Enterprise AI in 2025 — Year in Review

A year ago, the 2024 year-in-review reflected on patterns starting to clarify. A year later, those patterns have settled. Production deployments are no longer remarkable; they are routine. The disciplines have professionalised. The capability ceiling has continued to rise but the production envelope has shifted with it. The picture in late 2025 is more mature than at the end of 2024, in some ways less dramatic, and clearer about what works.

This is a year-end reflection on enterprise AI in 2025 — what stabilised, what surprised, and what's heading into 2026.

What stabilised in 2025

Production patterns

The patterns that work in enterprise AI have settled. Bounded scope, retrieval-augmented architectures, human-in-the-loop on consequential decisions, layered defences, governance encoded in platforms, evaluation as a discipline. The recipe is consistent; the teams that follow it ship.

Model selection

Multi-model architectures are standard. Model routing — different models for different workloads — is the default in mature deployments. The single-provider, single-model architecture is now uncommon for serious workloads.

Cost discipline

The wild variance in AI bills that characterised 2023 and early 2024 has settled. Cost monitoring, attribution, budgets, circuit breakers are now part of standard operations. Surprises are less common; the surprises that happen are smaller.

Evaluation as a discipline

Curated eval sets, automated regression testing, model upgrade as planned migration — these have become standard practice. The teams that built the discipline in 2024 are reaping its benefits in 2025.

Governance frameworks

The EU AI Act has entered into force; sector regulators have published more specific expectations; internal frameworks have codified. Governance is now part of the operating model, not an external constraint.

Open-weight credibility

Open-weight models are credible enterprise options for most workloads. The capability gap with closed frontier models has narrowed substantially. The choice is workload-specific, not ideological.

AI platforms

The platform pattern — shared infrastructure that line-of-business teams build on — has matured. The platforms that compound value across organisations have recognisable shape; the ones that didn't have been visibly disinvested.

What surprised us

Reasoning models becoming useful

The reasoning model category — o1, o3, equivalents — has shifted what's possible on multi-step problems. The cost is high; the use cases are narrower than the marketing; within those use cases the capability is meaningful.

MCP traction

Model Context Protocol gained more adoption than we expected at the start of 2025. The standardisation is reducing custom integration work; the ecosystem of MCP servers is growing.

Voice AI crossing the threshold

Real-time voice with current models has produced production deployments that previous generations couldn't. The category is now serious.

Computer use's slower production adoption

Computer-use AI got significant attention in late 2024 and early 2025; production deployment has been slower than the attention suggested. The reliability and audit gaps are real; the use cases are narrower than initially expected.

Self-hosting becoming mainstream

A year ago, self-hosting was niche. In 2025 it has moved into the mainstream for workloads with data residency, high volume, or latency requirements.

Multimodal becoming default

Document workflows, field operations, mixed-content support — multimodal has gone from a feature to a default expectation in many enterprise contexts.

What didn't materialise

A few predictions and expectations that didn't pan out:

Autonomous agents at scale

The autonomous-agent vision continued to be ahead of the production reality. Bounded supervisor-worker patterns are what ships; autonomous decision-making on consequential actions remains rare.

AI replacing significant headcount

Despite headlines, the actual headcount impact has been smaller than projected. Roles evolve; productivity rises; net headcount changes have been modest.

Specific frontier capabilities flowing into production faster

The frontier kept advancing; the rate of adoption into production has been slower. Enterprises adopt deliberately; the gap between frontier and production stays roughly constant.

Specific industries transforming wholesale

Categories like healthcare and education have absorbed AI more cautiously than the marketing projected. The regulated industries' caution is appropriate.

Custom foundation models

The vision of enterprises training their own foundation models hasn't materialised at scale. Fine-tuning of open models is the realistic capability; foundation training remains too expensive for most.

What got harder

A few areas where the work has intensified:

Governance complexity

Frameworks have codified; the work of complying with them has grown. Audit, documentation, validation are real and increasing efforts.

Vendor management

The vendor landscape is complex. Open ecosystems plus closed providers plus specialised vendors plus AI features in conventional software. Managing this requires capability that wasn't fully developed.

Cost management at scale

As AI adoption expands, the total spend grows. Managing it requires FinOps capability specifically for AI, which is still being built.

Talent

The competition for AI engineering and AI governance talent has intensified. Enterprises are paying more; capability building is still a challenge.

Cross-jurisdictional compliance

Different regulatory frameworks across the regions enterprises operate in. The compliance work is multi-track.

What we keep saying

Some recommendations that bear repeating:

Discipline beats capability

The teams that ship reliably are the disciplined ones, regardless of model choice. The teams that chase capability without discipline keep producing impressive demos and few shipped systems.

Foundations compound

The data layer, the evaluation infrastructure, the governance framework, the operational discipline — these compound. The teams that invest in foundations capture the value as capability evolves.

Match technique to workload

The right model, the right architecture, the right pattern depend on the workload. There is no universal best answer. The discipline of matching is the skill.

Human-in-the-loop is the production posture

The autonomy aspirations will keep going; the production reality stays supervised. The boundary between AI and human decision is the design choice.

Engage governance early

The governance work has its own timeline; engaging it late produces blocked deployments and reactive scrambling.

Going into 2026

What we expect:

Continued specialisation

Specialist models for narrow tasks, specialist agents for specific workflows, specialist tooling for specific domains. The general capability is good enough; the specialisation is where competitive advantage lives.

Deeper integration with conventional enterprise systems

AI becoming a feature of every major enterprise platform, not as a separate stack. The integration with existing systems and workflows will deepen.

Continued governance specification

Regulators will continue to publish more specific expectations. The institutions building governance capability will navigate the increasing specificity more easily.

Cost-curve continued evolution

We expect continued price competition, with the magnitude harder to predict than the direction. Build for cost discipline regardless.

Workforce continued evolution

Roles evolve; productivity rises; the skill profile of enterprise teams shifts. The change management of this is ongoing.

Selective transformation, not wholesale

The "AI transforms everything" framing continues to overstate. Selective transformation of specific workflows is the realistic pattern. The aggregate effect compounds over years.

What we recommend for 2026

For enterprise teams entering 2026:

1. Invest in operating and improving existing deployments before chasing new ones.
2. Maintain the disciplines — evaluation, governance, observability, cost.
3. Adopt new capabilities deliberately. The frontier evolves; production should follow with intent.
4. Plan governance and compliance work explicitly. The requirements are increasing.
5. Build workforce capability sustainably. The talent picture is competitive; internal capability matters.
6. Engage business sponsors as partners. The work succeeds when business and technology are aligned.
7. Take a multi-year view. AI value compounds; quarterly thinking misses the picture.

Enterprise AI in 2025 has become a routine enterprise capability. The drama of the early years has receded; the substance has settled. The teams that respect what 2025 taught us go into 2026 with the foundations to capture the value as capability evolves. The teams that keep chasing the next exciting capability without the discipline produce the same patterns of impressive activity and modest delivery. The choice, as always, is in the discipline.