FP&A copilot: driver-based planning, scenarios & variance commentary

Controlling teams spend the planning cycle and every month-end collecting numbers, rebuilding spreadsheets and writing the same "why did we miss" paragraphs. An FP&A copilot sits on top of the planning model (OLAP cube / driver model): it proposes budget and forecast drafts, answers "what if prices rise 3% and hiring slips a quarter?" in plain language, and explains variances with ranked drivers plus a first draft of the commentary. The calculation stays deterministic in the planning engine; the AI orchestrates, explains and drafts.

Problem & value

• Today: budgets and rolling forecasts are built in Excel or in planning tools (Anaplan, SAP, Jedox, Pigment and others) through slow top-down/bottom-up rounds. Variance analysis is manual drill-down across cost centres, and commentary is typed by controllers under deadline pressure.
• AI value: shorter planning cycles, more scenarios per cycle, consistent and faster commentary, and business partners who can query the model without knowing its structure. Controllers move from collecting to challenging.
• Evidence level: vendors publish case studies and agent launches (Pigment, Anaplan, SAP, Jedox), but independent, quantified benefit studies are scarce. Treat time-saving claims as vendor claims and measure your own baseline [1][2][3].

Variants

• SMB / Mittelstand: Excel or Power BI model plus an LLM assistant, or a lightweight planning tool (e.g. Jedox, Pigment) with built-in AI.
• Enterprise: connected planning across finance, HR, sales and supply chain in a multidimensional platform with governed agents.
• Departmental: cost-centre owner self-service ("why is my travel budget over?").
• Public sector / non-profit: budget execution reports with explanatory notes for councils and boards (DATA-002 for reporting layouts).
• Group controlling: multi-entity variance commentary with currency and consolidation effects.

How it works

1. actuals (ERP/GL) + planning cube + driver model + prior commentary
2. semantic layer (dimensions, accounts, hierarchies, driver definitions, business rules)
3. tools: query cube, run driver calc, run scenario, statistical baseline forecast
4. agent (LLM + reasoning): plan the analysis, call tools, never do arithmetic itself
5. outputs: forecast draft | scenario table | variance bridge | commentary draft
6. controller review & edit (human-in-the-loop)
7. approval
8. write-back to a draft version
9. publication to management report (DATA-002)

• Baseline forecasting: statistical or foundation time-series models (ETS/ARIMA, gradient boosting, TimesFM, Chronos, Nixtla) produce a driver-level baseline; the LLM does not forecast numbers itself.
• Scenario modelling: the LLM translates a request into parameter changes (driver, member, period) and calls the planning engine; results come from the cube, not from model text.
• Variance explanation: decompose the gap (price/volume/mix/FX/timing, headcount vs. rate) with deterministic calculations, then let the LLM rank drivers, pull context from contracts, tickets and notes, and phrase the narrative.
• Human-in-the-loop: every write-back goes to a draft version; owners approve; commentary is edited before release; a log stores prompt, tool calls, numbers and reviewer.

Data preparation

Data come from the ERP and general ledger (actuals by account, cost centre, entity, period), the planning cube or Excel models, HR plans (headcount, salary bands), price lists, contracts and earlier management commentary. In an SMB much of this lives in the ERP or DATEV exports, in Excel files with manual links, and in mail threads. Typical gaps: inconsistent account and cost centre hierarchies, plan and actuals in different structures, undocumented driver logic, hard-coded numbers in spreadsheets, and commentary only as Word files.

Preparation steps: (1) export actuals cleanly and automate the extract (PREP-005); (2) turn Excel plans into tidy tables (PREP-011); (3) write a glossary of KPIs, accounts and driver definitions as the semantic layer (PREP-012, PREP-015); (4) align hierarchies between plan and actuals; (5) collect past commentary as gold examples (PREP-016); (6) clean up permissions, especially salary data (PREP-023, PREP-024). Tools: Power Query, DuckDB or the planning platform's connectors.

