← Sakib Ahmed

What Has to Be True Before AI Works in FP&A

Most of the conversation about AI in FP&A jumps straight to the exciting stuff: automated variance commentary, agentic forecasting, natural language interfaces to planning models. But after six years working at the intersection of FP&A and technology, I've found that the teams who actually get value from AI aren't the ones with the flashiest tools. They're the ones who got the boring stuff right first.

This essay is about that boring stuff: the foundations that determine whether AI delivers real value or just generates plausible-sounding nonsense.

Where Current AI Falls Short

The typical FP&A analyst workflow has three stages: gather data from core systems, analyze the data, and act on the results.

Most AI progress in finance so far has accelerated the Gather and Act steps. Chatbots help people find information faster. Workflow automations move data between systems. But the Analyze step — the part where an analyst sits in a spreadsheet and applies judgment, variance analysis, scenario modeling, reconciliation, flux analysis — remains largely untouched.

That's where finance people spend the majority of their time. And it's the step current tools don't reach.

Closing this gap requires three foundations: clean data, connectivity, and meeting analysts where they actually work.

Foundation 1: Data Quality Is the Prerequisite

AI is excellent at middle-layer work: synthesis, pattern recognition, narrative generation. But it can only operate on what it can access and trust. For FP&A, that means:

  • Clean master data. Cost centers, GL accounts, department hierarchies, product area mappings. If your cost center structure is inconsistent or your GL accounts are miscoded, any AI built on top of that data inherits those problems. AI doesn't fix dirty data. It scales dirty data.
  • Structured outputs from planning tools. If your Anaplan or Pigment models export data in formats that require manual manipulation before they're usable, you've created a bottleneck that no amount of AI can route around.
  • Reliable data pipelines. AI agents that query Snowflake or NetSuite need to trust that the data is complete, timely, and reconciled. A forecast agent that pulls stale actuals will produce confident but wrong analysis.

The uncomfortable reframe: master data cleanup isn't maintenance work. It's unblocking AI. Every hour spent rationalizing cost centers or standardizing GL hierarchies directly expands the surface area where AI can operate reliably. Teams that treat data quality as a side project will keep wondering why their AI experiments don't scale.

Foundation 2: Connectivity

Even with clean data, AI agents need a way to reach it. The typical finance tech stack spans a planning tool, an ERP, a data warehouse, a reporting layer, and half a dozen spreadsheets. Without a standardized way for AI to connect to all of these, every use case requires its own bespoke integration, which means an engineering bottleneck, which means most use cases never get built.

This is where Model Context Protocol (MCP) matters. MCP is an open standard that gives AI models a consistent interface to external tools and data sources. Think of it as a universal adapter: instead of building a custom integration every time you want an AI agent to query Snowflake or interact with Anaplan, MCP provides one interface that works across all of them.

What FP&A teams can do themselves: personal productivity agents (Claude, ChatGPT for ad-hoc analysis and drafting), lightweight workflow automations (n8n, Zapier), and prompt engineering for recurring tasks. A technically curious FP&A analyst can drive a surprising amount of value here without engineering support.

What needs engineering: MCP implementations for enterprise systems, data pipeline work, and governance frameworks for agents that interact with financial reporting data. FP&A's role is to define the requirements and prioritize the use cases, not to build the infrastructure.

Foundation 3: Meet Analysts Where They Work

FP&A analysts live in spreadsheets. Not in chatbots, not in dashboards, not in Slack. Any AI strategy that asks analysts to leave their primary work surface to get AI help is fighting an uphill adoption battle. The biggest unlock for core finance functions is AI that works inside the spreadsheet, not outside it.

This leads to a distinction I find useful: sheet-native vs. sheet-accidental workflows.

Sheet-native workflows are ones where the spreadsheet is the legitimate work product: variance analysis workbooks, financial models that inform decisions, scenario analyses. The spreadsheet isn't a workaround. It's the artifact. AI that operates inside these spreadsheets is directly accelerating the actual deliverable.

Sheet-accidental workflows are ones where a spreadsheet became the system of record because no proper platform existed: master data tracking in Google Sheets, M&A deal tracking in ad hoc workbooks, pseudo-CRM files. For these, the right answer isn't smarter spreadsheets. It's migrating to a proper system.

The strategic implication: for sheet-native work, invest in AI spreadsheet tools now. For sheet-accidental work, use what you learn from AI experiments to define requirements for proper platform solutions. The short-term AI work becomes a discovery mechanism for what the long-term platforms need to do.

Will Enterprise Planning Tools Go Away?

Short answer: no. Not in the near or medium term.

The strongest evidence: AI-native companies themselves are choosing to implement enterprise planning tools. Companies like Anthropic and OpenAI, organizations with world-class engineering talent and direct access to the most advanced AI models, are implementing tools like Pigment rather than building their own planning systems. If anyone could replace Anaplan or Adaptive with AI, it would be them. They're choosing not to.

Why? Because these tools solve specific hard problems that AI doesn't address:

  • Large in-memory calculations. Planning tools are purpose-built to run multi-dimensional models across thousands of cost centers, departments, and time periods simultaneously. An AI agent can generate a forecast narrative, but it's not a replacement for an engine that calculates a full P&L rollup across 200 entities in seconds.
  • Concurrent multi-user editing. Budgeting and forecasting are inherently collaborative. Dozens of budget owners need to input data simultaneously, with version control, locking, and conflict resolution. This is an infrastructure problem, not an intelligence problem.
  • Security and access management. FP&A data is sensitive. Planning tools provide granular role-based access: a department head sees their cost centers, not the full company P&L. Building and maintaining this kind of access control layer from scratch is a significant engineering investment.
  • SOC compliance and audit trails. Enterprise planning tools come with SOC 2 compliance, change logs, and approval workflows out of the box. For public companies especially, these aren't nice-to-haves.
  • Opportunity cost. Even if a company could build its own planning system, the engineering hours required would be enormous. For most companies, including AI companies, that time is better spent on their core product.

What changes is the interface layer. Natural language will increasingly become how people interact with planning models, rather than learning proprietary UIs and formula syntax. The planning engine stays; the interaction paradigm shifts.

The spreadsheet didn't kill the general ledger. AI won't kill the planning tool. It will make the planning tool more accessible, and the teams that build their models with AI-ready outputs (structured, well-labeled, queryable) will get more value from this shift than those who don't.

The Bottom Line

AI in FP&A is real and it's coming fast. But the teams that will capture the most value aren't the ones rushing to deploy the latest agent framework. They're the ones investing in three things most people find unglamorous:

  1. Clean, well-structured data that AI can actually trust
  2. Connectivity standards (like MCP) that let AI reach enterprise systems without bespoke integrations
  3. Meeting analysts in their existing workflows rather than asking them to adopt new surfaces

Get these right, and the exciting AI applications — automated variance analysis, agentic forecasting, natural language planning — become natural extensions of what you've built. Skip them, and every AI project becomes a one-off experiment that doesn't scale.

The foundations aren't the interesting part. But they're the part that determines whether everything else works.