--- title: "AI Agents for Finance: The Context Layer for the CFO Office" url: "https://atlan.com/know/ai-agent/ai-agents-for-finance/" description: "A finance agent fails not because the model is weak but because \"revenue\" means three things. Learn the governed context layer the CFO office needs." author: "Emily Winks" author_role: "Data Governance Expert" published: "2026-07-01" updated: "2026-07-01" --- --- AI agents for finance fail in the office of the CFO for a reason that has nothing to do with the model: the agent does not know what your numbers mean. Ask a finance copilot for "revenue" and it may return bookings from the ERP, recognized revenue from the warehouse, or the ARR that marketing loaded into the planning tool, and it will pick one with total confidence. Atlan is the Context Layer for AI: the governed layer between your fragmented finance systems and every agent that reads from them, so a forecasting agent, a close agent, or a board-reporting agent all [reason](https://atlan.com/know/ai-agent/ai-agent-planning/) from the one certified definition finance will actually defend. The bottleneck was never intelligence. As models commoditize, the finance-specific [context that tells an agent](https://atlan.com/know/context-layer-for-ai-agents/) what "margin" means, how the close actually runs, and which number is approved for the board is the compounding advantage. Performance is a function of intelligence and context, and in finance the context is the part you own. > **Building for a bank, insurer, or asset manager?** This page is about the corporate finance function, the office of the CFO: FP&A, close, accounting, treasury, and management reporting. If your problem is model risk management, SR 26-2, or trading-desk conduct, read [AI agents for financial services](https://atlan.com/know/ai-agent/ai-agent-in-finance/) instead. The two problems rhyme, but the systems and stakeholders differ. --- ## How are AI agents being used across the office of the CFO? Adoption in the finance function is real and steady rather than explosive. According to Gartner's 2025 AI in Finance survey of 183 CFOs and senior finance leaders, [59% of finance functions reported using AI](https://www.cfodive.com/news/cfos-ai-adoption-slows-challenges-mount-gartner/805949/), up only slightly from 58% the year before, and [67% of those already using it are more optimistic](https://www.cfo.com/news/cfo-optimism-around-ai-rises-as-adoption-levels-off-gartner-/805868/) about AI than they were a year earlier. The intent is clearly agentic: reporting on the state of AI in finance, [Pigment notes that 17% of finance teams are already deploying generative agents](https://www.pigment.com/blog/the-state-of-ai-in-finance-10-statistics-fp-a-leaders-should-know), with roughly three quarters of leaders expecting agentic AI to be routine by 2028. McKinsey's State of AI work, [as reported by ChatFin, found 44% of CFOs now using generative AI](https://chatfin.ai/blog/mckinsey-44-of-cfos-now-using-ai-for-fpa-are-you-the-7-still-behind/) across five or more use cases, up from 7% a year earlier. The use cases already in production span the full breadth of the finance function: * **FP&A:** Agents build driver-based forecasts, run variance analysis against plan, and draft the commentary that explains why actuals moved. * **Month-end and quarter-end close:** Agents reconcile accounts, propose journal entries, flag exceptions, and assemble the close checklist so controllers review rather than assemble. * **Accounting operations:** Agents match invoices to purchase orders in AP, apply cash in AR, and clear the routine reconciliations that consume analyst time. * **Treasury and cash:** Agents forecast cash positions, monitor liquidity against covenants, and surface funding actions before a shortfall becomes a fire drill. * **Management and board reporting:** Agents pull the approved figures, assemble the recurring board pack, and draft narrative around the numbers finance has certified. According to [Deloitte's Q4 2025 CFO Signals survey](https://www.deloitte.com/us/en/about/press-room/deloitte-q4-2025-cfo-signals-survey.html), 87% of CFOs expect AI to be extremely or very important to their finance departments in 2026, and more than half say integrating AI agents will be a transformation priority. The gap between that intent and production is almost never the model. It is the context. --- ## Why the finance function is a hard context environment Finance runs on a stack of systems that were never designed to agree with each other. The ERP (SAP, Oracle, or NetSuite) holds transactions and the general ledger. The EPM or planning tool (Anaplan, Pigment, or Adaptive) holds the plan, the forecast, and its own version of every driver. The warehouse holds the modeled, transformed version that analytics teams built. BI holds the dashboards executives actually look at. Each of these systems has a table that some team calls "revenue," and none of them are guaranteed to mean the same thing. A finance [AI agent](https://atlan.com/know/ai-agent/what-is-an-ai-agent/) dropped into that stack inherits every one of those disagreements. It sees multiple candidate tables for the metric it