Deep|LLM: Enterprise AI Adoption Survey (Vol. 3) – Diverging Spending Outlooks, with Production Use Cases Scaling Alongside More Granular Cost Controls

This week, we continued our channel checks on AI spending at 10 enterprises. Companies are still extending AI access to more employees and embedding it in customer service, recruitment and internal business systems. At the same time, management teams are asking more pointed questions: which employees need the most expensive models, what additional output higher usage actually delivers, and whether time savings can translate into revenue growth or real cost reductions. These questions will directly shape the next round of AI budget allocation. Some companies plan to reinvest savings from optimization in new projects, some are shifting from fixed license fees to usage-based spending, and others are moving workloads to local models to bring paid-model bills down.

Account rollout alone is no longer enough to assess the next stage of enterprise AI growth. We also need to understand who is using the tools consistently, which business use cases merit additional investment, and which models will absorb the incremental workloads and how they will be deployed.

Our latest channel-check findings are below:

I. AI penetration

1. Broad access to basic capabilities, with tiered access to specialist tools, is common across this sample set, but paid coverage still varies widely

a) Mid-sized UK Payment Infrastructure Organization A provides basic AI tools to all ~350 employees, mainly for day-to-day work within the Microsoft ecosystem. About 30% can use development tools and more advanced models, while the highest access tier covers only ~10%. The company’s more complex internally developed use cases are still at the POC and prototype stages, and advanced users have not yet exhausted their existing quotas.

‘For the basic, 100% of employees have it. For more sophisticated, maybe there will be 30%, and for the top notch it will be maybe 10%.’‘Our super users are very few, and it’s more on the POC side at the moment - more building proof of concepts or prototypes. It’s not yet real production for the very sophisticated use cases that we are building ourselves.’

b) Large UK Legal Services Firm A provides ChatGPT Enterprise to all ~500 corporate employees. Weekly active usage is ~85% across paid tools and ~96% in priority teams such as sales, marketing, R&D and operations. About 10% - 15% of employees have both OpenAI and Anthropic accounts, mainly in development, marketing and roles where a clear use case has been established.

c) All ~1,000 employees at Mid-sized Middle Eastern Insurtech Company A can use Gemini within Google Workspace and the company’s internal AI tools; ~300 also have paid accounts for tools such as Claude. Access is tiered by role. General employees use basic versions, while developers can request higher Claude Code quotas or API access.

d) Large Asian Gaming Company A has opened a company-wide AI gateway to more than 7,000 employees, all of whom can use paid APIs. Access is split into Basic and Pro tiers. Basic covers non-R&D employees, while Pro provides access to premium models and requires a separate application.

e) Paid coverage is lower at the other companies. About 10% of white-collar employees at Large European Logistics Company A currently have paid Copilot accounts, with a target of ~20% by end-2027. Large Australian Pharmaceutical Distributor A has ~400 employees with additional paid Copilot accounts, or ~20% of its estimated office workforce. At Large Swiss Consumer Goods Manufacturer A, ~250 employees regularly use AI and hold relevant accounts, or ~10% of total employees, mainly in office roles. These companies are still rolling out accounts in phases, using training outcomes and actual usage to decide whether to expand access.

2. A clear gap remains between account coverage and actual usage depth, with engineers and a small group of heavy users accounting for most usage

a) Office employees in Europe and North America at Large European Retailer A can use or request access to an internal AI gateway, but weekly active usage is ~30% or slightly lower. Usage in software engineering is close to 100%, and Copilot is part of engineers’ day-to-day requirements. The small number of engineers who reach their token limit can request more quota.

‘For office employees, around 30% are active users, meaning they use it every week. For software engineering, usage is 100%; engineers are required to use it.’

b) About 10% - 20% of knowledge workers at Large Southeast Asian Telecom Operator A use paid AI tools. Among AI users, ~10% - 15% are heavy users but account for ~50% - 55% of total usage; another ~25% - 35% are medium users. The remainder use AI less frequently, with some light users active only four or five days a month.