Effort share is high, roughly 50-60 percent: the AI layer is thin if the semantic layer and the data quality are right. Check with GUIDE-003; see also DATA-001 and DATA-020.

Implementation paths

Which path for whom

• SMB (A): planning platforms with built-in AI (Jedox, Pigment, Cube, Vena) or Copilot in Excel respect existing roles and versions, and controlling is a standard process without own development capacity.
• Enterprise (B): planning logic, semantic layer and approval flows are core and specific, so a tool-calling agent on cloud APIs over the existing planning engine gives more control; finance data justifies EU region and strict logging.
• Developer (B): build the agent and commentary workflow with tool calling against the planning engine or SQL; self-hosting (C) needs HW3 and rarely pays off.

Path A — Buy (recommended for SMBs)

• Planning platforms with AI: Pigment (FR; analyst/modeller agents and an MCP server), Anaplan (CoPlanner and AI agents), SAP Analytics Cloud planning with Joule, Workday Adaptive Planning, Oracle EPM, Board (CH), Jedox (DE; AI features in its planning platform), Planful, Vena, Cube, Mosaic.
• Spreadsheet-centric: Microsoft Copilot in Excel and Google Gemini in Sheets for small models; Cube/Vena for Excel-based planning.
• Choose A when: a planning platform already exists or is due for replacement; vendor AI respects the existing security model (roles, cell-level access, versions) [1][2][3][4].

Path B — Build on cloud APIs

• Frontier LLMs with tool calling (Claude, GPT, Gemini, Mistral — EU region or EU provider for processing finance data) plus MCP connectors to the planning engine (Pigment's MCP server is one example) or SQL/MDX/REST APIs.
• Read-only tools by default; write tools only to draft versions. Use code execution for arithmetic and bridges, structured outputs for commentary.
• Orchestration: LangGraph, OpenAI Agents SDK, Microsoft Agent Framework, or plain function calling; data in Snowflake/BigQuery/Databricks/Postgres.
• Choose B when: the planning stack is custom, Excel-heavy or multi-tool and you want controlled prompts and logging.

Path C — Self-host (open source)

• Open-weight reasoning/tool-calling models (Qwen3, Mistral Small/Large open releases, gpt-oss, Llama) via vLLM or Ollama; time series with TimesFM or Chronos, StatsForecast, Prophet.
• Query layer over Postgres/ClickHouse/DuckDB with a semantic layer (dbt MetricFlow, Cube) so the model sees governed metrics instead of raw tables.
• Hardware: HW3 for strong reasoning models on sensitive finance data; HW2 works for narrow commentary tasks with a smaller model. Ops effort is moderate.

Path D — Custom training

Rarely justified. Possible: fine-tuning a small model on the company's historical commentary for house style, or training a tabular/time-series model on driver history (needs several years of consistent monthly data). Start with prompting plus retrieval of approved past commentary.

State of the art (2026)

• Agentic planning: vendors moved from chat assistants to role-based agents that build models, analyse variances and flag deviations inside the platform, and expose their data to external assistants via MCP [1][2][3].
• Pattern that works: LLM as planner/explainer, deterministic engine as calculator. Text-to-query over well-defined semantic models is far more reliable than over raw schemas.
• Time-series foundation models (TimesFM, Chronos) give decent zero-shot baselines for volumes and costs; they rarely beat tuned driver logic for strategic items and cannot know about planned price changes or contracts.
• Limits and failure modes: wrong aggregation level or hierarchy, double-counted intercompany or FX effects, plausible but unsupported causal stories ("due to weak demand") not backed by data, silent use of stale actuals, and prompt injection via free-text fields in source systems. Commentary must cite the figures and sources it used.
• Evidence gap: few independent benchmarks for finance-planning agents exist; run an internal evaluation set of past variances with known explanations.

Shared resources & combinations

• Semantic layer and text-to-query shared with DATA-001 (conversational BI) and DATA-006 (spreadsheet assistant).
• Narrative generation shared with DATA-002; same review workflow.
• Forecast engines shared with PRED-001 (demand), PRED-002 (sales) and PRED-016 (cash); OPT-008 supplies S&OP volumes as plan drivers.
• Downstream/adjacent: AGENT-006 (reconciliation and close), DATA-020 (regulatory reporting), DATA-012 (ESG plan figures).
• One reasoning-LLM server at HW3 can serve all of these.

