The energy transition depends on storage. Battery Energy Storage Systems (BESS) are being deployed at record pace, and investors compete to acquire the projects behind that growth. Before any acquisition comes screening: an investor receives a short teaser, later a data room full of documents, and must decide within weeks whether a prospect is worth pursuing. This thesis studies that screening step at GIGA Storage B.V. (hereafter: the BESS developer), a Dutch company that develops, acquires, owns, and operates utility-scale BESS, and asks:

How can a Large Language Model-based screening instrument be designed to convert unstructured prospect documentation into a screening record that supports early-stage investment risk assessment for utility-scale BESS acquisition?

The problem underlying this question is first established empirically. Interviews with nine experts and an analysis of the acquisition process reveal five recurring bottlenecks in current practice. Screening relies heavily on information supplied by the selling party (the prospect developer). Several specialists read the same documents in parallel without a shared overview. There is no cross-functional evidence gate that forces a structured go/no-go moment. Assumptions shift between phases and between advisors without being reconciled, and the evaluation is directed mainly at the current state of a project rather than at the state it must reach to become operable and profitable. The literature offers no ready method for these conditions: established techniques need historical data or quantified inputs, and early-stage BESS screening has neither.

Following Design Science Research, the thesis converts this problem into four results that build on one another. A structured literature review, merged with the interview findings, delivers a catalogue of 52 investment risk factors across 12 categories, mapped onto the 4P screening structure used in practice (Plot, Power, Permit, Profit); to the author's knowledge, it is the first factor-level risk catalogue for utility-scale BESS acquisition. Requirements elicitation delivers 41 requirements, of which seven define the design core. A conceptual process design then splits the work: assessing a prospect requires judgement, but most screening effort goes into something more tractable, turning unstructured prospect documents into structured facts. The design decomposes LLM-supported prospect analysis into four blocks that converge on the assessment, and this thesis realises the foundational block, filling the checklist, as a working prototype.

The prototype reads a teaser and a data room and fills a 152-item checklist in three phases. Every answer carries an evidence state (found, probably found, or not found), a source with page number, and a quote that is checked mechanically against the cited document. A human reviewer gates every phase: feedback and approval can be given on the extraction, critical values must be confirmed explicitly, and once a person edits or confirms an item, no later model output silently overwrites it. In the final phase, the tool verifies key claims against publicly available sources such as company registers and grid operators.

The frozen prototype was then evaluated on nine real prospects from the acquisition pipeline. A first quick extraction costs a few cents and completes within a minute; a complete deep run costs a few dollars. The public-source check corrected or extended 62% of the claims it verified, confirming the value of an independent leg in a process that leans on seller-provided information. Three near-identical sibling prospects enabled a consistency probe: 87% of items were classified identically, and the divergence traced mainly to gaps in the item specification rather than to the model. A group session with six practitioners endorsed the design and its scope, and the team asked to start using it. Two record walk-throughs with experts who know one acquired prospect first-hand completed the picture. In the first, the expert judged the record a usable screening basis and identified one substantive extraction error, which was corrected at the review gate. The second, an in-depth due-diligence audit, delivered the evaluation's sharpest finding: the extraction was frequently right in substance but drew on the seller's own overview documents rather than the underlying evidence (61% of quotations, against the 5 to 10% that the auditing expert considers adequate). That shortfall lies in the item specification and in the routing of documents to the model, both properties of the design rather than of the model, and both improvable by the same means this thesis demonstrates.

The result is therefore a tested proof of concept, which provides the foundation on which LLM-supported investment risk assessment can be built. The structuring layer proved technically feasible under the tested configuration, received positive formative feedback from the practitioners who would use it, and is governed by checks and balances that keep the human in control: evidence states that prefer an honest absence, verified sources behind every claim, human decisions that cannot be silently overwritten, spending capped in code, and a model that may not claim checks it did not perform. Two quantities that an adoption decision turns on remain unmeasured, and they define the follow-up study: extraction accuracy and time saved against a known baseline. Beyond the prototype, the thesis contributes the risk-factor catalogue, the requirements specification, a fully documented and reproducible prompt set, and eight design principles for LLM integration in confidential high-stakes document work.

Screening more prospects, more rigorously, with the same expert capacity means that viable storage projects are identified and funded sooner. This thesis contributes one component of that development: a working and formatively evaluated instrument that applies LLMs to the screening step, in support of a more efficient realisation of resilient electricity infrastructure.
... Read more