‘There are around 10% to 15% heavy users and around 25% to 35% medium users... The high-end users take around 50% to 55% of the overall usage.’

c) Usage is similarly concentrated at other companies. Heavy users at Mid-sized Middle Eastern Insurtech Company A consume ~8 - 10 times as much as general users. At Large Swiss Consumer Goods Manufacturer A, ~10% of platform users are heavy users, and most measurable productivity improvement is concentrated in this group; the impact on general office users remains limited. The top 20% of users at Large Asian Gaming Company A also consume most tokens, with developers and internal AI implementation teams using the tools most intensively.

d) Some companies are using training and license reallocation to raise actual usage. After several months of training, average weekly time saved per Copilot user at Large European Logistics Company A rose from ~30 - 40 minutes at the start of the year to ~3 - 3.5 hours. Accounts inactive for 1.5 - 2 months are reassigned. Large Australian Pharmaceutical Distributor A also removes accounts that remain idle and uses training and internal case studies to find employees with clear use cases. Both cases show that training and ongoing user selection still materially affect actual penetration, beyond simply adding more accounts.

3. Agent and business workflow penetration varies widely. A few companies have limited production deployments, while most remain at controlled pilot or selective expansion stages

a) Large UK Legal Services Firm A has connected Gong, Salesforce and its talent-matching platform through an internal integration layer. Sales, business development and talent teams can query information directly in ChatGPT or Claude and use it for client screening, outreach and meeting preparation. AI usage roughly doubled within 30 days of launch. Employees still have to connect the individual tasks manually, however, and there is not yet a single agent that can complete the full workflow end to end.

‘That caused basically a doubling in usage within 30 days because people saw this and said, okay, now I don’t have to go to all these platforms.’‘We’re not using it for workflows currently... It’s very disjointed. It’s not the same agent doing the different steps of the work.’

b) Large Australian Pharmaceutical Distributor A has put several agents into production. A service-desk agent identifies and routes tickets automatically, reducing reliance on an external provider. The cybersecurity team also uses an agent to collect vendor information and conduct an initial assessment, with an employee reviewing the final report. A separate customer-service agent can already handle ~30% of customer inquiries automatically.

‘We used AI to automate the triaging of the tickets from the service desk, which helped us to reduce the reliance on the external vendor.’‘Our AI agents can go and do the vendor assessment... and then create a report after that. Someone human at the end looks at it.’

c) Large Indian HR Services Company A uses a recruiter agent for the first interaction with candidates. The agent gathers job preferences, handles initial outreach and first-round screening, and then passes suitable candidates to recruiters. Our channel checks suggest this is one of the more mature business use cases because it expands the number of candidates the company can reach and process.

‘A lot of those starting touch points have now become recruiter agents... They tell their preferences. They have their first level of interview or screening, and then they move forward.’

d) The product teams at Large Asian Gaming Company A have developed agents for cross-language translation, LQA testing and animation configuration. They are used for version acceptance, translation quality checks and automated animation generation. Product staff have built and deployed some internal tools directly.

e) The remaining companies still rely mainly on central-team development or controlled pilots. The product team at Mid-sized Middle Eastern Insurtech Company A has put voice and conversational agents into production, but employee-built automations usually serve only one or two people and are difficult to replicate across the company. The main constraints are output consistency, data permissions and system integration. Large European Logistics Company A plans to train ~50 - 60 employees who understand specific business processes, then have them start building departmental agents in early 2027. Complex agent use cases at Mid-sized UK Payment Infrastructure Organization A also remain mainly at the POC and prototype stages.

‘My projects are already in production, but the individual projects which people are working on work best for one or two people; there is no path for them to scale it up.’

II. Open-source model usage

1. Most companies are still testing or using open-source models on a limited basis, although one has established meaningful internal production usage and another has a clear production migration plan

a) Large Southeast Asian Telecom Operator A has the clearest plan in this sample set. All production traffic still runs on third-party proprietary-model APIs, while open-source models account for ~5% - 6% of internal R&D testing. The company plans a phased production rollout from early 2027, starting with ~5% of low-risk internal workloads, then raising the share to 10% and 15% before migrating customer-facing applications. If model quality meets its requirements, it could move ~30% - 40% of current proprietary-model API usage to open-source models in 2027.

‘Currently we don’t have any workload on the deployed open-source models... Our whole usage is API-based usage, which we have taken from third-party closed models.’‘Maybe up to 30% to 40% of the usage which we are today doing on API-based closed models, we will be moving to open-source models.’

b) Large Asian Gaming Company A has deployed GLM 5.2, DeepSeek and K3 on company servers and changed the default model in its internal AI gateway from Sonnet 4.6 to GLM 5.2. In early August, locally deployed models accounted for ~16% of total token usage, mainly for day-to-day office work and lightweight tasks. Proprietary models still handle most higher-stakes or time-sensitive production tasks.

