Deep|LLM: Enterprise AI Adoption Survey (Vol. 6) - In Our Sample: AI Spending Growth Expected to Slow; Routine Work Shifts to Lower-Cost Models

Hi,

Welcome to the sixth edition of our enterprise AI adoption and spending study. This issue draws on expert interviews conducted between September 19 and October 11, 2026, covering six companies.

Our sample includes three mid-sized companies in sports media, fintech and auto lending, and three mega groups in pharmaceuticals, brewing, and pharmacy and health insurance.

Core findings

AI spending growth is expected to slow at most companies in this sample. The fintech group’s spending growth is expected to slow after a rapid initial AI rollout. At the pharmaceutical group, better model selection and procurement terms are expected to curb spending growth. Spending growth at the pharmacy and health insurance group is expected to remain broadly steady.

Claude-based workflows can replace specialist software for feature engineering. The subprime auto lender’s in-house workflow lets its team direct the search for variables, require explanations of their business meaning and account for compliance constraints. With that workflow meeting its needs, the lender plans to cancel or not renew a specialist tool costing ~$80K-100K a year.

Matching models to tasks can cut AI spending even as token use rises. After training, some teams at the fintech group roughly halved AI spending while using more tokens. The pharmacy and health insurance group’s expert estimates that better routing and model choice could reduce API costs by 20%-30% for the same workload; its automatic router is still under evaluation.

Some companies tie additional AI funding to expected cost savings, revenue gains or less manual work. The brewing group asks business owners to explain expected cost savings or additional revenue and whether they will commit to changing workflows and reducing headcount after full implementation. At the pharmacy and health insurance group, proposals to scale a use case must explain how it will reduce manual processing, handling time or rework.

Most companies in this sample have not adopted Fable for routine production; one interviewed company reports using Astra for cross-system commercial analysis. Cost, a lack of recurring tasks that justify it, and data-retention terms constrain broader Fable adoption. The mega brewing group uses Astra to combine market strategy, brand plans and financial information with customer-pricing information.

Self-hosted open-weight models already handle about half of API calls at one interviewed company. The pharmaceutical group has incorporated model downloading, scanning, verification and tuning into its internal process. Its expert roughly estimates model-use costs to be ~90% lower than with external APIs. These workloads avoid external per-token charges but still incur internal deployment and operating costs.

I. AI Spending Growth and Budget Expectations

1. AI spending growth and budget expansion

• AI spending growth is expected to slow at most companies in this sample.The fintech group expects slower growth as its initial rollout matures. The brewing group expects less hiring to slow subscription spending growth. At the pharmaceutical group, better model selection, fewer unnecessary loops and renegotiated procurement terms are expected to curb spending growth. The pharmacy and health insurance group is an exception, with growth expected to remain broadly steady.

2. Observations on AI-Related Software and Platforms

• The auto lender plans to cancel or not renew its specialist feature-engineering software subscription after building a Claude-based replacement.Its Claude-based workflow lets its team specify the tables and variable combinations to analyze, require explanations of their business meaning and account for compliance constraints, then turn the results into management reports. The purchased specialist tool generated complex statistical combinations whose business meaning was difficult to explain. With the in-house workflow now meeting its needs, the lender plans to cancel the ~$80K-100K annual subscription or let it expire. Its expert described the result:

  ‘The tool has clearly become redundant, hasn’t it? We no longer have a use for it.’

  Source: Full interview transcript on FUNDA expert platform.
• Enterprise AI applications are increasing demand for data and analytics platforms such as Databricks and Snowflake.At the school-sports streaming platform, Snowflake Cortex enables AI-assisted analysis within the company’s existing data and access-control environment. Employees can ask questions in natural language, reducing integration work and time spent writing SQL. The expert has seen more requests for reports and models, and the team uses the time saved to handle them without reducing staff. Cortex adds consumption charges within the Snowflake contract. Separately, a change to tools used on the email platform is reducing other Snowflake spending.
  The pharmacy and health insurance group uses Databricks, Snowflake and cloud data infrastructure for data preparation, model development, orchestration and some model serving. Their combined monthly cost rose from ~$20K-25K six months earlier to ~$35K-40K and is expected to reach ~$50K-60K in another six months. The total includes compute and storage for both AI and analytics.
  
• The sports streaming platform plans to give all employees access to Gemini through Google Workspace.Its Google Workspace agreement is awaiting completion. Gemini is expected to support routine documents, summaries and email. Developers currently use Claude Code, and production workflows rely mainly on Anthropic models. Adopting another model for those tasks would require comparing output quality and overcoming established habits. The platform’s expert explained why the team continues to use Claude for development and production:

  ‘The honest answer is it got here first. We started using it, everybody got used to it, and we didn’t make the change.’

  Source: Full interview transcript on FUNDA expert platform.
• The sports streaming platform has proposed cutting Atlassian spending as its use of Jira and Confluence declines.The proposed reduction forms part of the platform’s periodic review of software usage, technology spending and budget allocation.

