Deep|LLM: Enterprise AI Adoption Survey (Vol. 5) - Cost Optimization and Application Expansion Proceed Together; Business Returns Shape Further Investment

Hi,

Welcome to the fifth edition of our grand enterprise tokenomics study. This week, we interviewed 10 companies, spanning payments, telecoms, payroll and HR services, e-commerce, used-car sales, media agencies, digital marketing, pharmaceuticals, diversified holdings and industrial software. Most still plan to spend more on AI, but the source of the next increase is changing. At companies that have already equipped employees with subscription plans, growth increasingly comes from heavier use of APIs and business applications; elsewhere, expanding employee access remains part of the plan. Companies are also spending less per call, by routing work to cheaper models, negotiating prices and caching. That has slowed growth in total spending, but not stopped it; only the media agency’s spending is flat.

I. Where enterprise AI spending stands today

1. Spending growth over the past six months and the current monthly level

• a) Providing employees with AI tools drove spending growth.For the payment technology services company, the first substantial AI bill came from giving employees access. The rollout started in March and reached all ~2,000 employees by end-June, taking monthly spending from zero to ~US$200,000, entirely from subscriptions. At the payroll and HR services company’s European operation, employee tools carry both subscription charges and extra credit purchases. Those make up most of a bill that has roughly doubled since January to ~US$35,000-50,000/month. Only a small API component serves the customer-facing products.
• b) Wider deployment and business volumes continue to raise application spending.Rolling an established application out to more locations can raise spending without a new use case. At the used-car marketplace (Indian operations), expansion across cities drove ~40%-60% growth over the past 6 months. At the digital marketing technology and services company, spending rose ~15%-30% with new customers and traffic, even after previously implemented savings from model evaluation and switching. The industrial software company is extending agents across internal departments. Its ~18% spending increase, from ~US$170,000 to US$200,000/month, mainly came from APIs. This broader spending measure includes personnel and infrastructure.
• c) Tighter API approvals kept the media agency’s total spending broadly flat.The media agency was the only company in the sample whose total AI spending was broadly flat over the past 6 months. API costs fell ~10%-20%, while subscription spending rose ~10% as new users received accounts. The expert attributed the API decline to tighter use-case approvals for higher-tier, more expensive models. The agency also removed inactive users and keys. These reductions offset subscription growth at a group whose recent AI spending was ~US$9m/month.

2. Subscription versus API mix and employee coverage

• a) Spending mix reflects employee tool rollouts and production applications.The payment company’s agents run inside employee subscriptions, so their use does not yet create a separate API bill. At the digital marketing technology and services company, APIs account for ~60%-70% of spending, mainly for production applications such as customer communications. At the e-commerce operations services company, APIs and usage-based charges together account for ~85%-90%. Subscription costs may include platform licenses and extra usage allowances as well as seat fees, so a high subscription share can reflect both access and additional consumption.
• b) Broad access to AI tools does not mean everyone uses them.The gap between access and active use is visible at the diversified holding company. All ~1,000 employees have AI tools, but active usage is ~60%. The media agency has made Copilot available to almost everyone, with Gemini and Claude also broadly accessible; its active usage rate is ~60%-70%. The pharmaceutical company’s internal assistant reaches ~72,000 monthly users, equivalent to ~90% of employees. That is an active-use measure; the company has not disclosed paid subscription seat purchases.
• c) Incomplete rollouts require more seats, while procurement changes can raise the API share.The telecom operator still has a substantial employee rollout ahead of it. Paid seats currently reach fewer than 10% of employees, and subscriptions mainly cover IT coding tools. Its target of 30%-40% would extend access chiefly to finance, marketing and sales. The payroll company’s European operations, by contrast, plan to replace some tool subscriptions and extra credit purchases with access through an LLM gateway. The proposed solution is expected to cost ~US$40,000/month, broadly comparable to current total spending of US$35,000-50,000/month.

