The Vendor Landscape: A Healthy Scepticism Required
The AI transformation narrative in FP&A has been building since 2023. Vendors including Anaplan, Workday Adaptive, Pigment, Mosaic, Cube and a wave of newer AI-native tools all position their products around some combination of automated forecasting, natural language querying, intelligent variance commentary and real-time scenario analysis. The sales pitch is compelling: reduce the time your team spends in spreadsheets, enable non-finance stakeholders to self-serve financial insights, and produce better forecasts through pattern recognition across large data sets.
The honest assessment, two years into the AI FP&A wave, is that the best tools have delivered genuine value in specific use cases, that most implementations have taken longer and cost more than vendors projected, and that the transformative promise — a fully automated forecasting function requiring minimal human involvement — remains significantly ahead of current reality for most organisations.
This is not a counsel against investment in FP&A technology. The tools that work well genuinely save time and improve analytical capability. But the investment decision should be grounded in what these tools can do today, not what vendors say they will be able to do when the AI matures further.
What AI-Powered FP&A Can Actually Do Well
There are genuine capabilities in current-generation AI FP&A tools that are worth paying for and implementing. The key is to be specific about which capabilities you are buying and what problem each one solves.
Pattern Recognition in Large Datasets
This is where AI genuinely outperforms human analysis. A tool that can process three years of daily transactional data, identify seasonal patterns, correlate revenue movements with external variables, and flag anomalies that would take an analyst a week to find is genuinely valuable. The best tools in this category — particularly those with robust integrations to ERP and CRM systems — deliver measurable time savings in the data processing and pattern identification phases of the forecasting cycle.
Automated Variance Commentary
AI-generated variance commentary has improved considerably and is now genuinely useful as a starting point. A tool that can automatically compare actuals to budget, identify the top five drivers of variance, and generate a first draft of the written commentary for the management pack saves a finance analyst several hours per month. The commentary is not publication-ready without human review, but it reduces the time spent on the drafting task from two or three hours to thirty minutes. For teams that produce monthly packs for multiple cost centres or business units, this adds up.
Driver-Based Forecast Updating
In a well-configured driver-based model, AI tools can automate the process of updating the forecast when key input assumptions change. If you update your expected headcount plan, the model automatically recalculates payroll costs, benefit costs, T&E allowances and any other headcount-linked assumptions throughout the model. This is not conceptually new — the same outcome is achievable in a well-built Excel model — but the AI-powered versions typically handle a larger number of interdependencies more reliably and with less risk of formula errors.
What AI-Powered FP&A Still Cannot Do Well
The limitations are as important to understand as the capabilities, because the limitations determine where human judgement remains essential and where you should not rely on AI-generated outputs without careful review.
Judgement Calls and Qualitative Inputs
AI forecasting tools are pattern-recognition engines. They are good at extrapolating historical patterns and identifying statistical relationships in structured data. They are not good at incorporating information that has no historical analogue: a new product launch, a change in go-to-market strategy, a new competitor entering the market, a regulatory change with uncertain impact, or a macroeconomic shift that has not yet appeared in the financial data. These are precisely the inputs that drive the most significant forecast variances, and they all require human judgement.
The risk is that a finance team that has become reliant on AI-generated forecasts anchors its projections on historical patterns rather than forward-looking judgements. A tool that has learned from three years of data during a period of strong growth will tend to project continued growth. A CFO who understands the competitive dynamics, pipeline quality and customer sentiment may know that the outlook is more cautious than the historical data suggests. That judgement must override the model.
Novel Business Model Forecasting
AI forecasting tools perform best on established revenue models with significant historical data. They perform poorly on new products, new markets, or novel commercial structures. If your business is in a period of material transition — launching a new pricing model, entering a new geography, shifting from one-off to recurring revenue — the historical data that feeds the AI is not representative of the future you are trying to forecast. In these circumstances, a well-constructed driver-based model with transparent human assumptions outperforms an AI model that is extrapolating from an unrepresentative base.
"The biggest risk of over-relying on AI forecasting is not that it gets the numbers wrong. It is that it gets them plausibly wrong — producing outputs that look precise and data-driven but are anchored on historical patterns that no longer apply."
