When executives ask my team where they should deploy AI first, they usually point at the decision itself. Underwriting AI. Claims decision making AI. Lending decision making AI. The decision is the visible part of the workflow, and the cost of the decision is what shows up in the operating budget.
In my experience working with insurers and adjacent industries over 10 years, this is not where the highest-value early AI deployment lives. The decision is rarely the most time-consuming part of the workflow. The time consumption sits in the work that has to happen before the decision can be made at all.
I call this the data-collection tax. It is the time professionals spend gathering, reconciling, and verifying information before they can apply their judgment to it.
How the tax shows up
In claims, an adjuster spends most of the working day collecting facts, photographs, repair estimates, and policy details. The decision takes minutes. In underwriting, an underwriter spends the day collecting financial statements, loss histories, and exposure data. The decision takes minutes.
This pattern is not unique to insurance. It shows up wherever knowledge workers apply professional judgment to a defined question. Litigators collect facts and apply legal reasoning. Doctors collect symptoms, test results, and history and apply clinical reasoning. Risk analysts collect operational data and apply probabilistic reasoning. In each case, the collection time is the larger share of the workflow. The reasoning time is the smaller share, but it is where the value lives.
When my team published research from 20 US insurance leaders this June, we asked where the biggest gap was between how claims decisions get made today and how the leaders would want them made. The most common answer, named by close to two-thirds of respondents, was adjuster time spent gathering data instead of evaluating the claim.
That answer landed clearly ahead of every other operational complaint, including cycle time, decision consistency, and regulatory documentation effort. The bottleneck in their operations is not the decision. It is the work that has to happen before the decision.
Why AI fits here
The reasoning is operational and regulatory, and it generalizes across industries.
Operationally, AI handles structured but tedious data-collection work better than it handles judgment-intensive decision making work. Parsing a repair estimate in a PDF, extracting fields from a medical chart, reconciling a customer record across three systems of record, summarizing a deposition transcript, flagging a transaction against a watchlist. These are tasks where the input is unstructured, the output is structured, and the success criterion is verifiable. They are also tasks where the cost of an AI error is bounded. A wrong field extraction can be corrected by the human who reviews it. A wrong document summary can be checked against the source.
Compare this with AI deployed at the decision itself. The input is structured but contextual. The output is a judgment. The success criterion is contested. The cost of an error is high and often regulated. Most enterprise AI deployments that stall in pilot stall here.
Regulatorily, AI deployed on data-collection work attracts much less scrutiny than AI deployed at the decision. The NAIC Model Bulletin on AI in insurance, the EU AI Act, the NIST AI Risk Management Framework, and various federal and state-level expectations all focus their highest scrutiny on AI systems that make or substantially influence decisions affecting consumers. AI that pre-processes documents, intakes information, or routes work to the right human is generally outside the highest scrutiny tier. This means the deployment can ship faster, with less governance overhead, while still delivering measurable value.
Four categories of opportunity
Four categories of data-collection work cover most of the opportunity in most enterprises.
The first is document interpretation. Any workflow that requires a professional to read and extract information from unstructured documents has a data-collection tax. Policy documents in insurance. Loan applications in banking. AI can extract structured information in seconds, leaving the professional to verify it and apply judgment.
The second is intake and triage. Any workflow that begins with a customer providing information through multiple channels, in inconsistent formats, has a data-collection tax. First notice of loss in insurance. Application intake in lending. Patient onboarding in healthcare. AI can normalize the input, route it to the right specialist, and pre-populate the work queue.
The third is system reconciliation. Any workflow that requires manual reconciliation across multiple systems of record has a data-collection tax. Customer records across policy administration, billing, and claims in insurance. Patient records across electronic health records, billing, and pharmacy in healthcare. AI can perform this reconciliation continuously in the background.
The fourth is audit trail generation. Any workflow that requires documentation of how a decision was reached has a data-collection tax. Regulatory documentation in insurance and financial services. Clinical documentation in healthcare. Compliance documentation in any regulated industry. AI can generate the documentation as a byproduct of the decision rather than as a follow-on task.
Each of these has the same operational profile. The data foundation work is bounded. The output is verifiable. The regulatory exposure is low. The measurement is clear. The professional whose time is freed can spend more of the day on the higher-value work the organization is actually paying for.
A framework shift
The framework shift this enables is small in description and large in consequence. Most enterprises think about AI as automation of decisions. The more productive frame is AI as liberation of decision-makers. The decision itself is rarely the bottleneck. The data-collection tax is.
If you are sitting on an AI proposal that targets the decision itself, take a hard look at what happens before the decision. The faster, cheaper, less risky deployment is probably already in front of you.