II. How companies budget for AI and control costs

1. Budget status: AI’s share of the IT budget and how budgets are set

• a) Companies fund AI both within and outside IT budgets.At the diversified holding company, AI and IT share the same budget pool because introducing AI tools may reduce other IT spending. AI already accounts for ~20% of the combined allocation, and the company expects that proportion to reach 30%-40% in 3-5 years. The telecom operator also funds AI from an IT budget that management requires to stay flat or decline. A substantial share of this year’s AI spending was unplanned and was absorbed by cutting other IT costs; part of next fiscal year’s IT budget will be redirected to AI. The industrial software company has secured a different source of funding: a dedicated annual AI budget of ~$2.5 million outside IT, obtained after demonstrating AI spending returns to the CTO, CEO and investors. The pharmaceutical company also budgets AI separately from IT, with allocations for individual use cases as well as central spending on common requirements.
• b) Business units increasingly bear application costs.At the pharmaceutical company, the central AI budget pays for common requirements such as security, compliance and training. Business lines increasingly fund agent use cases, making new-project spending harder to forecast, according to the expert. The used-car marketplace (Indian operations) assigns cost burden according to whether a project can be reused: central IT funds shared projects, while business units pay for applications serving only their own operations. The payroll company’s European operations plan to move in the same direction from next fiscal year. Costs are currently centralized within IT, but each department will budget for its own subscription seats and developer usage going forward, within an overall AI budget held flat year on year.

2. Cost controls and model choice by task

• a) Usage controls vary by role and application value.The pharmaceutical company draws a distinction between general development tools and production agents. It caps development tools – tracking tokens and assessing value – but sets no deployment limits for agents with measurable value once safety requirements and task boundaries are clear. Work on those agents then centers on reducing token consumption and improving processes. At the used-car marketplace (Indian operations), controls follow employees’ roles. Token pools replenish daily; engineering, data science and AI teams are largely uncapped, while operations staff have allowances.
• b) Lower model and procurement costs have not stopped total spending growth.The industrial software company negotiated one API purchase program from ~$250,000 to $50,000 without losing service quality. Its broader AI spending, including personnel and infrastructure, nevertheless rose from ~US$170,000 to US$200,000/month as internal applications expanded. The digital marketing technology and services company has also seen total spending rise after cutting model costs. About a year ago, model selection cut its monthly token bill, previously ~$100,000, by more than half. Spending has since grown again with new customers and traffic and has risen ~2%-6% per month in recent months. Its expert described the model choice behind those savings:

‘sometimes the… the most expensive model, or the quote-unquote best model, is not the answer. Maybe it’s the cheap one who behaves better with our use case.’Source: FUNDA expert platform

• c) Proactive model selection and automatic routing play different roles in cost control.The telecom operator makes its model choice when designing each application. Each application has its own key, and a gateway on the cloud vendor’s model platform tracks token costs. Opus handles coding, while Sonnet or Haiku handles text-to-SQL after evaluation. The gateway does not dynamically route requests, and the expert doubts that intent- or keyword-based routing would necessarily deliver the desired results. Intent does determine routing in the diversified holding company’s smart hardware: conversations go to cloud models, while control commands go to small on-device models. For internal agents, the holding company buys routing from vendors.~70% of the e-commerce operations services company’s API spending goes to Opus, with no unified routing mechanism yet. It has not quantified what switching models might save. The payroll company’s European operations have a more explicit plan for changing model choice: a proposed LLM gateway would assign models by user and direct suitable tasks to smaller or open-weight models. Its expert estimates that the gateway and routing optimization together could reduce total spending by ~50%. Achieving that reduction still requires the planned migration; no such savings have been realized yet.