The Implementation Reality
The single most common reason FP&A technology implementations fail to deliver their projected ROI is data quality. AI tools require clean, consistent, well-structured data. Most finance functions run on a combination of ERP data, spreadsheet workbooks and manual adjustments that have accumulated over years. Before an AI FP&A tool can do anything useful, the underlying data must be cleaned, standardised and integrated into the new platform. This is typically the longest and most expensive part of any implementation, and it is almost always underestimated.
The integration effort is the second major cost driver. An FP&A tool that cannot connect reliably to the ERP, the CRM, the billing system and the HRIS is only as good as the manual data entry feeding it. The connectors exist for most major platforms, but getting them to work reliably in a real production environment, with real data quality variation, takes significantly longer than the vendor's integration timeline suggests.
Change management is the third frequently underestimated element. Finance teams that have built deep expertise in Excel-based models are often resistant to migrating to a new platform, especially if the early version does not replicate all the functionality they are used to. The implementation programme must invest in training, in demonstrating value quickly through early wins, and in managing the cultural transition from "our model" to "the platform." This requires committed senior sponsorship from the CFO, not delegation to a project manager.
How to Evaluate and Select FP&A Technology
Evaluating FP&A technology correctly is a process that should take three to four months, involve the actual users as well as the CFO, and be grounded in a proof-of-concept with your own data rather than vendor-provided demonstrations.
Key Questions to Ask Vendors
- What is the actual implementation timeline for a business of our size and complexity? Ask for three to five reference customers with similar ERP and data environments, and call them.
- What does the data preparation phase involve, and what does it cost? The implementation fee is not the total cost; data cleaning and integration often cost as much as the licence.
- What happens when the AI produces a forecast we disagree with? Can we override it, and does the override feed back into the model's learning?
- What is your system uptime and disaster recovery provision? A planning platform that goes down during the budget cycle has catastrophic consequences.
- How are the AI model's assumptions surfaced and audited? Can we see why the model is producing a particular forecast?
Red Flags
- A vendor who will not allow a proof-of-concept with your own data before contract signature
- Implementation timelines shorter than six months for a business with more than one ERP source
- AI forecasting capabilities that are described in marketing terms but cannot be demonstrated on your data
- A customer reference list that does not include businesses of comparable size and complexity
- Pricing that is not transparent about the total cost of ownership, including implementation, data migration and ongoing support
The ROI Case for FP&A Technology
The ROI case for a well-implemented FP&A platform is genuine, but it is not primarily driven by AI capabilities. The primary source of ROI is elimination of the spreadsheet model maintenance burden: the hours spent each month reconciling linked spreadsheets, chasing inputs from business unit owners, rebuilding broken formula links after a model update, and managing version control across multiple copies of the budget. A platform that eliminates this overhead and produces a reliable, version-controlled single source of truth for the financial plan typically delivers a measurable time saving within the first financial year.
The AI capabilities — automated commentary, pattern recognition, scenario analysis — add incremental ROI on top of this base. They are typically worth having, but they should not be the primary justification for the investment.
A realistic ROI calculation for a Series B or C fintech with a three-person FP&A function might look like this: two to three hours saved per analyst per month on data assembly; one to two hours saved per analyst per month on variance commentary; four to six hours saved per budget cycle on consolidation and reconciliation. At an all-in cost of £80,000 to £90,000 per analyst, the time savings translate to £25,000 to £40,000 of annual analyst time recovered per year — which typically justifies a platform investment of £40,000 to £80,000 per year at a two-year payback horizon.
Key Takeaways
- AI-powered FP&A tools have genuine capabilities in pattern recognition, automated variance commentary and driver-based forecast updating. These are worth paying for.
- The tools still cannot replicate human judgement, incorporate qualitative forward-looking inputs, or reliably forecast novel business model transitions. These remain CFO responsibilities.
- Most implementations take two to three times longer than vendors project, primarily because of data quality and integration complexity. Budget the time accordingly.
- The primary ROI case for FP&A platforms is elimination of spreadsheet model maintenance burden, not AI capabilities. The AI adds incremental value on top.
- Always run a proof-of-concept on your own data before contracting. Vendor demos on curated data sets are not representative of production performance.
- A two-year payback horizon is a realistic target for a well-implemented platform at a fintech with a three-person FP&A function. Shorter payback projections are typically not credible.
- Change management is as important as the technology itself. CFO sponsorship and active engagement throughout the implementation is essential.