‘The locally deployed models accounted for around 16% of the company’s total token usage.’

c) Most other companies have not reached scaled production use. Open-source models account for ~20% of API testing at Large Indian HR Services Company A, but this is limited to engineering POCs and represents a smaller share once platform spending is included. Mid-sized UK Payment Infrastructure Organization A cited only one limited use case for extracting external data.

‘On the API part, we are still testing with like the 80/20... But on the platform and all, it’s all closed.’‘Nothing that we have scaled.’

2. Compliance and security remain the main barriers to production use, with industry-specific constraints

a) Large European Logistics Company A has the clearest restriction. Employees are currently prohibited from using open-source models because of information security, compliance, data residency and limits in the company’s current technical capabilities. The company does not plan to adopt them even through Azure Foundry or AWS Bedrock, and expects no change in 2027.

‘We decided not to go ahead with any kind of open-source models... Nobody should be allowed to use open-source models.’

b) For legal-services and payment-infrastructure companies, risk reviews also cover client data, service regions and inference location. Large UK Legal Services Firm A may explore open-source models next year, but information security and legal reviews for existing vendors usually take two to three months, and some clients restrict which models and providers can be used. Mid-sized UK Payment Infrastructure Organization A requires inference to run in Europe and preferably in the UK.

‘OpenAI’s review took like 2.5 months and Anthropic’s review took three... It’s going to take a few months just for all of the different legal reviews we have to do.’

c) Large Swiss Consumer Goods Manufacturer A does not use open-source models. In addition to information security, the company considers model origin and geopolitical risk. It is relatively more comfortable with US open-source models but does not currently have dedicated time or staff for an assessment.

3. Some enterprise platforms already offer open-source models, but actual usage and production deployment remain difficult to verify

a) Large European Retailer A can already access Kimi and other open-source models through Copilot and its internal AI gateway, and some software engineers have started experimenting with them. The platform displays cost tiers, token prices and credit usage for each model, with Kimi consuming materially fewer credits than some proprietary models. The channel check could not confirm actual usage shares, however, and most usage still runs on OpenAI and Anthropic.

‘Copilot also includes Kimi. I increasingly hear that many software engineers have started using it, but I think the majority still use OpenAI and Anthropic.’

b) Large Australian Pharmaceutical Distributor A has the technical ability to adopt open-source models through Azure, AWS or Databricks, but the channel check could not confirm which specific models are used in production today. Current usage and spend are therefore recorded as N.A.; platform access alone is not counted as production usage.

4. Most companies still prefer to access open-source models through cloud platforms, while a few with stronger technical capabilities have deployed models themselves or set out clear plans

a) The internal AI team at Large Southeast Asian Telecom Operator A has deployed and tuned open-source models on Tencent Cloud and Alibaba Cloud infrastructure and VPCs in Jakarta to keep data in-country. Self-deployment in this case means that the internal team manages the models on cloud infrastructure, not that the company owns on-premises GPU servers.

‘The open-source models we are doing by ourselves... Currently the model which we are building is on Tencent, but the team working on the model building and tuning is our internal AI team.’

b) Large Asian Gaming Company A has deployed GLM 5.2, DeepSeek and K3 on local servers, mainly for day-to-day office work and lightweight tasks.

c) Other companies are more cautious. Mid-sized Middle Eastern Insurtech Company A has assessed on-premises GPUs and self-deployment, but current usage is too small to support dedicated infrastructure, and some open-source models still fall short in complex agent use cases. It therefore prefers hosted models from regional cloud providers. Large Indian HR Services Company A currently tests mainly through Azure and APIs and has no on-premises deployment plan.

‘The best way to work is to get a provider who has hosted models and give us access to it on a cloud basis... Billing would be lower, but again, it’s not zero.’

d) As in our first two rounds, most companies considering open-source models still prefer to use their existing cloud platforms or enterprise AI gateways. Few manage models themselves. Adoption depends mainly on usage scale, internal technical capabilities, model performance and data residency requirements.

III. Token maxing or budgeting

1. As in the first two rounds, all companies in this sample set use a budgeting approach; none explicitly described token maxing as its current model

a) Some companies have implemented budget controls at the team or project level. Large European Logistics Company A sets a quarterly cap for its software development team. Projects that need to exceed the cap must explain the business value to IT and finance at a quarterly review. Large Southeast Asian Telecom Operator A allocates quotas by department, monitors them monthly against project progress and reviews them quarterly. Once the budget is exhausted, additional quota requires an internal assessment.

