Production is the hard part. Why CRE and investment managers aren't seeing AI ROI
Most AI programs stall between the demo and the P&L. The fix usually starts in the data layer, well before anyone picks a model.
Most prospective clients we talk to in the Built Environment and Investment Management verticals have a version of this story.
The demo pulled every lease term correctly. Six months later, abstraction still runs on spread sheets through the same analysts. The pilot never connected to the system of record, and no one would sign off on output they couldn't trace back to the source document.
The pilot proves the model can do the work. Then production asks harder questions about where the data lives, who can see it, and who answers for the result. The program loses momentum while those questions get worked out one meeting at a time.
The model is rarely what's missing
Claude, powered by Anthropic, can read a 90-page lease, a data room, or a borrower's financial package and pull out what matters., Claude Sonnet 5.5 released last month does this even better than previous models. That capability is widely available now. What separates firms that produce ROI from AI from those still running pilots is the layer underneath. That means clean ingestion from the systems of record, a data model and ontology that reflects how the business actually works, and clear rules about what the AI can touch.
When that foundation is in place, the application layer moves quickly. When it isn't, every new use case rebuilds the same plumbing, and the business case never catches up to the cost.
Start from the outcome and work backwards
The most useful planning exercise we run with clients is simple. Name the decision or report you want to change, then trace it back to the data it depends on. That tells you what has to be built, what it will cost, and which later use cases will ride on the same work.
A few examples from the work we've scoped:
In commercial real estate, lease abstraction is usually the first stop. Terms come out of the document in under an hour instead of several days, and every field links back to the clause it came from. The same extraction foundation then feeds lease-rollover scoring and ESG reporting, both of which are just as document-heavy.
For fund managers, the pressure point is NAV and the reconciliation behind it. A governed, auditable calculation replaces spreadsheet rollups across fund structures. Book-to-tax reconciliation that took days runs in minutes against one audited source. Carry waterfalls and LP reporting draw from that same ledger.
Across both examples, the payoff executives feel most directly is the ability to talk to their data. That dream only becomes a plausible reality when the data under it is governed.
Governance your board will ask about
Executives evaluating an AI partner usually ask what the system can see, what it's allowed to do, and who is accountable when it acts. We answer those questions in the configuration, before any work starts.
Work runs in your environment. Permissions live in a managed configuration file, so access doesn't depend on any one engineer's judgment. An agent that needs to read a database gets read-only access to that database and nothing else. Every agent action is logged, so you have a record of what the system touched and when. Where your agreement with Anthropic supports it, we configure zero data retention so nothing persists after the session.
The same rules apply to the business agents we build for contract review, finance reporting, hiring support, and customer inquiries. Scope is narrow and ownership is explicit. Any decision involving money, legal exposure, or a client relationship goes to a person.
How we deliver
Proxet runs AI-assisted engagements through a framework we call the Intent-Driven Lifecycle. Claude handles the structured work at each stage. It turns meeting notes into user stories, generates tests from requirements, and keeps documentation current as code changes. A separate Claude instance reviews code without the original context, which catches issues a single pass tends to miss. Engineers review every change before it ships.
For you, that means the fifth project runs the same way the first one did. Repeatability is what turns a list of use cases into a roadmap you can plan and budget.
The team behind it
127 Proxet engineers hold Anthropic's Claude credential. It's a proctored exam built around production judgment: agent loops that fail under load, tool integrations scoped too broadly, and systems that can't check their own output. We required hands-on production experience with Claude before anyone sat for it.
Proxet is a Preferred Partner in the Claude Partner Network Services Track. The program holds partners to an ongoing standard of certified people, live customer deployments, and published customer results. That keeps the team you meet at the start of an engagement current when the next one begins.
Where AI helps less
AI doesn't speed up every kind of work equally. A 2025 randomized study by METR found that experienced open-source developers took 19% longer on tasks in codebases they knew well when they used AI tools. The gains are real in structured work like test generation, documentation, scaffolding, and multi-pass review. They're thinner in open-ended problem-solving and judgment-heavy decisions.
Knowing which kind of task you're looking at is most of the skill. That judgment, more than any credential, is what a year of production work has taught our team.