AI in finance covers three different technologies, namely predictive models, generative models, and agentic systems. It produces measurable results where the input is unstructured, the policy is written down, and the answer is checkable afterward. Accounts payable and expense audit fit that shape. Estimates and reserves do not.
Treating the three technologies as one is how budgets get misallocated.
Predictive models learn patterns from historical data and score new cases. Fraud scoring, payment-date prediction, collections prioritization, and demand-linked forecasting all belong here. They are mature, well understood, and unglamorous, which is why they get overlooked in favor of newer options that do the same job worse.
Generative models read and produce language. In finance they read documents, extract meaning from unstructured input, draft explanations, and summarize. Their weakness is that they produce fluent output whether or not they are correct, so they need a checkable answer behind them.
Agentic systems pursue a goal across steps, choose their own tools, act, and escalate when uncertain. They usually contain generative models. What differs is that they take action, which is why they carry a governance requirement the other two do not.
Most finance problems are best solved by one of these three technologies in particular. Vendors selling all three under one label make that harder to see.
The pattern across deployments producing hard numbers is consistent. The input is a document or a message, the correct answer is written down somewhere, and the answer is checkable after the fact.
Accounts payable fits that pattern exactly. Qualcomm reported moving autonomous invoice processing from 14 percent to 61 percent with 21 agents across six categories, cutting manual work 40 percent. Applied Industrial Technologies reached 87 percent autonomous AP on more than 500,000 invoices a year.
Expense audit fits the same pattern equally well. Every expense report includes a receipt, and a written policy defines what to test it against. Where teams once sampled 10 to 20 percent after reimbursement, full pre-payment coverage is now standard practice. Takeda reviews 100 percent of expenses with AI. Spectrum Brands reached 72 percent auto-approvals and cut expense processing from three weeks to three hours.
Compliance screening fits the pattern wherever a named rule exists, covering Sarbanes-Oxley evidence collection, Foreign Corrupt Practices Act screening, and value-added tax validation.
Reconciliation and close support are partly automated today. Forecasting improves at the pattern level and not at the judgment level. Reserves, estimates, and hedging decisions remain advisory almost everywhere, correctly.
Two 2026 surveys, read together, describe the state of AI in finance honestly.
Deloitte's Q2 2026 CFO Signals survey, published July 2026 from 200 North American CFOs at billion-dollar companies, found 93 percent reporting AI use across key operations, 44 percent for financial planning and budgeting, and 41 percent for analyzing financial data. It also found 46 percent naming cost uncertainty as their biggest internal concern, and 59 percent naming the balance between deploying quickly and managing risk as their top challenge.
McKinsey's State of AI survey, published August 2026 from 1,719 respondents across 97 countries, found 40 percent of respondents at organizations above $1 billion in revenue scaling AI agents, up from 27 percent. It also found 37 percent attributing at least some earnings impact to AI, about the same share as the year before, and 6 percent reporting 5 percent or more.
Use is near universal, and value is concentrated. The variable separating the groups was workflow redesign, present in nearly three-quarters of high performers and about one-quarter of the rest.
Four constraints hold across every deployment, and stating them upfront saves a quarter.
Gartner predicted in June 2025 that over 40 percent of agentic AI projects would be canceled by the end of 2027, citing unclear business value and inadequate risk controls, and estimated roughly 130 genuine agentic vendors among the thousands using the label.
Horizontal AI platforms ship without a finance opinion. A general system holds no view on a three-way match tolerance, a fapiao, a Sunshine Act disclosure, or a Section 174 capitalization test, so your team encodes all of it and maintains it as the business changes. That work never appears in the license cost.
The second gap is evaluation vocabulary. Task completion rate and tool-call success are engineering measures. A controller needs to know whether the system's general ledger coding matched what the team would have done, and most products cannot answer that.
We build for the workflows that fit the shape above rather than for the category. Our platform runs on ZenLM, a family of finance-specific models covering document understanding, semantic categorization of financial data, and routine task execution, so the model holds the domain knowledge instead of configuration your team maintains.
Every Agent is benchmarked against historical and live data and compared against human expert decisions before it is granted authority. Every action is visible and auditable, and each Agent escalates at a stated uncertainty threshold. Data is de-identified before any model training, described under trustworthy AI. Across our customer base, CFOs report reductions in finance operating costs of up to 50 percent and automation rates above 80 percent, running on pre-trained agents for AP, expense, and compliance.
AI in finance produces measurable results where the input is unstructured, the policy is written, and the answer is checkable, and it produces confident noise everywhere else. Take your three largest workflows, test each against those conditions, and fund only the ones that pass. Start with the one where you already measure a cost per touch.
Reading and validating documents, scoring risk, and completing transactional workflows end to end. Accounts payable and expense audit are the most mature, because both start with unstructured input and test it against a written policy. Deloitte found in July 2026 that 44 percent of CFOs use AI for financial planning and budgeting and 41 percent for analyzing financial data.
McKinsey found in August 2026 that 37 percent of respondents attributed at least some earnings impact to AI while only 6 percent reported 5 percent or more. The separator was workflow redesign, present in nearly three-quarters of high performers and about one-quarter of everyone else. Tools applied to unchanged processes deliver speed rather than structural cost reduction.
Anything without a checkable correct answer. Reserves, estimates, hedging decisions, and judgments depending on business context a model cannot see should stay advisory. A generative model will produce a fluent answer for each of them, and there is no way to grade it.
Predictive models score cases against learned patterns. Generative models read and produce language, including reading documents. Agentic systems pursue a goal, act, and escalate when uncertain, and they usually contain generative models. Agentic systems take action, which is why they carry governance requirements the other two do not.