‘We are putting a maximum budget, a cap per quarter... If they want to go beyond that budget... they need to come to this dedicated stage-gating meeting.’

b) Large Indian HR Services Company A does not set a fixed hard cap for engineers, but checks whether token spending corresponds to specific output. Under last year’s flat subscription model, the company focused more on expanding usage. It has now started reviewing how much AI each engineering team uses and what it delivers, with budgets adjusted through annual planning and quarterly reviews.

‘There is no cap at the moment, but... let’s say you spend a thousand dollars, that needs to match to an output.’‘Till last year... spend as much... because it was a flat subscription model... But now we are being mindful on the engineering side.’

c) Controls remain preliminary at some companies. Large Australian Pharmaceutical Distributor A has only recently started using employee tools priced by token and needs several months of data before it can forecast spending reliably. Large Swiss Consumer Goods Manufacturer A still uses a shared subscription quota with no individual limits, while Mid-sized Middle Eastern Insurtech Company A relies mainly on the usage limits built into subscription products.

‘We want to give a couple of months to understand the patterns... then we will be able to predict for the next six months or 12 months.’

2. Only a few companies set fixed individual token limits; most use shared quotas, subscription tiers or overage approval to control usage

a) Large European Retailer A is one of the few companies with monthly token limits for software developers. An estimated ~5% of users reach the limit each month. Employees can submit an internal form to request more quota, and most reasonable requests are approved. The limit is intended to prevent sudden spending increases while the company continues to encourage AI adoption.

‘Only around 5% of users may hit the monthly limit. If they do, they can apply for more, and many requests are approved.’

b) Large Asian Gaming Company A introduced individual points quotas in September, with 7,000 points per day and 100,000 per month. Daily quotas can be raised temporarily within the weekly and monthly limits; exceeding either limit requires department-head approval. Premium models carry higher point multipliers per token, while locally deployed models use no points.

‘There is a daily limit of 7,000 points and a monthly limit of 100,000 points... If the weekly or monthly limit is exceeded, an expansion request requires approval.’

c) Mid-sized Middle Eastern Insurtech Company A controls individual usage mainly through subscription tiers. Service stops when a user exhausts the quota, and employees cannot top up accounts themselves. Developers and other users with a clear need can request a higher tier. Because spending is still mostly subscription-based, the company has not seen material billing surprises.

‘Till we have the subscription model, I don’t see spikes... If your subscription hits the limit, it just stops. Then you can top it up.’

d) Other companies prefer shared quotas or paid overages. Large Swiss Consumer Goods Manufacturer A still uses a shared quota; once it is exhausted, the company must pay more or accept a model downgrade, and it plans to set different permissions by user type. Commercial subscriptions at Large UK Legal Services Firm A allow overages priced close to API rates. The company therefore uses default models and access rights to prevent excessive use of premium models.

3. Automated model routing remains rare; most companies rely on default models, access rights or manual model selection to control costs

a) Large Southeast Asian Telecom Operator A is one of the few companies in this sample set using an automated router in production. Its customer-service system sends complex questions to Sonnet and basic questions to lower-cost models such as DeepSeek and Llama. Routing runs centrally in the backend, so customer-service staff do not select models themselves.

‘We have some intelligent routing... High-end and complicated queries... Sonnet... Basic questions... DeepSeek... Llama.’

b) Large UK Legal Services Firm A has not disclosed a router that switches models for each request, but has moved model selection for internal tools into the backend. The company tests model quality and cost, then presets the appropriate model for each tool. When premium models first launched, users applied them to many routine tasks and monthly spending reached ~1.8x the expected level. Default models and access controls brought this down to ~1.2x.

‘For every new tool... we get to set the model... There’s no model selection in our tools.’

c) Other companies still rely mainly on defaults and manual selection. Mid-sized UK Payment Infrastructure Organization A sets a default model for routine tasks and compares several models only for critical or complex work. Large Swiss Consumer Goods Manufacturer A currently relies on training to correct users who default to high-cost models and plans to limit model access by user type.

‘For very critical or sophisticated models, we do comparisons, but for the basics, we keep a default type of model.’

4. Caching has delivered measurable savings at a few companies, but most still rely mainly on budgets, access rights and model selection to control costs

a) Large European Logistics Company A has deployed an internal cache between user endpoints and models, reducing model consumption by ~20% - 25%. The longer-term target is ~40% without reducing accuracy or stability. Large Southeast Asian Telecom Operator A also uses caching and estimates a further ~15% - 20% optimization opportunity. Savings remain within the AI budget and are reallocated to other AI projects.

