Research highlighted by UC Davis shows why a reviewer's accounting knowledge matters when AI extracts information from contracts. For firms, adoption depends on defining the task, preserving the evidence and measuring the cost of qualified review
A finance team can spend considerable time locating the terms of an agreement before it can decide how to account for them. A payment may be described in one section, qualified in another and altered by a schedule near the end. An AI assistant can make the search much faster. The business still needs someone to judge whether the passages identified are complete, which rules apply and whether a plausible explanation follows from the evidence. This creates an adoption question with consequences for staffing and training: what expertise must remain available when the first reading of a contract is delegated to software?
A UC Davis research account published on 10 November 2026 brings that question into focus. Ayung Tseng and C. Janie Chang examined workflows for extracting accounting information from debt, lease and sales contracts, varying human verification and accounting expertise. The university’s account reports that a non-accountant’s verification process reduced accuracy in one comparison, while accountant-guided refinement improved it. This is a report of a bounded research exercise, rather than an enterprise rollout. Its importance lies in showing why appointing a reviewer is only the beginning of designing a review process.
The underlying paper, Domain Knowledge and Human-AI Collaboration, describes nine workflow variants and concludes that verification intensity and expertise were complementary in the tasks studied. The SSRN record identifies a version written in August and revised in September, so the November communication brings attention to existing research. The study addresses a particular class of information-extraction work. A firm considering it needs to ask how closely its own contracts, accounting questions and review arrangements resemble those conditions before treating the findings as a basis for operational change.
Define the question before asking for an answer
In an illustrative pilot, a small finance team could begin with one recurring question about a familiar type of contract. It might ask the assistant to identify specified payment provisions, quote them and show the surrounding clauses needed to interpret them. This would give the reviewer a manageable task with visible evidence. Asking for a complete accounting conclusion at the outset would bundle several decisions together, making it harder to see whether a failure came from missing text, misunderstanding a term or applying the wrong rule. A defined question helps the team decide what a correct answer requires.
The source document also needs a clear boundary. The team may have the signed contract, an earlier draft, an amendment and a schedule stored under similar names. If the assistant receives only some of those documents, its answer may be coherent within that limited record while being wrong for the agreement the business is actually using. The reviewer should be able to see which files and versions were supplied. Where an amendment changes the interpretation, the output needs a traceable connection to it. A model’s fluency cannot compensate for an incomplete set of materials.
The accounting framework supplies another boundary. A business operating under one set of reporting rules cannot assume that an answer grounded in another is applicable. For an adopting team, the question should identify the framework, the relevant period and the exact issue being investigated. The authorised source for any rule should be accessible to the reviewer. The research discussed by UC Davis uses US accounting contexts. Its broader workflow lesson can inform other settings, while the substantive conclusion for a different framework requires the appropriate local expertise and authoritative material.
Expertise changes the review itself
The university’s account gives a concrete warning about apparent precision. In one lease analysis, the system supplied an accounting-standard reference whose content was fabricated. A reviewer familiar with the standard could identify the problem and provide the correct material. The example matters because a citation’s shape can look reassuring while providing no support for the conclusion. A review process can require the relevant authoritative text and a clear connection between that text and the proposed treatment. The reviewer then has something to verify beyond the presence of a credible-looking reference number.
The researchers also describe refinement that made instructions shorter and more focused. That observation suggests a practical distinction between adding more directions and identifying the directions that matter. An expert can recognise which exception is relevant, which definition is being confused and which extra wording distracts from the question. The business case therefore includes the time needed to build a reliable task definition. Treating instruction-writing as a one-off administrative step could leave the team repeatedly correcting the same misunderstanding without changing the process that produces it.
A manager should also consider who can disagree with the output. A junior employee may be able to compare a quote with a page and still lack the experience to challenge the proposed interpretation. A review queue can separate those activities, allowing evidence checks to be completed before a specialist resolves a disputed conclusion. Such a division is a proposed operating arrangement, not a result reported by the study. It would need testing to ensure that the specialist receives enough context and that uncertain cases are escalated rather than silently accepted because the first check found the words in the document.
