The Token Stops Here
Enterprise AI shifts from model magic to margin math
My favorite genre of conversation is with CTOs of very large companies who make decisions that move thousands of employees, billions of dollars, and occasionally entire vendor categories - and yet have absolutely no social presence.
They are not posting hot takes or chasing the hypegeist. They are usually three abstraction layers beneath the public conversation, thinking less about which model is best than how to keep a 40-year-old enterprise running while introducing entirely new ways of working.
A few takeaways from conversations this week with industry leaders, especially relevant for anyone selling AI into Fortune 100-scale enterprises:
1. “Build” is back - not always as a strategy, but as a threat.
In the perennial build-vs.-buy debate, build is becoming a serious negotiating tool. For the first time in a while, large enterprises can credibly tell vendors: if you are not passing through AI-driven productivity gains, we may build the replacement ourselves.
That does not mean every bank, insurer, or healthcare company wants to become a software company. It means they want pricing leverage. AI has made the internal-build option credible enough to change the tone of large contract renewals. AI gives the buyer a credible BATNA.
2. Consumption based pricing is introducing unpredictability in the expense base
Enterprise buyers understand seats. They understand predictable annual contracts. They are much less enthusiastic about signing contracts whose economics ultimately depend on token consumption that neither side can accurately forecast.
If the vendor absorbs inference costs, their margins get squeezed. If the customer absorbs them, budgets become harder to forecast. If everyone pretends it is fine, procurement eventually shows up in a bad mood. The next SaaS pricing battle may be about who owns inference risk.
3. The best enterprise AI products may be the ones that know when not to use AI.
The winning pattern in regulated, complex, cost-sensitive environments may not be “make everything agentic.” It may be 90% deterministic workflow automation + 10% AI judgment.
Let software do the repeatable things repeatably. Use AI for ambiguity, classification, summarization, exception handling, and judgment calls. That architecture is cheaper, more reliable, easier to explain, and much easier to govern.
4. Agents are not yet an architecture. They are a capability looking for a control plane.
Today’s agents feel a bit like stored procedures before API governance: powerful, fragmented, hard to observe, and terrifying once they start multiplying inside a large organization.
API governance took years to mature into identity, permissions, observability, policy, lifecycle management, versioning, and rollback. Agent governance will follow a similar arc.
Which is one reason the incumbents have a better story than people sometimes acknowledge. Microsoft, ServiceNow, Salesforce and Atlassian already own many of the systems that answer enterprise questions like “Who are you?” and “Are you allowed to do that?”
“Agentic chaos” favors platforms with enterprise context and distribution rather than only net-new AI startups.
5. Slack-centric AI overfits to tech-company workflows.
A lot of AI products are designed around how technology companies work: Slack, docs, tickets, code, meetings, and fast-moving teams with relatively legible workflows.
But banks, insurers, healthcare companies, industrials, and governments do not run their core operating fabric in Slack.
Their context lives in email, ServiceNow, Microsoft 365, Teams, SharePoint, internal portals, ticketing systems, mainframes, bespoke databases, policy docs, approval chains, and institutional knowledge accumulated over decades.
The most valuable employee in many enterprises is still the person who knows which undocumented process breaks every quarter.
The enterprise AI market is over-indexed on Silicon Valley collaboration patterns. The real TAM is messier, older, and less chat-native.
The overarching takeaway is that sophisticated enterprise buyers are not anti-AI. They are anti-vague-ROI, anti-unbounded-token-cost, anti-agent-sprawl, and anti-generic-workflow. They want AI embedded into systems that are governed, observable, economically rational, and connected to proprietary enterprise context. Much harder business to build, also a much bigger one.