Frontier AI models are legitimately impressive. They draft quickly, summarize well, and can spot issues fast. Used properly, they increase individual productivity.

That has led to an understandable - but incomplete - conclusion: that specialized legal AI platforms are unnecessary, and that working directly in ChatGPT or Claude is enough.

It isn’t.

Yes, one part of the market deserves criticism. Many “legal AI” products are wrappers: a nicer interface on top of third-party models and generic retrieval. They may look polished, but they don’t fundamentally change how legal work is organized, verified, and preserved over time.

But rejecting wrappers doesn’t mean raw frontier AI is the answer.

Legal practice doesn’t run on output quality alone. It runs on reliability - the ability to preserve privilege boundaries, maintain continuity across matters that last months or years, keep client information isolated, and produce work product that remains traceable and defensible under scrutiny.

Those aren’t prompt-engineering problems. They’re architectural requirements.

Here’s what breaks when legal work relies on raw, general-purpose AI:

In litigation, a “good” summary that loses transcript line references or misses one key exhibit can be unusable, because you can’t cite it and you can’t trust that it reflects the record.
In deals, an AI that doesn’t carry forward the actual deal history - fallback positions, defined terms, prior concessions - creates inconsistency across drafts, and the lawyer becomes the one stitching the record back together.

More broadly, a few predictable risks show up:

Client information gets processed more broadly than intended because context isn’t tightly scoped. Context fractures across sessions, forcing lawyers to reconstruct the matter manually. Authorities appear without clear provenance, increasing verification burden and professional risk. And work product can look polished while still requiring major reconstruction to align with posture and record.

The result isn’t just inefficiency. It’s higher risk in the exact matters where firms can’t afford drift.

Lawyers also have a professional obligation to understand how client-sensitive material is handled. ChatGPT and Claude are general-purpose systems, optimized for broad use cases. Legal work demands tighter control over what is processed, where it travels, and how outputs remain grounded.

This is where the next phase of legal AI is emerging: infrastructure-centric systems.

Infrastructure determines what the model sees, how context is organized, how facts and authorities stay linked, how matters remain isolated, and how outputs remain traceable. Done correctly, it changes the risk profile of AI-assisted legal work: less review burden, stronger confidentiality controls, and more confidence that outputs reflect the actual matter record - not a reconstructed narrative.

That’s the approach we’ve taken at Irys: not a wrapper around a frontier model, and not a raw interface to one, but legal infrastructure built around matter continuity, scoped retrieval, and traceability - so advanced reasoning is usable in real legal workflows.

The market is moving beyond wrappers. But the future isn’t simply “use a frontier model directly.”

The future is legal intelligence built on infrastructure that makes reasoning safe, reliable, and aligned with professional obligations.

Wrappers improve access.
Frontier models accelerate cognition.
Infrastructure makes legal AI trustworthy.