Effort, cost & KPIs

• Path A: weeks to a few months for AI features on an existing platform; licences are typically per user or per model size, AI add-ons priced separately by vendor.
• Path B: 2-4 months for a read-only analyst plus commentary drafting; write-back and scenarios add effort. Cost drivers: integration, semantic-layer work, evaluation, tokens.
• Path C: 3-6 months; GPU server plus data/ML engineering.
• KPIs: planning cycle time, close-to-commentary lead time, scenarios per cycle, share of commentary accepted with minor edits, forecast accuracy (WAPE/bias) vs. baseline, number of numeric errors found in review, user adoption by budget owners.

People & change

Involve management and the CFO as sponsors, because plans, forecasts and wording of commentary carry responsibility. The department lead (controlling) owns scope; budget owners and controllers are the end users who review drafts. IT and IT security handle ERP connections, roles and cell-level access, data owners release the ledger and HR plan data.

Works council: depends. Personnel plan and cost centre data can touch individual employees, and Copilot-like assistants log user activity, so the system can be suited to monitor behaviour or performance (§87(1) no. 6 BetrVG, objective suitability suffices). Inform under §90(1) no. 3; if the topic is complex, an expert is necessary under §80(3). Salary and headcount detail needs a DPO review, AVV (Art. 28 GDPR) with the vendor and, depending on granularity, a DPIA.

AI literacy: train controllers to verify numbers against the cube and to treat the LLM as drafting aid, not calculator (GUIDE-006). Pilot with one cost-centre group for a monthly close, compare commentary time and numeric errors, and keep approval and write-back to draft versions with humans. Concern "AI replaces controllers": it removes drafting routine, not judgement. See GUIDE-008, GUIDE-009 and GUIDE-004.

Risks, regulation & compliance

• EU AI Act: minimal risk — internal financial analytics is not an Annex III use. AI literacy (Art. 4) applies. It becomes transparency-relevant if AI-generated text is published externally; if outputs feed credit scoring or employee evaluation the classification changes [5].
• GDPR: cost-centre data often includes salaries and named employees; use aggregates where possible, role-based access, DPAs with vendors and EU processing.
• Financial governance: listed companies and ICS under SOX-style or HGB/IDW controls need change logs, segregation of duties and version control; AI-edited numbers must stay traceable. External reporting (CSRD, ESEF) stays under DATA-012/DATA-020 controls.
• Confidentiality: unpublished forecasts are inside information for listed firms — restrict model access and logging.
• Mitigations: draft-only write-back, tool-computed numbers, cited sources in commentary, owner sign-off, evaluation set, prompt-injection filtering of free-text inputs.

Getting started

1. Pick one process (monthly variance commentary for one business unit) and baseline the hours and quality.
2. Clean and document the hierarchy, account mapping and driver definitions; this is the semantic layer.
3. Pilot read-only: variance bridges plus commentary draft, reviewed by the controller; keep an error log.
4. Add scenario execution to draft versions once numbers reconcile with the cube in 100% of tests.
5. Roll out to budget owners with usage limits, access rules and a monthly accuracy review.

Sources

1. Pigment — AI agents and FP&A software overview — https://www.pigment.com/blog/best-fpa-software
2. Anaplan — AI and CoPlanner product pages — https://www.anaplan.com/
3. SAP — SAP Analytics Cloud planning — https://www.sap.com/products/data-cloud/cloud-analytics/planning.html
4. Jedox — AI in planning and analytics — https://www.jedox.com/en/
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Google Research — TimesFM time-series foundation model — https://github.com/google-research/timesfm
7. Amazon Science — Chronos forecasting models — https://github.com/amazon-science/chronos-forecasting
8. Model Context Protocol specification — https://modelcontextprotocol.io/