was asked about, no signal about which one is authoritative, and no record of how any of them were built. Left to its own devices it retrieves the first plausible match, or worse, [blends two definitions](https://atlan.com/know/ai-agent-hallucination/) into a number that reconciles to nothing. This is the finance version of the general problem behind [why AI agents fail in production](https://atlan.com/know/why-ai-agents-fail-in-production/): the model is fine, but the context is missing. The cost of that missing context is not hypothetical. Gartner research, [widely cited across industry analyses, puts the average cost of poor data quality at $12.9 million per year](https://www.integrate.io/blog/data-quality-improvement-stats-from-etl/) per organization. In finance, that cost shows up as restated forecasts, close cycles that slip, and a board number that gets challenged in the room. --- ## The definitional conflict problem: when "revenue" means three things Every finance organization has a definition problem it has learned to live with in the era of human analysts, because a human knows which "revenue" to use for which audience. An agent does not, unless you tell it. Consider the canonical case. Finance's "revenue" may mean recognized bookings for a period. Marketing's "revenue," loaded into the same planning tool, may mean ARR. Both are labeled revenue. A forecasting agent asked to project next-quarter revenue has no way to know that the two are different constructs, so it either picks one at random or averages them, and produces an answer nobody in the room will stand behind. The same conflict repeats across every metric the CFO office reports: * **Revenue:** Bookings versus recognized revenue versus ARR, differing by system and by team. * **Margin:** Gross versus contribution versus operating, with different cost allocations behind each. * **EBITDA:** Reported versus adjusted, where the adjustments themselves are a matter of policy. * **Headcount:** Filled positions versus approved positions versus full-time equivalents, differing between HR and the plan. * **Cost center:** Hierarchies that were reorganized mid-year, so the same code rolls up differently depending on the reporting period. For close and reporting, resolving this requires source-to-report lineage and KPI certification: the agent must be able to trace a number back to its origin and know which version is approved. For management reporting, it requires the finance-approved definition to be applied every single time, without exception. A [semantic layer](https://atlan.com/know/semantic-layer/) that encodes one certified meaning per metric is what turns an agent from a plausible guesser into a reliable one. --- ## Knowledge, Expertise, and Norms: the three kinds of context a finance agent needs Definitions are necessary but not sufficient. A finance agent that knows what "operating margin" means still fails if it does not know how the close is run or what it is allowed to touch. Useful [context for an AI agent](https://atlan.com/know/ai-agent/ai-agent-context/) has three parts, and missing any one of them breaks the agent. * **Knowledge, what the entities and metrics mean.** The certified definition of revenue, margin, EBITDA, headcount, and cost center, plus the entity model that says how a legal entity rolls up to a segment and a segment to the consolidated group. This is the layer most teams think about first. * **Expertise, how finance work actually gets done.** The close calendar and its dependencies, the order in which reconciliations run, the allocation methodology behind contribution margin, the accrual rules a controller applies. An agent that does not encode this reconciles accounts in the wrong sequence and produces a close that does not tie out. * **Norms, what the agent is allowed to do.** Who can see unreleased earnings figures, which numbers require controller sign-off before they leave the building, what an agent may draft versus what it may post to the ledger. These are [policy rules](https://atlan.com/know/ai-agent-governance/), and in finance they are not optional. Miss the knowledge and the agent uses the wrong number. Miss the expertise and it does the right steps in the wrong order. Miss the norms and it exposes a pre-announcement figure or posts an entry no human approved. A [production finance agent](https://atlan.com/know/how-to-build-an-ai-agent-step-by-step-guide/) needs all three, delivered together. --- ## What a governed finance AI architecture looks like A production-grade [context layer](https://atlan.com/know/what-is-context-layer/) for the office of the CFO has four layers. Each resolves one of the failure modes above. ### Layer 1: The canonical finance glossary and semantic layer Every reported metric, revenue, margin, EBITDA, headcount, cost center, gets one certified definition with a named owner in finance, mapped to the specific source tables that produce it. This is the layer that ends the "revenue means three things" problem, because every agent resolves the term the same way regardless of which system it entered through. Encoding it as [business context for AI](https://atlan.com/know/business-context-for-ai/) makes the