Context
This solo episode of The AI Daily Brief synthesizes three recent views of enterprise AI behavior: Databricks' AI Index, Arcana Research's enterprise AI study, and Ramp's AI Index. The useful question is not simply whether AI adoption is growing. It is how companies are getting more output without letting spend rise at the same pace, and what that says about product strategy, vendor selection, and rollout priorities. For product managers, the episode matters because it shows enterprise AI shifting from early experimentation to a more disciplined phase built around model switching, cost control, and narrower production use cases. These notes focus on that enterprise AI segment of the episode.
The Big Idea
Enterprise AI is becoming a portfolio and procurement problem, not just a model capability race. The companies furthest ahead are building the ability to switch providers, compare models on task outcomes, and use price competition to get much more AI for almost the same spend.
That changes what good product judgment looks like. A PM now needs to care not only about whether a model can do a task, but whether the team can route work to the cheapest model that still succeeds, measure the result, and avoid getting trapped by one vendor.
Key Insights
Usage grew faster than spend
- Ramp found that October token usage rose 36% while AI spend rose only 2%.
- Across OpenAI and Anthropic, the average price per million tokens in October fell 21%.
- The product implication is simple: use cases that looked marginal on ROI a quarter ago may now work if the team can take advantage of cheaper models or more efficient routing.
Switching providers became normal
- Databricks said that in July, 64% of companies using Unity Gateway relied on a single model provider.
- By early November, 72% were using two or more providers.
- Arcana found that a full third of OpenAI customers were open to switching, often considering Anthropic. Arcana also found that 43% of GitHub Copilot customers were open to switching, generally naming Claude Code.
- PM takeaway: model portability is turning into a buyer expectation, not an edge case.
Outcome cost beat token cost
Databricks argues that raw token price is the wrong number to optimize. The better metric is cost per successful task: the total cost to complete a job at acceptable quality. A cheaper model can still be more expensive if it fails more often, needs retries, or creates review work for humans. This is the right frame for PMs evaluating AI for search, support, code generation, or workflow automation.
AI budgets lack visibility
- Arcana found that two-thirds of surveyed Global 2000 leaders did not have visibility into model and API spend.
- At the same time, those leaders described AI as an eight-figure budget line.
- Arcana also found that 91% expected to spend more on AI in 2026, with most expecting growth of 20 to 40%.
- Large budgets without clear spend visibility usually lead to blunt cost cutting later. PMs who instrument spend early are in a much better position to defend high-ROI use cases.
Spend is extremely concentrated
- Ramp found that the top 1% of firms spend about 625x what the median firm spends per employee per month.
- The median Ramp firm spends $16.32 per employee per month. The top firms spend over $10,000 per employee per month.
- Databricks saw a similar shape. The median user spends $2.49 per week, while the 99th percentile user spends almost $700 per week, a 281x gap.
- This matters for PMs because average adoption numbers can hide a market where a small power-user segment drives most revenue, most feature requests, and most urgency.
Coding leads, support rises
- Ramp found that 72% of AI spend comes from engineering teams.
- Arcana found that AI coding tools are the single biggest pilot category among surveyed Global 2000 companies, with a full third piloting them.
- At the same time, Ramp saw customer support grow from 0.8% of AI spend in June to 13.3% in November.
- Arcana found that 38% of surveyed Global 2000 companies already use AI in customer service.
- The pattern is useful. Coding remains the main entry point, but support is emerging because the work is repetitive, measurable, and easier to evaluate than open-ended agent workflows.
Agents lag real deployment
Arcana found that autonomous agents ranked fourth on its heat index of pilots and plans, but dead last on the adoption index of what companies had actually deployed. That gap matters more than the headline enthusiasm. For PMs, it is a warning not to overread executive curiosity or conference buzz as proof that buyers are ready for production agent systems.
Mental Models & Frameworks
Cost per successful task
- What it is: measure AI by the total cost of completing a job at the required quality level, not by token price alone.
- How it works: include model cost, retries, latency penalties, review effort, and failure recovery.
- When to use it: any time a team is comparing models for a production workflow such as support replies, code generation, classification, or summarization.
- Why it matters: a low-cost model that fails often can be more expensive than a premium model that succeeds on the first attempt.
Model optionality as leverage
Model optionality means building enough abstraction into a workflow that a team can change providers without rewriting the whole product. The episode shows why that matters. Once switching becomes common, technical flexibility turns into commercial leverage. A team that can switch has more negotiating power, more resilience when quality shifts, and more freedom to match each task to the best model.
Heat versus deployment
This is a simple check for separating interest from reality.
- Heat: what buyers say they are excited about, piloting, or planning.
- Deployment: what buyers have actually put into production.
- Use it for: roadmap decisions, sales narratives, and market sizing.
Arcana's agent data is a good example. Autonomous agents score high on interest and low on production use, so a PM should treat them as a future category with current friction, not as a mature demand signal.