‘With [internal caching], we do see close to around 20%, 25% reduction in consumption... The goal... is up to 40%.’

b) Technical cost controls remain limited at other companies. Mid-sized Middle Eastern Insurtech Company A uses caching only in some conversational systems. Large Australian Pharmaceutical Distributor A is looking for a governance tool that can inventory internal agents and identify duplicate applications. Large Swiss Consumer Goods Manufacturer A and Mid-sized UK Payment Infrastructure Organization A do not yet use caching. The more common approaches are to restrict access to premium models, reclaim unused licenses and monitor spending by project and user.

IV. Optimization headroom

Across this sample set, AI cost controls are concentrated in model tiering, backend defaults, caching and agent governance. Large UK Legal Services Firm A cut monthly spending from ~1.8x budget to ~1.2x after adjusting default models and access rights. Internal caching at Large European Logistics Company A has reduced model consumption by ~20% - 25%, with a target of ~40%. Large Southeast Asian Telecom Operator A estimates a further ~15% - 20% reduction is available. As usage and production deployments continue to grow, most of the savings are being redeployed to new use cases and higher volumes. Large Asian Gaming Company A, however, expects paid model spending to decline by year-end through individual points limits and shifting traffic to local models.

1. Permissions and default settings for premium models are the most direct cost-control levers today

a) When premium models first launched at Large UK Legal Services Firm A, employees used them for many routine tasks and monthly spending reached ~1.8x budget. The company then introduced default models and access controls, reducing spending to ~1.2x budget. Models for internal tools are tested and preset in the backend, so employees generally do not need to select one.

‘We had about 1.8 times the cost we had expected in the month... We can now set people’s default models and the models they’re able to access... We’re at roughly 1.2 times what we would expect.’

b) Large Swiss Consumer Goods Manufacturer A still uses a company-wide shared subscription quota. In some months, general users defaulting to premium models consumed ~50% - 60% of the shared quota. The company has started reviewing high-usage users individually and is considering user tiers, with lower quotas and no access to top-tier models for general office users.

‘In some months, 50% to 60% of the budget has been burned like that... The C users are just the general users... They should never be able to use the top models.’

c) Large Southeast Asian Telecom Operator A uses backend routing in customer service. Complex questions go to Sonnet, while basic questions use lower-cost models such as DeepSeek and Llama. The system selects the model automatically based on the type of question.

‘For the very high-end and complicated queries, we are using Sonnet... For basic questions, we are going for the basic models. We are using DeepSeek as well... and at some places Llama.’

d) Large Asian Gaming Company A began assigning different point multipliers to models in August. GPT-5.6 Sol costs ~80x the base points and Opus 5 ~100x, while local models use no points. Premium models remain available, but users need department-head approval to exceed weekly or monthly quotas.

2. Caching has delivered measurable savings at a few companies

a) Large European Logistics Company A has deployed an internal cache between employee endpoints and models, reducing model consumption by ~20% - 25%. The company continues to improve the system and targets a reduction of ~40% without affecting accuracy or stability.

‘With [internal caching], we do see close to around 20%, 25% reduction in consumption... The goal... is up to 40%.’

b) Large Southeast Asian Telecom Operator A also uses caching and continues to test other optimization methods. Our channel checks suggest that current model spending could still be reduced by ~15% - 20%.

‘We already have implemented caching, but there are multiple other optimization techniques which we are applying. I believe that up to 15% to 20% we can improve.’

3. As agent counts rise, duplicate and idle applications are becoming part of cost management

a) Large Australian Pharmaceutical Distributor A has seen employees and departments build more agents, but does not yet have a complete view of what has been deployed. It is looking for a unified agent-governance tool to inventory existing agents, identify overlapping projects and retire applications that remain unused.

‘If two things are doing the same thing, then we’ll just need to remove it... We are looking for some kind of governance tool that can really understand what agents have been deployed across the company.’

4. Savings are usually reinvested in new use cases and higher volumes, although a few companies expect paid spending to decline outright

a) Large European Logistics Company A plans to reinvest savings from technical optimization in AI, expanding existing projects, developing new use cases and building reusable capabilities. Our channel checks suggest that if lower costs free up more budget, the number of use cases that can reach scaled deployment could rise by ~20% - 25%.