definition machine-readable rather than trapped in a policy PDF. ### Layer 2: Source-to-report lineage Column-level lineage traces each figure from source systems through every transformation to the number in the board deck. This gives the agent the ability to explain where a number came from and what changed since last period, and it gives the controller and audit committee the [provenance](https://atlan.com/know/enterprise-data-graph/) they need to sign off. Without it, an agent-produced number is a black box no one will approve. ### Layer 3: KPI certification and the AI asset registry A review-and-approve workflow marks the finance-approved version of each metric, and an [AI asset registry](https://atlan.com/know/what-is-ai-registry/) records every agent, the definitions it consumes, and the policies it operates under. Together they give the CFO office a single point of truth for what each agent is allowed to report and where it got its numbers. ### Layer 4: Policy enforcement and activation Access and sensitivity rules are enforced at the layer that delivers context, before anything reaches the agent, not reimplemented inside each copilot. That same layer [activates context through MCP, SQL, and API](https://atlan.com/know/ai-agent/ai-agent-tool-use/), so every finance copilot, NL2SQL bot, and reporting agent consumes identical governed context. Every agent action is captured as a [decision trace](https://atlan.com/know/what-are-decision-traces-for-ai-agents/) linking the data, definitions, and policies behind the output. --- ## How Atlan supports finance AI agents in production Atlan operates as the governed context layer between your finance systems and your agents, connecting ERP, planning, the warehouse, and BI to every copilot and reporting agent through one policy-enforced infrastructure. * **Canonical finance glossary and semantic definitions:** Certified definitions for revenue, margin, EBITDA, headcount, and cost center, each with explicit ownership, so every agent queries the finance-approved meaning. * **End-to-end column-level lineage:** Source-to-report provenance reverse-engineered from SQL, pipelines, and BI, so an agent can explain where any number came from and what changed. * [**Context Agents**](https://atlan.com/know/context-agents/): AI agents that mine descriptions, metric definitions, and process maps from your existing SQL and lineage, then keep them current as the estate changes. Context Agents have generated 690,000+ descriptions across 50+ enterprise customers, with 87% rated on par with or better than human-written. * [**Context Engineering Studio**](https://atlan.com/context-engineering-studio/): The workspace where finance and data teams build, test, review, and certify the context layer before agents reach production. * [**Context Repos**](https://atlan.com/know/ai-agent/context-repository-for-ai-agents/): [Versioned](https://atlan.com/know/ai-agent/context-versioning-for-ai-agents/), portable bundles of finance context, so what was approved for the Q3 board pack is preserved exactly as it stood when a number is later questioned. * **MCP Server** and policy enforcement: [Atlan's MCP Server](https://atlan.com/mcp-server/) is the governed endpoint every finance agent consumes. Before context reaches an agent, it enforces what a metric means, whether the data meets the freshness threshold, and which policies apply. --- ## A finance customer building governed AI context: Workday Workday built a revenue-analysis agent that could not answer the questions finance asked of it. The agent was capable; what it lacked was a translation layer between the raw data and the business language that finance teams had spent years agreeing on. When Workday exposed that shared language to AI through Atlan's MCP Server, the same agent started producing answers finance could trust. "Atlan captures Workday's shared language to be leveraged by AI via its MCP server. As part of Atlan's AI Labs, we're co-building the semantic layer that AI needs." - Joe DosSantos, VP Enterprise Data & Analytics, Workday Watch Now "With Atlan, we cataloged over 18 million data assets and 1,300+ glossary terms in our first year, so teams can trust and reuse context across the exchange." - Kiran Panja, Managing Director, CME Group Watch Now Grounding the revenue-analysis agent in that shared context drove a 5x improvement in AI-analyst accuracy. The lesson generalizes to any office of the CFO: the model was ready, the context was not, and closing that gap is what made the agent usable. Workday has since gone on to [build developer tooling for AI agents across HR, finance, and IT](https://newsroom.workday.com/2026-06-02-Workday-Launches-New-Tools-for-Developers-to-Build,-Connect,-and-Verify-AI-Agents-For-HR,-Finance,-and-IT) on that same foundation. --- ## Moving forward with AI agents for the finance function The path to production finance agents is an [architectural](https://atlan.com/know/ai-agent/ai-agent-architecture-explained/) one, and the sequence matters. Build the context first, then build the agents. Start with a canonical finance glossary and certify the definitions of the metrics that end up in front of the board. Add source-to-report