III. Frontier models: where Fable / Astra are used

1. Where Fable / Astra are used, and where existing models are enough

• a) Production favors cheaper models, while explorations can be supported by subscription allowances.The digital marketing technology and services company’s expert uses frontier models for initial planning, then switches to cheaper models for execution. Employees also use frontier models for exploration and coding, almost entirely within subscriptions. Fixed fees per seat mean that frontier spending cannot be separated from the subscription bill. Production calls face a different cost test: customer replies mainly use mid-tier models and DeepSeek, and the expert considers frontier models uneconomical for that work. The telecom operator makes a similar distinction, using Fable / Astra for planning and ideation, Sonnet or Haiku for text-to-SQL, and lower-tier models for summaries.
• b) Capability gains do not automatically trigger large-scale model upgrades.The payroll and HR services company’s European operations tried Fable with developers for ~14 days, then disabled it when credits were consumed too quickly. Its planned LLM gateway addresses a related concern: the company believes vendors’ default use of expensive models weakens cost control, and wants to select models for each use case. The payment technology services company identifies no current need for Fable / Astra. Existing models already handle pricing proposals, data queries and code refactoring. Its expert gives greater priority to data quality, arguing that unreliable underlying data directly limits what AI can do.
• c) R&D and data science support current frontier use; further applications of frontier models remain exploratory.Fable has a defined role in the industrial software company’s R&D work: codebase analysis, complex reasoning and code review, with adoption conditional on outperforming human benchmarks on real engineering tasks. The media agency also has current frontier usage, at ~15% of AI spending, while its APIs mainly support data science. Beyond these current uses, the software company is exploring 3D design generation from text, images and other inputs. The pharmaceutical company is exploring drug design and clinical protocol optimization, with tasks approaching the limits of reasoning still largely experimental. Its expert believes biological and chemical complexity makes these problems difficult to solve through model upgrades alone.

2. Impact of Fable’s zero data retention change and other data control concerns

• a) With very limited Fable usage, no company tied usage to its ZDR change; data retention is handled through contracts and compliance.Fable is little used in the sample: the payment technology services and diversified holding companies do not use it, the payroll company dropped it after a 14-day trial, and the e-commerce operations services company does not use Fable 5.1. None tied usage to Fable’s ZDR change. The three companies that discussed data retention described how they contain the risk. The used-car marketplace (Indian operations) has kept usage steady under its data retention policies. It continues to audit access permissions, personal data masking and model performance roughly every two months, while monitoring agents’ access and actions continuously. The payroll company’s European operations review model suitability for each use case and likewise report no effect on usage. The e-commerce operations services company relies on contractual prohibitions against exposing company data and will not cut usage on that basis, although it cannot verify vendor compliance.
• b) Compliance and security requirements constrain implementation.The payroll company has had to wait on some implementations because of compliance requirements, even though data retention policies have not visibly affected usage. Each use case still requires review. Client data adds another approval step at the media agency, where clients must confirm permitted uses and legal approval constrains model choice. The expert believes clearer protection and legal boundaries could make more room for open-weight adoption. Such restrictions can also determine which models a client can use. Some clients of the digital marketing technology and services company reject models from certain sources; others cannot use particular models because of privacy regulations. The expert says most clients nevertheless use the same model set.

IV. Open-weight versus closed models

1. Open-weight versus closed models: current split, stance and cost

• a) Open-weight models already carry production workloads.The industrial software company’s open-weight plans build on a division of work already in place. It keeps complex reasoning on closed-source models and assigns high-throughput processing, extraction and private-data tasks to open-weight models. These account for ~40% of API workloads; the company plans ~60% next year and targets ~80% within 5 years. The digital marketing company tests model fit against its existing production choice. It adopted DeepSeek after its own evaluation sets showed it outperforming the GPT model then in use, and repeats these evaluations every 4-6 months. Its expert remains skeptical of general benchmark claims.
• b) Open-weight models’ cost advantages are task-dependent.The payroll company’s estimates separate cheaper calls from a lower overall bill. On suitable tasks, small or open-weight models could cost one-third to one-sixth as much as more expensive options. Its ~50% potential reduction in total spending also depends on gateway and routing optimization. Running models in-house brings its own costs: the telecom operator lacks the hardware and skills to do that today and still considers API access cheaper, despite improving open-weight capability. Its expert attributes the current advantage to token subsidies and doubts that it will last:

‘Currently, I believe a lot of the… the API token is still subsidized by the… by the AI labs, right? ... Because it seems to be cheaper if you are consuming API, but I think it will not be forever like that’Source: FUNDA expert platform

2. On-premises and local hosting

• a) Self-hosting provides data control but may cost more.Some of the industrial software company’s clients have data residency requirements, and the company also needs private processing for internal financial information. It meets those needs partly through leased data center capacity in the US and Europe, with related costs of ~US$50,000-60,000/month for hosting, power and hardware maintenance. The move toward self-hosted open-weight models for internal tools is intended to support customization and reduce dependence on a single cloud platform. The company nevertheless believes GPU, hosting and maintenance costs may make this more expensive than cloud solutions. It continues to use the public cloud; self-hosted DeepSeek is still in testing.
• b) Hardware and implementation know-how of employees constrain local inference deployment.The telecom operator considers small models with ~27 billion parameters already competitive, yet lacks the hardware and skills in model architecture and inference serving needed to run them locally. Its expert sees on-premises deployment or rented cloud GPUs as possible options for highly sensitive workloads, but gives no timeline. Renting cloud GPUs primarily entails operating expenditure, while on-premises deployment is more capital-intensive. Current production APIs access closed-source models through two cloud providers’ model platforms, with no in-house inference in place.
• c) Data requirements are handled case by case, mostly without self-hosting.The payroll company’s European operations plan to prioritize cloud deployment after the gateway launches, with compliance reviewed use case by use case. Small-scale local models remain an option for sensitive workloads, but none have been adopted. The telecom operator masks sensitive prompts through a gateway on its cloud provider’s model platform before calling closed-source APIs, and treats on-premises or rented GPUs as a future option only. At the media agency, the only open-weight project known to the expert was specified by a client and ran on a local laptop to keep data inside the company. The location followed that one client’s data requirement rather than a company policy; the agency’s other work runs on vendor-hosted closed-source models. The expert also sees internal work involving no client data as a potential source of greater open-weight spending.

V. ROI: spending decisions, measurement and realized returns

1. Who decides on AI spending, and whether decisions are ROI-driven

• a) Additional funding depends on evidence of business value.Existing project results can support further funding; proposed uses must establish a business case. At the industrial software company, budget increases become available once projects begin to recoup investment, with a 30% ROI hurdle for each solution. The pharmaceutical company makes further funding available after a PoC or MVP demonstrates value, while keeping budgets capped and requiring justification for increases. For proposed API use cases, the media agency requires a business case identifying the client served and sponsorship from a senior manager outside the direct reporting line. Its expert described the aim:

‘let’s create the workflows that actually add value, rather than keep prototyping till the end of time’Source: FUNDA expert platform

• b) AI spending decisions involve management and business teams.Business teams help select AI applications, while management retains final spending authority. At the payment company, regional teams set API and in-house development priorities for roughly the next six months, subject to group approval. Subscriptions follow a different route: group management decides centrally, with no current decision-making role for regional teams. The pharmaceutical company also separates use-case selection, shared by digital and business teams, from final decisions by senior management. At the e-commerce company, technology, product, data and business teams jointly redesign workflows, while decision-makers include the CEO. It has yet to establish a company-wide ROI framework.

