The PSV Production Standard: How AI Work Ships in Commercial Real Estate Operations

PSV publishes workflow guides for the jobs commercial real estate teams actually run with AI: appraisal review, underwriting, due diligence, leasing, asset management, property management, brokerage work, and tool selection. Each guide covers its own workflow in its own terms, and each closes by applying one shared operating standard to that workflow. This article is the canonical statement of the standard itself, in one place, so a team can read the controls once and then judge any individual workflow against them.

The standard exists because the failure pattern in this industry is consistent. Teams rarely get hurt by a model that writes badly. They get hurt by fluent work whose sources nobody checked, whose assumptions nobody surfaced, and whose approval nobody owned. Every control below targets that pattern, and none of them depends on which model or vendor a team uses.

A usable production AI workflow starts with a named job. Define the event that enters the queue, the source material that may be used, the finished output, and the person who owns the next decision. A request such as “help with underwriting” is too loose to test or govern. A request such as “prepare a cited first pass for this packet, flag every missing or conflicting item, and route the exceptions to the named reviewer” is concrete enough to inspect. That difference is how a team prevents a useful tool from becoming an unofficial process that nobody can explain, improve, or safely hand to the next person.

Keep the input boundary as disciplined as the output. List the approved files, systems, and public records; preserve each source’s date and version; and make the workflow stop when a record is missing, stale, or outside the permission model. Do not reward a system for filling a gap with a plausible answer. The stronger behavior is to surface the missing evidence in an exception register, attach the original request or document, and assign an owner and due date. That turns uncertainty into work the team can close, rather than letting it move invisibly into a memo, a model, or an external conversation.

The output should separate four things that are often blended together: the sourced fact, the deterministic calculation, the working assumption, and the recommendation. A reader needs to know which sentence can be traced to a document, which number can be recalculated, which premise is awaiting confirmation, and which judgment a person is making. That separation is a practical control, not a formatting preference. It gives a second reviewer a fast way to test what matters, lets a committee challenge the correct layer of the work, and keeps a confident narrative from disguising an unsupported input as a settled fact.

Design the exception path before allowing the workflow to run at volume. Define what happens when sources conflict, the request falls outside policy, the model cannot identify a required field, or the proposed action has legal, financial, tenant, lender, or reputational consequences. Low-risk preparation can move quickly; consequential decisions should stop for accountable review. The handoff should identify the specific issue, the evidence considered, the unanswered question, and the person with authority to resolve it. A visible stop condition is what makes automation trustworthy in practice, because it gives people a way to intervene before an error becomes a commitment.

The approval record belongs with the work product. Capture who reviewed the sources, who accepted or changed the draft, what assumption was approved, and what was ultimately sent, entered, priced, or decided. When a team later asks why a conclusion changed, it should be able to open the record rather than reconstructing decisions from chat history and email. This is also how the workflow survives staff changes: the next reviewer inherits the evidence, the open questions, and the previous decision logic instead of only a finished answer with no path back to its source.

Measure the system on decision quality and review burden, not on how impressive its first draft appears. Track cycle time, the rate of material corrections, unresolved exceptions, reviewer effort, and whether the finished package made the next decision faster and more defensible. Review a small set of real examples on a regular cadence, update the instructions or template when the pattern is clear, and widen the scope only after the controls hold. That is the institutional standard PSV applies: AI handles structured preparation and evidence management; accountable people keep the judgment, the approval, and the responsibility.

Set a review cadence before the workflow becomes routine. A weekly operating check can clear exceptions and resolve ownership; a monthly quality review can sample completed work, identify repeated corrections, and decide whether the instruction, source set, or approval gate needs to change. Keep the record of those changes with the workflow itself. That gives the team a controlled improvement loop instead of a collection of private prompts, and it ensures that a system that was reliable at launch remains reviewable as the people, files, and tools around it change.