Measure accuracy alongside the cost of review
The UC Davis account reports perfect accuracy on new samples after refinement in the workflows it describes, while noting that each test sample contained only ten contracts. That sample size makes the result informative about those cases and leaves considerable uncertainty about broader performance. A firm should preserve that distinction when deciding what to test next. A new contract type, an unusual amendment or a document with poor extraction quality could introduce failures absent from the small samples. The goal of local validation would be to find where the process remains dependable and where the reviewer needs additional scrutiny.
A useful trial could compare ordinary review with assisted review on a set of appropriately cleared contracts whose answers are independently established. It could record missed terms, unsupported interpretations and the time required to resolve each disagreement. A particularly important category would be an answer that appears complete but omits a clause changing the conclusion. An average accuracy figure alone could hide a small number of consequential failures. The team would need to examine those cases individually and determine whether a change to the task, source preparation or review arrangement addresses them.
Costs need the same complete view. Model charges per contract are one input. The cost of preparing documents, checking passages, consulting a specialist and maintaining the workflow can be much larger. A faster initial answer may save little if it takes longer to verify than the conventional approach. Conversely, an assistant that reliably finds the relevant evidence could free a specialist to spend more time on difficult judgements. The firm can distinguish those possibilities by tracking the whole job, including the contracts for which the assistant provides little help, rather than measuring only successful demonstrations.
Keep a record that another professional can inspect
A reviewed output should retain the question, the documents used, the supporting passages and the conclusion reached by the responsible professional. That record makes it possible to revisit the work after a contract changes or a later reviewer asks how a decision was made. A system update can also change the answer to the same request. Keeping a stable record of the version and settings used helps the team investigate such differences. The purpose is practical continuity: a later colleague should be able to understand what evidence was available and which part of the conclusion involved judgement.
Broader professional experience supports attention to these arrangements. A July 2024 PCAOB staff report on outreach to auditors and preparers describes concerns about confidentiality, reliability, supervision and review as firms explored generative AI. It reports observations from that period and is explicitly a staff publication rather than a Board rule. It provides historical context for the same adoption problem. Access to a tool creates opportunities for research and drafting, while the organisation must decide which information it can supply and how the resulting work will be checked.
For a smaller practice, a restricted initial service may be more manageable than integrating every client file. It could test extraction on one category of agreement and use an approved environment with suitable confidentiality arrangements. An independent professional or sole trader could assess a narrow task against work they already know how to perform, preserving a manual route for unusual cases. These are possibilities that require validation, not claims of guaranteed savings. The scale of the implementation should reflect how much qualified review and maintenance the practice can sustain through its ordinary workload.
Invest in the judgement the tool depends on
Training is part of the long-term business decision. If experienced staff use AI to locate evidence, they can also explain why a proposed answer fails and make that explanation available to less experienced colleagues. A team that gives junior staff only the final output may save immediate reading time while reducing their opportunities to develop the judgement needed for later review. An adoption pilot could therefore examine how people learn from corrected cases, whether they can identify recurring mistakes and when they feel able to escalate uncertainty. Those questions connect current efficiency with the future capability of the practice.
The research communicated in November gives businesses a concrete reason to treat domain expertise as an input to AI adoption. The relevant investment includes defining the task, checking evidence and learning from failure. A successful workflow would make contract information easier to retrieve while leaving a clear account of the conclusion and the person responsible for it. Whether that arrangement improves performance has to be established on the firm’s own work, with the costs of review included. The strongest result would be more reliable decisions made with a better use of professional time, supported by records another accountant can understand.
Image credit: Accounting Calculator by DoubleCritch, CC0 1.0. Illustrative image. The source file is unedited. WordPress may resize or crop display versions.