‘Every cost that we can reduce, we will redirect... into use cases that we can develop and, most importantly, into capabilities that can be reused across the organization.’

b) Large Southeast Asian Telecom Operator A also keeps savings within the AI budget for other projects. Large UK Legal Services Firm A expects annual usage commitments to reduce fixed OpenAI and Anthropic license fees to zero in 2027, but the saved amount will shift to usage-based spending and total AI software spend is still expected to rise.

‘The cost going down is not actually going to cause our AI spend to go down because that spend is just going to be allocated into usage rather than a license fee.’

c) Large Asian Gaming Company A is one of the few companies expecting a direct decline in paid model spending. As individual points limits take effect and light users and some high-volume users move to local models, paid API and third-party hosted-model spending could fall from more than $14.3M/month to ~$11.4M - $12.9M/month by year-end.

V. ROI

Some companies in this sample set have incorporated ROI directly into project approval and post-launch reviews. Large European Logistics Company A generally uses a four-year horizon and a ~3x return target. Large Australian Pharmaceutical Distributor A requires projects to be ROI-positive within two years. At Large Southeast Asian Telecom Operator A, a central value team works with finance to assess projects both before approval and after launch. Production use cases tied directly to sales, external service costs or avoided hiring are easier to measure. Employee tools are still assessed mainly through adoption, time saved and output quality, making their benefits harder to translate into financial returns. Large Asian Gaming Company A does not yet have a formal ROI framework; department heads currently approve overages based on the use case.

1. Some companies already use ROI as a key input to project approval and budget decisions

a) Large European Logistics Company A typically assesses project returns over four years and generally requires cumulative returns of ~3x the investment at approval. Projects expected to return ~2x are left to the relevant business-unit leadership. The business case includes developer, software-license, infrastructure and token costs. The business unit proposing the use case owns both the funding and the final business outcome.

‘We always look for a four-year TCO... The ROI has to be at least 300%... which means we expect our returns to be at least three times the investment.’‘There are use cases where it may not always be 300%. It may be, let’s say, 200% or so, and then it’s at the discretion of the leadership of that business unit if they would still like to invest.’

b) Large Australian Pharmaceutical Distributor A calculates investment and expected savings before starting any AI project. Projects are generally expected to pay back within one to two years and must be ROI-positive within two years. Use cases without a clear benefit do not proceed to implementation.

‘We calculate ROI for all of the use cases. I never start a use case without the ROI.’‘The use case will need to be ROI-positive within two years.’

c) Large Southeast Asian Telecom Operator A has a dedicated value team within its central AI organization. When a business unit proposes a use case, the value team defines metrics such as revenue, cost savings or cost avoidance and reviews the business case with finance. The same team then assesses actual results after launch. For a call-center project, the company first looks at response time, answer accuracy and queue time, then tracks NPS, escalation to human agents and labor costs.

‘We have a small sub-team, which is the value team... We develop a business case, and that business case goes to finance.’‘After execution and implementation, the same value team sits with those departments and tries to understand what benefit they have got.’

2. Returns from production use cases mainly come from revenue growth, avoided hiring and lower external service costs

a) Large UK Legal Services Firm A places more weight on AI’s contribution to revenue growth. In 2027, the company expects to spend ~$500K - $600K on AI tools and another ~$3M - $4M in developer costs, with a target of ~$15M in incremental revenue.

The company modeled the impact on a sales team of ~150 people. Cutting ~25 sales roles would save ~$5M at an estimated cost of ~$200K per employee. Instead, the company prefers to retain the team and use the time saved for more client outreach and meetings, which it expects to generate ~$10M of revenue. Tool improvements and better client screening could contribute another ~$5M.

‘We expect to spend about 600K on AI next year... roughly 3 to 4 million dev time. And then we expect to get roughly 15 million out of that.’‘Instead of basically 5 million in savings, we can actually generate, solely from the time, about 10 million in revenue. And then the additional 5 million in revenue is coming from actual advancements of the tool.’

b) A customer-service agent at Large Australian Pharmaceutical Distributor A already handles ~30% of customer inquiries automatically, reducing the need to hire more customer-service staff as volumes grow. A service-desk agent identifies and routes tickets, lowering fees paid to an external provider. In both cases, token and software costs can be compared directly with avoided hiring or lower external-service costs.

