Six shifts defining enterprise AI right now
A field note on what has actually changed recently — and which of it matters if your goal is getting systems into production rather than following the discourse.
- Agents are moving into production, but only where scope is narrow and autonomy is gated.
- Evaluation has become infrastructure — the durable asset most teams still do not have.
- Governance shifted from principles to enforceable regulation with real deadlines.
- Document and multimodal AI quietly delivers more enterprise value than anything on a leaderboard.
Reading the field honestly
Enterprise AI discourse runs about a year ahead of enterprise AI reality. The useful exercise is to separate what changed in capability from what changed in practice — what teams are actually able to put and keep in production. These are the six shifts we see mattering most from inside the work.
1. Agents crossed into production, narrowly
Two years ago agentic systems were demos. Now they are running real workloads — but the ones that survive contact with production share a profile: tightly scoped, given well-designed tools, heavily instrumented, and permitted to act autonomously only on reversible, low-blast-radius steps. General-purpose autonomous agents remain unreliable in ways that matter, and the pilots that failed mostly failed by attempting breadth.
The practical read: agents are ready for jobs you can describe precisely and evaluate rigorously. They are not ready to be handed an objective and left alone.
2. Evaluation became infrastructure
The most important shift is also the least glamorous. Teams that ship reliably have stopped treating evaluation as a testing phase and started treating it as a permanent system: a versioned suite built from real cases, running in CI, gating deploys, and extended every time production surfaces a new failure.
This is also, increasingly, where the durable asset is. Models change; your evaluation suite is the thing that lets you adopt a new one in an afternoon with evidence instead of a quarter with anxiety. We would rather inherit a team's eval suite than their prompts.
If you take one thing from this piece: build the evaluation suite before you build the system. Every other decision — model choice, architecture, whether to grant autonomy — becomes evidence-based once it exists, and remains a matter of opinion until it does.
3. Open weights reached good-enough parity
For a large share of enterprise tasks, open-weight models are now close enough to frontier systems that the deciding factors are deployment location, cost, latency, and control rather than raw capability. That changes procurement: the question moved from "which provider is best" to "which model, in which place, for which workload."
It also made portability a design requirement. Teams building on a gateway with a portable evaluation suite can move workloads as the frontier shifts. Teams welded to a single vendor's API cannot, and pay for that rigidity every time the market moves.
4. Governance stopped being a principles document
Responsible-AI statements have been replaced by enforceable obligations with dates attached. The EU AI Act phases duties in over several years, with the heavier requirements landing on high-risk systems. Management-system standards such as ISO/IEC 42001 give organizations something certifiable. Voluntary frameworks like the NIST AI Risk Management Framework are increasingly written into contracts. Sector regulators — financial, medical, employment — have made clear that existing rules apply to AI-assisted decisions without needing new statutes.
The practical effect is that documentation, lineage, human accountability, and evaluation records are becoming table stakes rather than differentiators. The teams that built these in are unbothered; the teams retrofitting them are discovering it is expensive and sometimes architecturally impossible.
5. Document and multimodal AI does the quiet heavy lifting
Leaderboards reward reasoning benchmarks. Enterprise value is disproportionately created by reading documents accurately — invoices, claims, contracts, clinical notes, benefit applications, inspection photographs. This work is unglamorous, extremely common, and has improved substantially with vision-language models that read a page as a page rather than as a stream of OCR fragments.
If you are looking for near-term return, the highest-probability place to find it is usually a document workflow that currently consumes a large number of human hours in mechanical assembly.
6. Cost moved from footnote to design constraint
As systems reached real volume, inference cost became an architectural input rather than a finance conversation. Routing between model sizes, caching aggressively, specializing small models for hot paths, and budgeting cost per interaction are now standard practice rather than optimizations for later. Related, and underrated: context engineering — deciding precisely what enters the model's window at each step — has displaced prompt phrasing as the craft that most affects quality, latency, and cost simultaneously.
What has not changed
Worth stating plainly, because it survives every hype cycle. Data foundations remain the bottleneck; most stalled AI programs are stalled on data quality, lineage, and access rather than on models. Change management remains the difference between a working system and a used system. Narrow, well-instrumented deployments still outperform broad ambitious ones. And human accountability for consequential decisions is holding — technically, legally, and, in our view, correctly.
The organizations doing well are not the ones tracking every release. They are the ones with an evaluation suite, a governed data foundation, a small number of workflows in production, and the operational discipline to keep them healthy. That has been the answer for a while, and each of these six shifts has made it more true rather than less.
More perspectives.
How AI is redrawing market structure
Model capability is commoditizing faster than almost anyone predicted. That doesn't mean the value disappears — it means it moves. A map of where it is going, and what that implies if you are buying rather than building.
Read the perspective →Agentic AI grows up: from impressive demo to system that holds
Most agent pilots stall in the gap between a demo that works once and a system that works ten thousand times. The difference is almost never the model.
Read the perspective →The quiet shift to small, specialized models
The largest model is rarely the right model. Why serious work is moving down the size curve — and what enterprises gain besides a smaller bill.
Read the perspective →Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.