II. Subscription and API Spending

1. Sources of additional subscription and API spending

• Expansion into core business workflows is driving API demand in two interviewed groups.At the brewing group, API use is expanding beyond proofs of concept, knowledge search and chatbots into core workflows, multimodal tasks and agents. Subscriptions mainly support employees’ personal productivity. Some Microsoft subscriptions carry two- to three-year commitments, whereas API costs vary with consumption. The brewing group’s expert described the distinction:

  ‘the subscription one is just more like personal productivity. kind of use cases, whereas the API ones are more changing the workflow, or the business process, etc.’

  Source: Full interview transcript on FUNDA expert platform.
  At the pharmacy and health insurance group, more workflows are calling models in production. Monthly API spending rose from ~$20K-25K to ~$45K over the past six months, increasing its share of subscription and API spending from ~20% to 30%. Subscription spending rose more gradually, from ~$90K-95K to ~$105K over the same period.
• Broad employee use and new model releases have shifted the pharmaceutical group’s AI spending toward subscriptions.Subscriptions’ share of combined subscription and API spending rose from ~20% at the start of 2026 to ~75% at the time of the interview. The expert links the change to new model releases, broad employee use and limited visibility into token consumption within subscriptions.

III. Budget and Usage Management

1. Budget formation and approval

• All six companies have budget or approval controls in place for AI spending.
  Five companies are classified as Budgeting and the school-sports streaming platform as Budgeting and Maxxing. The platform encourages adoption while using contracts, production budgets and overage alerts to manage costs. Encouraging use and controlling spending can operate together.
  
• Some companies assess further AI investment against expected savings, revenue gains or reductions in manual work.Before approving additional AI spending, the brewing group asks whether each dollar could generate $2-$3 in cost savings or additional revenue. Management also asks business owners whether they will commit to changing core workflows and reducing headcount after full implementation. The group’s expert described the requirements:

  ‘are we going to get two, three times the value for every dollar we spend? on AI, and is the business going to also take responsibility to change their core process, workflow when the AI is fully implemented, and will they commit to make that headcount reduction?’

  Source: Full interview transcript on FUNDA expert platform.
  The pharmacy and health insurance group allows small exploratory pilots before returns are clear. To scale a use case, business owners must explain the expected reductions in manual steps, handling time or rework. The expert explained how a strong use case can obtain additional funding:

  ‘IT sets the guardrails and enables the platform for us, but ROI drives the incremental spend. If the use case has a strong payback. The budget can expand even if the base IT budget is literally flat.’

  Source: Full interview transcript on FUNDA expert platform.
  The fintech group plans quarterly reviews comparing departmental AI spending with expected and actual outcomes to decide whether to continue funding. The CEO retains final budget authority.
  The auto lender’s expert expects AI access to extend to eligible employees and simpler applications to be explored over roughly the next year. If revenue gains or ROI remain unclear by then, the expert believes management may pause further investment. The company currently encourages small trials and has no general quantified economic-return threshold for approval.

2. Usage and cost management

• Matching models to tasks can cut AI spending even as token use rises.
  Employees at the fintech group had been choosing expensive models, including Opus, for routine reports. After model-selection training, some coached teams roughly halved spending while token use rose. An in-house gateway also assigns models by request, and daily dashboards track departmental spending and tokens. Separately from the observed team results, the expert estimates at least 30% potential savings from better model choice. The expert described the results:

  ‘some of the team we, we help on, they actually reduced the so-called dollar value by half, but the token actually goes up.’

  Source: Full interview transcript on FUNDA expert platform.
  The brewing group’s in-house router uses Microsoft capabilities to send simple tasks to lower-cost models. Some workflows also hide model choices from employees, preventing them from manually selecting the highest-priced option. At the pharmaceutical group, a gateway directs routine work such as PowerPoint preparation and summaries to cheaper models.
  The pharmacy and health insurance group still relies mainly on manual model selection, with an automatic task router under evaluation. Its expert estimates that moving suitable classification and extraction work from Sonnet or Opus to Haiku or open-weight models could reduce API costs by 20%-30% for the same workload, or ~$9K-14K against the current ~$45K monthly API bill. Cleaner context, cached results, shorter prompts and fewer unnecessary agent loops offer a separate potential 10%-20% reduction in token volume.

IV. ROI Measurement and Outcomes

1. Measuring returns and reported outcomes

• The pharmacy and health insurance group’s expert reports faster handling of complex claims and less time spent on nursing documentation and administration.The expert reports more than 20% shorter handling time for complex claims. In the nursing workflows described, assistance with documentation and administration frees up to ~90 minutes per nurse per day for care, case management or other work. These are operating outcomes; net financial ROI has not been quantified.
  
• The brewing group’s expert believes ERP simplification and SAP standardization may deliver higher returns than AI projects.The group continues to invest substantially in these non-AI projects. Its expert considers them necessary for AI to deliver its full benefits.

V. Frontier Models

1. Frontier-model adoption and task value

• Most companies in this sample have not adopted Fable for routine production.Recurring tasks that justify its cost remain limited, while data-retention terms restrict its use with sensitive data. The pharmaceutical group’s trials remain confined to non-sensitive data. Its expert highlights Fable’s capabilities in agent orchestration, coding and executive communication, and Astra’s in mathematics; actual Astra use at the group was not established.
  