2. How ROI is measured and what has been realized

• a) Returns on AI spending show up first as more clients served per head.Serving more paying clients with the same team is a tangible benefit at the media agency, which has not reduced its data science team. The expert illustrated the opportunity with a product that could release 30%-40% of employees’ time for other valuable work. At the e-commerce operations services company, a three-person team now runs three stores instead of one; nobody has been let go, and the focus is redeployment. The payment company has a similar goal of serving more merchants without adding account managers. The diversified holding company counts hours saved, especially overtime paid at 1.5x, as its cost metric. All four are describing gains from subscription tools, and none has converted them into a financial figure: the payment company has not formally calculated or tracked ROI, and the holding company estimated returns before launch without precise calculation. In the payroll company’s European operations, no efficiency baseline exists. The company plans to register use cases and track model usage and costs through an LLM gateway before measuring benefits.
• b) Companies do measure returns: avoided software costs, monetized hours saved and project-level tracking.The telecom operator can compare cash outlays directly because it replaced a software subscription. Its former marketing automation platform charged ~US$200,000-250,000 a year in license fees. The in-house replacement costs US$3,000 a month, plus ~US$5,000 in one-off development spending, excluding labor. The industrial software company’s calculation puts a financial value on employees’ time: it combines hours saved with task volumes and hourly rates. Its ROI formula deducts total ownership costs from financial benefits, then divides the result by those costs. On that basis, reported ROI was ~35% last year and year-to-September. The measure captures the value of saved work even where that value is not a reduction in cash spending.Project results can also differ within the same company. In the used-car marketplace’s Indian operations, ~35%-40% of projects generate positive returns and 40%-45% are near breakeven. Production applications are assessed on revenue, customer experience and efficiency. AI-related business activity is ~1.5x its previous level; the expert did not give a revenue figure. The digital marketing company assesses its customer-facing products differently: dashboards connect AI costs with output quality and track API costs per call, while returns primarily reflect customer retention and conversation quality. Its expert described returns above 2x, without providing the exact multiple or full calculation methodology.

VI. Use cases: which models, which departments, and which industries

1. Model choice by workload types and department

• a) Internal AI workflows have expanded from individual tasks to full processes.The pharmaceutical company’s procurement agent executes the full purchasing process from an employee’s request. This agent and an incident-logging agent, which creates a ticket from a single natural language instruction, are deployed at scale within its internal AI assistant. At the payment company, sales proposal agents follow templates and use a designated pricing file organized by card network, debit or credit, and domestic or international card mix. Their corporate proposals set out payable fees and potential savings. The sales team already uses the office-suite AI assistant widely.
• b) AI use is spreading across departments, while direct adoption in payroll processing remains limited.The industrial software company’s expanding legal, finance and HR applications track client news, analyze clients’ financial information to support renewals and generate onboarding materials. Some of its internal tools use open-weight models, and tasks involving private or financial data are handled in leased data centers. In the payroll company’s European operations, direct AI use in payroll processing remains limited. It has added chatbots and customer service capabilities to client products, while employees use AI for development, research and writing, and presentations.

2. Industry-specific use cases

• a) Domain data and workflow integration shape industry application adoption.The industrial software company aims to generate 3D designs from text, images and other inputs, an application that requires multimodal capabilities to fit customers’ design workflows. The pharmaceutical company is exploring target discovery, genomic and single-cell analysis, generative chemistry and clinical protocol optimization, yet its expert believes biological and chemical complexity makes drug design difficult to solve through model upgrades alone. In health Q&A, the constraint is output quality: the holding company’s expert sees substantial risk in the overly long health answers general-purpose models tend to produce. To obtain reliable, concise responses, the company hosts health Q&A in a dedicated cloud environment, using its own retrieval-augmented generation system and specialized databases.
• b) In payments and livestreaming, error costs and unit economics, not budgets, limit adoption.A transaction-security error can cause losses that are difficult to recover. The payment company therefore remains cautious about real-time transaction decisions, keeping AI mainly in supporting processes outside transactions and post-event analysis, with people involved in complex processes. API projects for agentic payments, transaction protection, fraud and chargeback management are still in preparation. The e-commerce company has not adopted AI livestreaming after discussions with several providers. It found existing solutions expensive compared with human-hosted livestreams, with no clear improvement in returns. Any future offering would primarily serve brand clients interested in trying the technology, rather than target profitability for now.

The next section brings all the above findings together to assess the outlook for total AI spending. We compare each company’s expected spending growth over the next six months with growth over the past six months, alongside its expectations for year-over-year growth in 2027. Our institutional clients also receive a more detailed analysis of the 50+ companies surveyed to date, including comparisons across sectors, industries and regions.