the capabilities they purchase justify the total cost of operating them.
Why this is trending
- Capgemini Research Institute: Published its AI economics research on 12 November, documenting budget overruns, indirect costs and plans for multi-model AI strategies.
- Reuters: Reported the survey's findings on 12 November, highlighting the growing interest in lower-cost open-weight AI models.
No matching prominent Hacker News or Reddit discussion about this specific new report was verified. X was inaccessible, and no matching current daily.dev story was found. This is an important research-driven industry trend, not a claim of viral community traction.
Sources
- Most organizations plan to adopt a multi-model AI strategy, as nearly six in ten faced significantly higher AI costs than expected — Capgemini, 12 November 2026. Primary source for survey results, methodology and proposed AI Yield approach.
- Research Snapshot: The Cost of Intelligence — Capgemini Research Institute, 12 November 2026. Research summary covering cost overruns, visibility gaps and enterprise AI economics.
- Most large firms plan higher open-weight AI use to cap bills, study shows — Reuters, 12 November 2026. Independent reporting on the survey and the open-weight model trend.
Further reading
- Most organizations plan to adopt a multi-model AI strategy, as nearly six in ten faced significantly higher AI costs than expected — Capgemini, 12 November 2026. Primary source for survey results, methodology and proposed AI Yield approach.
- Research Snapshot: The Cost of Intelligence — Capgemini Research Institute, 12 November 2026. Research summary covering cost overruns, visibility gaps and enterprise AI economics.
- Most large firms plan higher open-weight AI use to cap bills, study shows — Reuters, 12 November 2026. Independent reporting on the survey and the open-weight model trend.
Building a product in this space?
MORE FROM THE NEWSROOM
Related news
SaaSUsage-Based Billing Becomes the Default for SaaS Teams
Seat-based pricing is steadily giving way to metered and hybrid models across the SaaS industry. The shift is as much an architecture decision as a revenue one — usage metering has to exist somewhere in the stack.
Software DevelopmentJPMorgan Chase Traces Banking Delays to Software Bug in On-Premise Systems
JPMorgan Chase experienced transaction-processing delays on 16 November after a software bug affected its on-premise hardware systems. The bank said customers retained access to their accounts and funds, while some transactions involving other financial institutions could also experience delays.
CybersecurityMalicious Optimizely Scripts Reached 178 Organizations, Exposing Website Visitors to Data Theft
A malicious JavaScript campaign distributed through Optimizely Web reached visitors across 178 organizations, according to security researcher Scott Helme. The payload included cryptocurrency address substitution, payment-card skimming, cookie theft and iOS exploit delivery mechanisms, highlighting the security risks associated with trusted third-party website scripts.