‘We automated, like, 30% of customer calls or customer inquiries by AI. That means you don’t need to hire more people to deal with the demand.’‘We used AI to automate the triaging of the tickets from the service desk, which helped us to reduce the reliance on the external vendor.’

c) At Large European Logistics Company A, ~10% of AI use cases directly affect revenue, ~40% - 50% reduce costs and the remaining ~40% mainly improve employee productivity. When a parcel has a barcode exception, manual correction takes ~5 minutes on average; computer vision reduces this to ~45 seconds. Given peak-season parcel volumes, the time and labor savings from this type of industrial use case can feed directly into project-return calculations.

‘If the barcode is wrong, then people spend on average five minutes per package to rectify that. And now with computer vision, we can bring that time to almost 45 seconds.’

d) Mid-sized UK Payment Infrastructure Organization A uses both financial and operating metrics for AI projects. Internal projects track time saved, processing capacity and new services. Payment-infrastructure projects also track fraud losses, cybersecurity incidents, system downtime and operating resilience. Each project uses metrics that match its business objective.

3. Returns on employee tools remain difficult to tie to financial outcomes, and time saved must be net of review and rework

a) At Large Indian HR Services Company A, employee-tool measurement does not focus only on content-generation speed; it also includes the time needed to review and revise the output. A presentation that previously took ~10 hours may be drafted quickly by AI, but if ~20% of the output still needs changes, the calculation must include ~2 hours of review and rework. The company therefore considers net time saved, rework and employee costs before increasing the related budget.

‘It’s very easy to say... you used to spend 10 hours on a slide and now you get the slides done in 10. But then you always have to fix a lot of that output... even, let’s say, 20% of that output needs to be fixed.’

b) Large European Logistics Company A assesses employee tools such as Copilot mainly through weekly time saved. If an employee saves ~2 hours per week, the function head must direct that time toward analysis, planning or other higher-value work and include the subsequent output in the assessment.

‘If somebody saves two hours every week, for example, we want to make sure what that person will do with that two hours that he has saved.’

c) The finance team at Large European Retailer A began building separate business cases for software engineering and knowledge work this year, shifting the focus from request volume to efficiency and output quality; the calculation still varies by use case. Large Swiss Consumer Goods Manufacturer A can directly compare pre- and post-AI costs in processes such as marketing agency work, translation and equipment maintenance. Other small office use cases are still assessed mainly through productivity and data quality.

d) Large Asian Gaming Company A does not yet have a formal AI ROI framework. The company first uses individual points limits to identify extremely heavy users. When weekly or monthly quotas are exceeded, the department head decides whether to approve more quota based on the applicant and intended use. A more formal assessment process is expected over time.

‘There is no formal ROI mechanism yet... For now, the department leader looks at who is applying and the reason for the request.’

4. Projects directly linked to business outcomes are more likely to receive follow-on funding

Large European Logistics Company A and Large Australian Pharmaceutical Distributor A have made return requirements part of project approval, while Large Southeast Asian Telecom Operator A also reviews actual results after launch. Revenue growth, lower external-service costs and avoided hiring are relatively easy to include in financial calculations. Employee subscriptions are still valued mainly through adoption, time saved and output quality.

VI. New use cases and budget growth

Among the faster-growing budgets in this sample set, spending is going to two main areas: high-frequency applications such as customer service and voice that are already in production and preparing to scale, and integration of AI into more business systems to build multi-step agents and workflows. New models rarely create entirely new production use cases by themselves, but they have materially improved output quality and completion times in a few areas such as talent matching and complex engineering tasks. Where employee subscription coverage is already high, incremental budget is shifting toward APIs and internal applications. One company that rapidly expanded paid model usage at the start of the year is now tightening spending through quota controls and local-model offload.

1. New models mainly improve existing use cases, with material gains in a few complex tasks

a) Large UK Legal Services Firm A uses AI to identify lawyers in its internal talent pool who match client requirements and to generate the rationale for each recommendation. Earlier models could complete the screening, but the resulting profiles were usually suitable only for internal use. About 80% of materials generated by the newer model can be sent to clients with only minor edits.

‘The first model we’ve used that can actually do that process well... About 80% of the time they’re ready to send off to the client, whereas the previous models just didn’t understand the client language and context.’

b) Large Indian HR Services Company A has also seen larger gains from new models on long-running, complex tasks with sufficient context. Some engineering work that previously took three to four days can be completed in five to six hours once the full background material is prepared. These tasks account for only ~10% - 20% of workloads, however, and routine work still uses lower-cost models.