• The brewing group reports using Astra for commercial analysis across systems.Market strategy, brand plans and financial information are spread across multiple systems and were previously difficult to assemble into a complete analysis. The expert says Astra brings that material together with customer-pricing information to support commercial strategy development. The expert described Astra’s role:

  ‘So, what Astra unlocks is it kind of becomes your, you know, chief of staff, which can do multi-step reasoning. And it is able to exercise judgment across different systems.’

  Source: Full interview transcript on FUNDA expert platform.
• Lower-cost models can already handle routine summaries, retrieval and extraction.The brewing group’s expert considers mid-tier models or smaller open-weight models sufficient for ~70%-80% of tasks. Microsoft MAI and older GPT versions handle routine reasoning, knowledge retrieval, RAG and summaries.
  The pharmacy and health insurance group’s expert considers current models sufficient for summaries, coding assistance and extraction. As capabilities improve, the expert expects the strongest models to be reserved for multistep reasoning, long-context analysis and difficult work where better answers reduce subsequent human review. The expert estimates that these difficult tasks account for ~10%-20% of all tasks.

VI. Open-Weight and Closed-Source Models

1. Open-weight use and expansion plans

• Self-hosted open-weight models handle about half of API calls at the pharmaceutical group.The group runs an automated process that downloads models, scans and verifies them, then tunes them for internal use in regional data centers and on company servers. Open-weight models handle ~50% of API calls, with those calls up ~20% from the preceding month. The expert roughly estimates model-use costs to be ~90% lower than with external APIs. These workloads avoid external per-token charges but still incur internal deployment and operating costs.
• Open-weight use is expected to expand substantially at two companies by end-2027.At the fintech group, use remains mainly AI-team trials, accounting for ~5% of API spending. Its expert expects the share to reach 5%-10% at end-2026 and 50% or more at end-2027. The group plans to deploy more open-weight models on cloud servers, replacing some external per-token charges with server costs and gaining greater data control; it has no plan to build a company-owned data center.At the pharmacy and health insurance group, open-weight models already handle classification, extraction and some internal summaries, accounting for ~25% of API spending and 35% of usage. Its expert expects the API-spending share to reach 35% at end-2026 and 45% at end-2027. Expansion must meet requirements for task quality, security, reliability and support. Use remains mainly in approved enterprise cloud environments such as Azure and AWS, with negligible on-premises deployment in the expert’s view.

2. Barriers to adopting open-weight models

• Technical-support concerns are holding back open-weight adoption at the sports streaming platform.The expert has considered self-hosting, but management prefers familiar vendors and wants to know who would help if a model failed. The concern remained when the interviewer raised access through Amazon Bedrock. The platform’s expert summarized the issue:

  ‘the problem always comes back to, if something breaks, who are we calling?’

  Source: Full interview transcript on FUNDA expert platform.
• The auto lender is not considering open-weight models in the near term because of security and data-privacy concerns.When asked about US-origin or Microsoft-hosted options, its expert still reported no plan to try them.

VII. Applications

1. Adoption and expansion in industry workflows

• The pharmaceutical group uses knowledge graphs to help models and agents retrieve drug-research information faster.
  Its graphs connect diseases, genes, proteins, pathways and prior studies, helping researchers find material relevant to their hypotheses. The expert sees the current benefit mainly in less search and analysis time; the hypotheses still require experimental validation.
• The pharmaceutical group uses OpenAI and Claude vision models to check packaging seals and other quality requirements.The expert says these checks help verify that packaging meets the group’s quality requirements.
• LLMs speed up credit-model research at the auto lender, while lending decisions remain with traditional models.The lender uses LLMs to read research papers, write code to reproduce their algorithms, and pull internal data to test whether the algorithms improve its credit models. In the expert’s experience, it can now test algorithms from 5-10 papers in 30-60 minutes, work that once took about a week per paper. Loan approval and pricing still rely on traditional machine learning because applicants must be able to understand rejection reasons and models must withstand regulatory audits. The company is designing a process to generate model-audit documentation with AI; direct LLM analysis of raw credit-report text for lending decisions remains an untested idea.
  
• The auto lender’s expert credits voice AI payment reminders with reducing the need to expand its service team.Third-party voice AI handles scripted reminders for payments 1-2 weeks late. The expert believes this has reduced the need to keep expanding the service team and lets existing staff move to other work. Longer arrears and negotiations over extensions or repayment commitments remain with people because the company does not yet trust AI to handle them independently.
  
• The sports streaming platform plans self-service, AI-edited season highlights to increase use between games.The platform plans to let schools and students create athletes’ season highlights for coaches through a self-service feature. It aims to increase non-game-day use and explore higher subscription prices. Creating individual highlights currently requires someone to supply footage and timestamps before automated editing. Scoring-play replays are already generated and sent to fans automatically. The planned self-service offering would initially retain human review.

The final section compares the spending outlooks: how growth is expected to change, which costs drive the increase, and what the companies have said about 2027.