The Problem With Bolt-On LLM Features
Most PropTech platforms started their AI journey the same way: a chatbot here, a document summarizer there, a lease-clause extractor wired into one workflow. Each feature works in isolation, but none of them talk to each other, share context, or compose into anything larger. The result is a patchwork of point solutions that is expensive to maintain and impossible to govern.
The next stage of maturity is an agentic orchestration layer: a single control plane that coordinates specialized AI agents, gives them safe access to your systems of record, and executes multi-step real estate workflows end to end. This is the architecture VSBD built for a production PropTech platform — and the work behind our PropTech 2026 Awards nomination.
What "Agentic" Actually Means Here
An agent is an LLM given a goal, a set of tools it can call, and the autonomy to decide which tools to use and in what order. "Agentic orchestration" is the layer above the individual agents that decomposes a goal, routes sub-tasks to the right specialist agent, manages shared state and memory, and enforces guardrails on everything that crosses a boundary.
The mental model is a control plane, not a single mega-prompt:
(Diagram in the original article.)
The Five Components Every Orchestration Layer Needs
1. The orchestrator. This is the brain that receives a goal, breaks it into steps, and decides which agent handles each step. Keep it deterministic where you can — explicit routing rules and state machines beat a fully autonomous planner for reliability and auditability. Reserve open-ended planning for the genuinely ambiguous cases.
2. Specialized agents. A document agent that classifies and extracts from leases is a different beast from a valuation agent that reasons over comparables. Narrow agents with focused tool sets and tight prompts outperform one generalist agent trying to do everything. They are also far easier to evaluate and improve in isolation.
3. Typed tool contracts. Every action an agent can take — query the CRM, draft an email, schedule a contractor — is exposed as a tool with a validated input and output schema. This is the single most important reliability decision you will make. Typed contracts turn "the model said something weird" into a caught validation error instead of a corrupted record.
4. Shared memory and state. Agents need short-term working memory for the current task and longer-term context (this tenant, this property, this portfolio). Externalize state into a store the orchestrator owns rather than stuffing everything into the prompt — it keeps token costs bounded and makes runs replayable.
5. The guardrail and approval layer. Between the agents and your systems of record sits a band of validation, policy enforcement, and human-in-the-loop gates. Irreversible or high-value actions — sending a payment, signing a document, emailing a tenant — pause for human confirmation. The agent proposes; a person approves.
Why Real Estate Is a Perfect Fit for Agentic Workflows
Real estate operations are full of multi-step, document-heavy, cross-system workflows that are too variable to hard-code but too repetitive to keep doing by hand. Consider tenant onboarding: parse the application, verify income documents, run a background check via a third-party API, generate the lease, route it for signature, and provision access. That is five systems and four document types — exactly the kind of workflow an orchestration layer collapses into a single supervised run.
Other high-value targets we see across European and US PropTech platforms:
Lease abstraction at portfolio scale — extract key terms, dates, and obligations from thousands of agreements and push structured data into the asset management system.
Maintenance triage and dispatch — classify an inbound request, check the SLA, find an available contractor, and schedule — with human approval on anything above a cost threshold.
Valuation and comps assembly — gather market data, retrieve comparable transactions, and draft a defensible valuation narrative for an analyst to review.
Tenant communication — draft grounded, policy-checked responses to inbound tenant queries, escalating anything sensitive to a human.
Build vs. Bolt-On: The Architecture Decision
The temptation is to keep adding LLM calls inside existing services. That works until you need two features to share context or coordinate — at which point the lack of a central orchestrator becomes a wall. Investing in the orchestration layer early means every new agent you add composes with the ones already there, rather than becoming another island.
The orchestration layer is also where your governance lives. Auditors and compliance teams do not want to hear "the AI decided." They want a trace: what triggered the run, what the agent proposed, what data it grounded on, who approved the action, and what changed. A control plane gives you that for free; scattered LLM calls never will.
Where to Start
You do not need to orchestrate everything on day one. Pick one workflow that is painful, repetitive, and cross-system — tenant onboarding and lease abstraction are common first wins — and build the orchestrator, one or two agents, typed tool contracts, and a human-approval gate around just that. Get the observability and evals in place from the first run. Once the pattern proves out, every additional agent is incremental.
VSBD designs and ships agentic orchestration layers for PropTech platforms across Europe and the USA — from the first supervised workflow to a full multi-agent control plane. If you are weighing whether to bolt on another LLM feature or build the layer properly, we can help you make that call.
Originally published on the VSBD blog.