‘Tasks that used to... take three days or four days... take five, six hours.’‘If I had to put a percentage number on it, 10% to 20%.’

c) Other companies have less need for new models. Mid-sized UK Payment Infrastructure Organization A mainly uses AI for document extraction, summarization, rule queries and information organization, and existing models already meet its needs; recent model upgrades have made only marginal improvements. Mid-sized Middle Eastern Insurtech Company A prioritizes response speed and output consistency in voice and conversational systems, usually choosing lightweight, proven models and switching production versions only when a new release delivers a material improvement. Large Asian Gaming Company A has not adopted Fable 5 because its internal model team does not currently see a need.

‘More and more, we’re seeing it as a marginal improvement, at least with the type of use cases we have now.’‘We don’t necessarily use the latest and the best... We generally work on a model which may be four months, six months old, which does the job for us, and stick to it until and unless there’s some change, really a 10x gain.’

2. The clearest new use cases connect AI to more systems and link multiple steps into a workflow

a) Large Southeast Asian Telecom Operator A is extending agents from development and testing into commercial functions. The customer-value management team plans to connect audience selection, offer design, marketing-content generation, campaign execution and performance analysis. The customer-service team is building an end-to-end process from receiving a complaint and routing it to a back-office team through to generating the customer response. Legal and procurement teams are also applying agents to contracts and tender documents.

‘There are multiple steps in the campaign process, starting from analyzing which customer we have to send the campaign to, what should be the offer, what content we have to send... Once the campaign has been executed, what is the response? This is where we are trying to apply all those agentic workflows.’

b) Large UK Legal Services Firm A has connected call records, CRM and talent-matching systems, but AI still handles individual steps in client identification, outreach, meeting preparation and talent screening. Employees must connect the steps manually. By end-2027, the company expects AI to execute more steps sequentially while retaining context. These workflows will cost more than current applications, and the company will require correspondingly higher returns.

‘Right now... it’s very disjointed. It’s not the same agent doing the different steps of the work.’‘I’d expect that to be much closer to solved by the end of next year... That will, of course, cost more. But we would also expect that to have far more ROI than where we are now.’

c) Some new use cases at Mid-sized UK Payment Infrastructure Organization A are aimed at banks and fintech companies. The organization is exploring agents that can provide anti-fraud capabilities to banks without directly sharing the underlying sensitive data. It is also building a digital rules repository that can query and explain payment rules, standards and regulatory requirements. Budget growth will depend on whether banks and other participating institutions adopt the services.

‘Now you can share AI agents that help banks control fraud. So you build an AI agent on a data set; you don’t share the data.’‘If the participants like what we show, definitely it’s going to increase, and if not, it should stay constant.’

d) Large Asian Gaming Company A uses AI in game production and version acceptance. Product teams have developed a multi-agent platform for cross-language translation and LQA testing to check translation quality in game builds. Animation teams also use agents for animation configuration and automated generation. Non-programming teams have built and deployed some tools for internal use.

‘We build cross-language translation agents and LQA testing agents - a multi-agent platform for version testing and acceptance, and for checking translation quality.’‘We now use agents to assist with animation configuration and automate parts of the animation-generation workflow.’

e) Large European Logistics Company A plans to train ~50 - 60 employees who understand specific business processes by the end of October. They will begin developing departmental agents in early 2027.

‘By the end of October, we should have trained these 50, 60 people... Once they go back to creating agents, that would be from January next year onwards.’

3. High-frequency use cases such as customer service and voice are starting to scale, and pricing models will affect budget growth

a) The next phase of growth at Mid-sized Middle Eastern Insurtech Company A will come mainly from voice services, with a target to increase current volumes by ~20 - 30x. The company is still improving the underlying capabilities and plans to prioritize annual subscriptions or fixed contracts to limit the billing volatility of usage-based APIs.

‘The goal is to go at least 20, 30x from the volume perspective... A lot of the spend and the focus would go into subscriptions. We would not work on API-based billing because API-based billing becomes a little unpredictable.’

b) Large Australian Pharmaceutical Distributor A has put customer-service, service-desk and cybersecurity agents into production, with the customer-service agent able to handle ~30% of inquiries automatically. The recruiter agent at Large Indian HR Services Company A handles initial candidate outreach, gathers preferences and performs first-round screening, increasing the number of candidates the company can reach and process at the same time.

AI usage, model selection and cost management provide the context for changes in spending. The table below summarizes the change in AI spending since the start of the year across 10 enterprise samples, along with their expectations for 2H26 and 2027.

VII. AI spending growth and outlook

This report covers 10 enterprise samples. The table below shows how their AI spending has changed and their budget outlook.