Most firms approach AI integration the wrong way. They treat it as a technology selection problem. The team evaluates vendors, runs pilots, and selects a tool. The tool gets deployed, sees inconsistent adoption, and produces marginal value. The conclusion is that AI is overhyped or that the firm is not ready. The actual conclusion should be that the firm started in the wrong place.
AI integration is not a technology selection problem. It is a workflow design problem. The technology determines what is possible. The workflow determines what is real. A team that selects the right tool and deploys it into the wrong workflow will produce the same result as a team that selects the wrong tool and deploys it into a strong workflow. The tool is necessary. The workflow is what makes the tool useful.
The starting point matters more than the destination. Firms that start with the workflow and select the tool to fit produce results that compound. Firms that start with the tool and ask the workflow to accommodate produce pilots that never become production.
The Wrong Starting Points
Three common starting points consistently produce poor outcomes.
The flagship demo. A senior partner sees an AI tool demonstrated. The demo is impressive. The firm acquires the tool and asks the team to find uses for it. The team identifies adjacent workflows, attempts to adapt them to the tool, and produces inconsistent value. The tool sits unused after the initial enthusiasm fades because no workflow was redesigned around it.
The biggest visible problem. The firm identifies the most painful workflow. It is the one everyone complains about. The firm assumes that solving the most painful workflow first will produce the highest visible value. This is sometimes true. It is more often false. The most painful workflows are painful because they are complex, contested, or both. AI applied to a complex contested workflow without the underlying redesign will fail visibly and damage the broader integration effort.
The everywhere-at-once rollout. The firm decides to integrate AI across all functions simultaneously. The change management overhead consumes the team. Adoption is partial in every workflow and complete in none. The integration becomes a perpetual project rather than a series of completed deployments.
Each starting point reflects an assumption that AI value is a function of the tool. The value is a function of the workflow that the tool enables. The starting point should be the workflow.
The Right Starting Points
The workflows that produce the highest return on initial AI investment share three properties.
The first is high frequency. The workflow is performed many times per quarter, by multiple people. The volume creates compounding value from incremental efficiency. A workflow performed once per year produces no compounding regardless of the efficiency gain.
The second is structured input. The workflow consumes documents, emails, or data with predictable structure. The predictability allows extraction and processing to be reliable rather than experimental. A workflow that consumes idiosyncratic inputs produces unreliable outputs and erodes confidence in the integration.
The third is verifiable output. The workflow produces a result that can be checked against a known standard. The verifiability allows quality to be measured, errors to be caught, and the system to improve. A workflow whose output cannot be verified produces value that cannot be defended.
Property | What It Enables |
|---|---|
High frequency | Compounding efficiency value |
Structured input | Reliable extraction and processing |
Verifiable output | Measurable quality and improvement |
In a CRE firm, the workflows that meet all three properties are concentrated in document processing and analytical preparation. Lease abstraction. Rent roll reconciliation. Operating statement normalization. Comparable property compilation. Each is performed many times per quarter, consumes structured documents, and produces output that can be checked against source.
The First Workflow
The first workflow to integrate is rarely the most strategic. It is usually the most boring. The reason is that boring workflows are the ones where the team has the clearest sense of what "correct" looks like, the highest tolerance for incremental change, and the lowest political stakes for failure.
Lease abstraction is the most common first workflow for CRE firms. It is high-frequency, structured-input, and verifiable. It is also undervalued, which means the team performing it has both the most to gain and the least political resistance to change. The integration succeeds because the workflow was waiting to be improved and the team is incentivized to make it work.
First Workflow Candidate | Frequency | Input Structure | Verifiability |
|---|---|---|---|
Lease abstraction | High | Structured | High |
Rent roll reconciliation | High | Structured | High |
Comparable compilation | Medium | Semi-structured | High |
Operating statement normalization | Medium | Structured | High |
IC memo drafting | Medium | Unstructured | Low |
Investor reporting | Medium | Mixed | Medium |
Deal screening | High | Mixed | Medium |
Each candidate is a defensible starting point. The wrong starting points are the ones at the bottom of the list. IC memo drafting consumes unstructured input and produces output whose quality is contested. Investor reporting involves mixed input and political sensitivity. These workflows are integration candidates after the team has built credibility, not before.
The Sequence Matters
The first integration is not the integration that produces the most value. It is the integration that produces the credibility for the next integration. A firm that integrates lease abstraction successfully has demonstrated to its team that AI can be deployed without breaking existing processes. The demonstration is what allows the next integration to be approved.
The sequence that consistently works has three stages.
Stage | Workflow Examples | Purpose |
|---|---|---|
1. Document processing | Lease abstraction, rent roll reconciliation | Build extraction and verification discipline |
2. Analytical preparation | Comparables, operating statement normalization | Build trust in AI-prepared inputs |
3. Workflow orchestration | Deal screening, IC memo drafting | Integrate AI across multi-step workflows |
A firm that attempts stage three without completing stages one and two will fail. The team has not built the discipline or the trust required to use AI in workflows where the inputs and outputs are less structured. The integration looks like overreach, and the broader effort loses credibility.
What Has to Be True at Each Stage
Each stage requires specific conditions to succeed.
Document processing requires source citation discipline. The team has to expect that every extracted value carries a source. The discipline is what allows the AI output to be trusted. Without it, the AI is producing values that the team second-guesses or ignores.
Analytical preparation requires verification standards. The team has to define what "correct" looks like for each AI-prepared input. The standards are what allow the AI output to be checked. Without them, the team accepts whatever the AI produces or rejects whatever they cannot easily verify.
Workflow orchestration requires judgment calibration. The team has to define where AI executes and where humans decide. The calibration is what allows the orchestration to scale. Without it, the AI either makes decisions it should not or escalates decisions that humans should have delegated.
Each requirement is a workflow design decision, not a technology selection decision. The technology supports the requirements. The requirements determine whether the integration produces value.
What "Done" Looks Like for the First Integration
A successful first AI integration meets the following criteria:
The selected workflow is performed enough times that the team can measure improvement.
The AI output is verified against source for every material field, with citations preserved.
The team has defined what "correct" looks like and has a process for handling exceptions.
Adoption is consistent across the team performing the workflow, not selective.
The integration produces measurable time or quality improvement that the team can defend.
If the integration is used by some team members and avoided by others, the integration is incomplete.
Conclusion
AI integration succeeds when it starts in the right place and proceeds in the right sequence. The right place is the workflow that is high-frequency, structured-input, and verifiable. The right sequence builds discipline before scope and credibility before complexity. Firms that start small, succeed visibly, and expand deliberately compound the value of every integration. Firms that start with the most ambitious workflow or the broadest rollout will produce pilots that never become production. The first integration is not the most valuable. It is the precondition for the next one.