Power-law enterprise demand
Enterprise AI demand is not evenly distributed. A small slice of firms spend vastly more than the median, which means market averages can hide the real shape of opportunity. Use this model when setting pricing, enterprise packaging, or roadmap priorities. A product may need one plan and one feature set for the broad middle, and a very different offer for heavy AI users who create most of the revenue.
Decision Principles
Principle: Preserve vendor flexibility
- When: choosing a foundation model, designing an AI architecture, or signing a long vendor agreement.
- Why: the episode's data shows that multi-provider usage is becoming normal, and switching creates both cost leverage and risk protection.
Principle: Instrument spend early
- When: an AI workflow is moving from pilot to budgeted product investment.
- Why: Arcana's finding that two-thirds of leaders lack model and API spend visibility shows how easily cost can scale before understanding does.
Principle: Start with measurable work
- When: deciding which AI workflows deserve product investment first.
- Why: coding and support lead adoption because success is easier to define, failure is easier to catch, and ROI is easier to prove than in broad autonomous agent systems.
Trade-offs & Nuance
Savings versus integration complexity
Multi-model routing can reduce cost and increase negotiating leverage, but it also adds real product and platform work: evaluations, routing logic, fallback rules, governance, and monitoring. The savings are real only if the team can manage that complexity without degrading reliability.
Ambition versus operational certainty
Autonomous agents offer larger upside on paper, but narrower workflows such as coding assistance and customer support move first because the tasks are bounded and the output is easier to judge. The trade-off is between a bigger future promise and a smaller present workflow that can actually survive procurement and production review.
Signal versus representativeness
The studies point in a similar direction, which makes the pattern worth taking seriously. Still, each dataset has limits. Ramp reflects a narrow group of tech-forward companies, Databricks reflects customers routing through its gateway, and Arcana's Global 2000 sample is small. A PM should treat these numbers as strong directional evidence, not universal market truth.
Common Mistakes
Mistake: Optimizing token price alone
A lower token price can look like a win while total workflow cost gets worse. If the cheaper model needs retries, produces weaker output, or forces more human review, the real cost per resolved support case or completed coding task can rise.
Mistake: Confusing interest with rollout
Teams often overreact to what buyers say they are excited about. Arcana's data shows why that is dangerous. Autonomous agents generated strong interest but ranked last in deployed adoption, so roadmap weight should come from production behavior, not pilot buzz.
Practical Application
Audit model spend visibility
- Break down AI spend by team, workflow, and model provider.
- Include retries, failed runs, and human review time where possible.
- Review the data monthly, not quarterly.
- This closes the exact visibility gap Arcana found across much larger enterprise budgets.
Add a second provider
For one high-volume workflow, build a fallback or side-by-side test with a second model provider. Compare quality, latency, and total cost to completion. Even if the primary provider stays the same, the ability to switch creates leverage in both product design and vendor negotiations.
Define task success clearly
Before scaling an AI workflow, write down what success means in operational terms. Examples include a support reply accepted without edit, a code suggestion merged, or a summary approved by a reviewer. Then measure cost per successful task instead of cost per prompt or cost per token.
Pilot support before agents
If the roadmap jumps straight from coding assistants to broad autonomous agents, insert a narrower support or operations use case first. Customer service work has clearer metrics, repeated patterns, and more immediate evaluation loops, which makes it a better proving ground for production AI.
Renegotiate with real options
Take usage data and switching capability into vendor conversations. The episode's central point is that competition is intense and providers are under price pressure. A team that can credibly move workload elsewhere is more likely to get better pricing or better terms.
Questions to Consider
- Which AI workflows in your product have a clear definition of a successful task, and which ones still rely on vague judgments such as users liked it or the demo looked good?
- If a foundation model price changed or performance dropped next month, could your team switch providers for one important workflow without a full rewrite?
- Where is autonomous agent excitement on your roadmap outrunning what customers have actually put into production?
- Are you planning around average enterprise AI adoption, or do you understand whether a small group of heavy AI users is driving most of the opportunity in your market?
Bottom Line
The enterprise AI market is getting smarter about buying, not just using. The strongest signal in this episode is that companies are learning to switch models, measure outcome-level efficiency, and focus deployment on bounded workflows like coding and support, while autonomous agents remain more promise than production.
Resources Mentioned
| Resource | Type | Why it was mentioned |
|---|---|---|
| Databricks AI Index | Report | Used to show the rise of multi-provider model usage, the importance of model routing, and the shift toward cost per successful task. |
| Arcana Research state of enterprise AI study | Research study | Used to show Global 2000 budget expectations, low spend visibility, vendor switching openness, and the gap between agent interest and deployment. |
| Ramp AI Index for October | Index | Used for the October data point that token usage rose 36% while spend rose only 2%, alongside falling average token prices. |
| Ramp AI Index for November | Index | Used to show team-level spend concentration, vendor share, and the growth of AI usage in customer support. |