lineage so every one of those numbers can be traced and explained. Pick a first workflow where correctness is easy to verify and the payoff is clear: variance commentary in FP&A, account reconciliation in the close, or the recurring board pack in management reporting. Use that deployment to prove the context layer works, then reuse the same certified definitions and lineage to [earn broader autonomy](https://atlan.com/know/ai-agent-stack/) across forecasting, treasury, and consolidated reporting. As Deloitte's [work on generative AI in finance transformation](https://www.deloitte.com/us/en/what-we-do/capabilities/finance-transformation/articles/generative-ai-in-finance-transformation.html) makes clear, the finance functions that win with agents are the ones that treat trusted context as the foundation, not an afterthought. The context is your IP. Keep it portable, keep it certified, and every agent you deploy compounds on the last. Book a Demo --- ## FAQs about AI agents for finance ### What are AI agents for finance? AI agents for finance are autonomous or semi-autonomous systems that support the office of the CFO across FP&A, month-end and quarter-end close, accounting, treasury, and management reporting. They plan, forecast, reconcile, and draft reporting by reasoning over data in ERP, planning tools, the warehouse, and BI. Unlike a general chatbot, a finance agent must produce the specific number the CFO will defend, which depends on certified definitions and lineage rather than model capability alone. ### Why do finance AI agents give inconsistent numbers? Finance agents give inconsistent numbers because the same term is defined differently across systems. "Revenue" may mean bookings in one table and recognized ARR in another; "margin," "EBITDA," "headcount," and "cost center" each carry competing definitions across ERP, EPM, the warehouse, and BI. An agent that picks whichever table responds first, or blends two definitions, invents an answer nobody in finance will defend. The fix is a canonical definition per metric that every agent queries. ### How is this different from AI agents for financial services? AI agents for finance address the finance function inside any company: FP&A, close, accounting, treasury, and management reporting for the office of the CFO. AI agents for financial services address the banking, insurance, and asset-management industry, where model risk management, SR 26-2, and market-conduct rules apply. The context challenges rhyme, but the systems, definitions, and stakeholders differ. If you are building for a bank, insurer, or asset manager, read the [financial services guide](https://atlan.com/know/ai-agent/ai-agent-in-finance/) instead. ### What is source-to-report lineage and why does a finance agent need it? Source-to-report lineage is a column-level record of how a reported figure was produced, from source systems through every transformation to the number in the board deck or filing. A finance agent needs it so it can explain where a number came from and what changed since last period. Without lineage, no controller or audit committee can sign off on an agent-produced figure, because there is no way to trace it back to its origin. ### What does the office of the CFO need before deploying finance AI agents? Before deploying finance AI agents, the office of the CFO needs a canonical finance glossary with certified definitions and named ownership, KPI certification that marks the finance-approved version of each metric, source-to-report lineage for provenance, policy enforcement applied before context reaches an agent, and a governed activation layer so every copilot and reporting agent consumes the same context through MCP, SQL, or API. --- ## Sources 1. [CFOs' AI adoption slows as challenges mount: Gartner, CFO Dive](https://www.cfodive.com/news/cfos-ai-adoption-slows-challenges-mount-gartner/805949/) 2. [CFO optimism around AI rises as adoption levels off, CFO.com](https://www.cfo.com/news/cfo-optimism-around-ai-rises-as-adoption-levels-off-gartner-/805868/) 3. [The state of AI in finance: 10 statistics FP&A leaders should know, Pigment](https://www.pigment.com/blog/the-state-of-ai-in-finance-10-statistics-fp-a-leaders-should-know) 4. [Technology Transformation Emerges as a Top Priority for CFOs in 2026: Deloitte Q4 2025 CFO Signals Survey, Deloitte](https://www.deloitte.com/us/en/about/press-room/deloitte-q4-2025-cfo-signals-survey.html) 5. [McKinsey: 44% of CFOs Now Using AI for FP&A, ChatFin](https://chatfin.ai/blog/mckinsey-44-of-cfos-now-using-ai-for-fpa-are-you-the-7-still-behind/) 6. [Data Quality Improvement Stats: 50+ Key Facts (Gartner $12.9M/yr), Integrate.io](https://www.integrate.io/blog/data-quality-improvement-stats-from-etl/) 7. [Generative AI in Finance Transformation, Deloitte](https://www.deloitte.com/us/en/what-we-do/capabilities/finance-transformation/articles/generative-ai-in-finance-transformation.html) 8. [Workday Launches New Tools for Developers to Build, Connect, and Verify AI Agents for HR, Finance, and IT, Workday](https://newsroom.workday.com/2026-06-02-Workday-Launches-New-Tools-for-Developers-to-Build,-Connect,-and-Verify-AI-Agents-For-HR,-Finance,-and-IT)