AI is a tool. And like every tool before it, it will solve some of your problems in a novel way.

Take double-entry accounting, invented in Italy in 1494. It allowed businesses to thrive by building trust in financial systems. Later came Excel in 1985, making analysis and financial modeling accessible. In 1990, ERPs transformed how companies operated by integrating every major business function. Most companies have been using parts of AI for years (excluding invoicing automation which has been around since 1965).

But here’s the truth: none of these tools make strategic decisions,  and AI shouldn’t either.

At the end of the day, it will always be people who make decisions about directions and priorities. And, It is Finance’s job to enable that by providing strong data.

When finance embraced double-entry accounting, ERPs, and Excel, the function grew into the strategic powerhouse it is today. And the teams that embraced these tools early moved faster and achieved success more quickly.

So, paradoxically, AI is both the most underhyped and overhyped technology of our time. Often, this confusion leads to severely mis-matched expectations.

In terms of a framework, AI’s general capabilities are: 

• Great: Automating tasks with clear rules (e.g., data cleanup, consolidation, accounting entries)

• Good: Giving quick answers to questions where data is structured (e.g., analyses, reports, scenario planning, alert systems) thanks to chat-based interfaces

• Not Great: Tackling problems with high uncertainty or requiring persuasion

• Awful: Making strategic decisions

And now, how this applies across FP&A work:

Important:

1. Nearly all of AI’s value in FP&A derives from having all of the information from all your systems in one place, which is critical for nearly any FP&A work. 

2. Your current data stack will define how you use AI. If it’s in Excel, load it into an LLM and then generate a Python file if you want to use it. If it’s across large Google Sheets, connect a big LLM model to your sheets through an API. If it’s in a data warehouse, integrate tools over the platform. If it’s in an FP&A tool, expect most of these features are either present or on the roadmap.

Let’s get to work.

1. Data Consolidation & Consolidation

People have been harping on about data consolidation since the first ERP. Then, consultants began selling the wonderful benefits of integrations. But the sad fact is, forty years later, finance teams still spend a ridiculous amount of time reconciling systems, importing spreadsheets, and fixing data errors.

Goal: All your systems are the source of truth. AI constantly checks across them all to give you perfection.

Ways AI can help now:

1. Cleaning data for Import: For a single import, you can use simple tools to sift through thousands of transactions, identify errors, and suggest fixes across widely different formats. A good system will use AI to validate across all of your systems and uncover underlying issues.

   1. Example: Import all of your CRM, ERP, HRIS, and billing data. Define issues: Customer X is missing in HubSpot, Employee Y isn’t reflected in the HRIS from last June, or 15 companies show no usage but haven’t paid bills and have technically churned.

2. Monitoring issues: AI agents can run on top of each system and across them as an overview layer. They can tell you when, say, a customer has the wrong ID in your billing database.

3. Fixing issues: AI agents can propose and execute fixes, once approved, by matching across several criteria to clean transactions quickly and accurately.

You know you’re doing it well when this is automated:

The CRM logs a $20k ARR drop due to a churn. The system checks billing and ERP to confirm alignment. It checks whether this is within expected churn. If not, it flags a Slack channel with recommended actions, then confirms once the issue is fixed.

Done well, AI checking your data and proposing fixes is like having someone reconciling all your data non-stop. Once the data is in a good place, we can turn to reporting.

Tip: If you’re hesitant, try this: upload a particularly tough reconciliation to a general LLM (Google, OpenAI, etc.) and ask it to analyze.

2. Reporting

Reporting is a balance between showing a single version of the truth (the company view) and slicing reports in different ways to answer specific questions. Traditional reports always face this tension, sometimes solved with drilldowns and filters, which inevitably leads to people pulling the wrong data.

Goal: All slices of reports and decks match effortlessly, past, present, and future. Anyone can ask a question, like “How is my customer, Perplexify, doing in terms of usage and MRR over time?”

Ways AI can help now:

• Chat-based user interface: People are increasingly comfortable asking questions in context. This UI is much friendlier than clicking through dashboards.

• Generate reports: Creating one report isn’t that time-consuming. But what about 20 slight variations? Or reformatting for the board? Or pushing the same metrics to 20 different stakeholders? AI can cut this time from weeks to hours.

• Look beneath the reports: Perhaps the most interesting capability. Reports often struggle with answering “Why did sales increase by 10%?” AI can now help. (E.g., “X new customers bought Y product; marketing conversion rose Z%.”) When done right, these systems spot the actual events without needing to drill-down for hours.

• Automate explanations: Many tools now summarize reports to help those less comfortable with numbers understand what’s going on.

• Take KPIs to the next level. You can track more data, and track more relevant data without needing frequent updates, which actually allows people to use KPIs effectively.

You know you’re doing it well when this is automated:

“Sales conversion this past week is 10% lower than average, even adjusted for seasonality. Root cause: fewer ICP leads routed to the Europe sales team based on two underperforming marketing campaigns” Or an immediate answer to: “Break down margin by region, product, and month for the past two years.”

Summarizing large volumes of data (and let’s be honest, this was mostly done with SUMIFs) is something AI does incredibly well.

Tip: Building too many models upfront can miss the point. First, make sure there’s a central repository where major models are validated, controlled, and created. Otherwise, you risk having your data tell completely different stories.

3. Scenarios

Building on reporting, you’ll eventually want to explore future scenarios, whether it’s budgeting, pricing, or go-to-market strategies. Like reporting, you can generate these scenarios easily with the right tools.

Goal: Answer any What If? question instantly. Track hundreds of decisions live to evaluate their impact.

Ways AI can help now:

• Question, Answer: Pair question-asking with AI-generated reporting, and your CEO can get answers to the kinds of questions that drive real decisions.

• Tweaking many variables: Generative AI can iterate thousands of scenarios across hundreds of variables. It helps you choose the most realistic paths—and shows which assumptions might derail you.

• And much more: Planning and budgeting…

You know you’re doing it well when this is automated:

“Show me a downside scenario where next quarter’s new sales are 20% below plan, churn increases by 2 percentage points, and we roll out the new import feature as an add-on to customers who joined in the last year.”

By saving time on building scenarios, your team can focus on which scenario is most realistic and what targets are truly worth aiming for.

4. Agents/Business Partners

Okay, now you have the data consolidated, the reports set up, and you can run scenarios at a touch. It’s time to let agents act as analysts across the business.

Goal: Everyone on the team stays aware of how the business is performing, with real-time alerts on what needs attention.

Ways AI can help now:

• Monitor rules: Set up AI to monitor rules across systems and send alerts. For example, if reconciliations are off or cash from a key customer is late.

• Work as business partners: Assign one agent per team to field data questions, send tailored reports, and provide a custom dashboard view.

• Receivables partners: Similar to customer service bots, agents can send overdue payment reminders. These can be customized with language based on usage trends, making them more effective.

• An extra example: BCG report

You know you’re doing it well when this is automated:

Morning Report: “Good morning!

• Cash balance is $12.4M (down $0.3M day-over-day, reflecting yesterday’s payroll run).

• ARR: $50.2M, up $200k from new bookings yesterday (2 deals closed in Salesforce).

• Churn: No cancellations recorded yesterday. Current MoM churn is 1.8%, slightly above our 1.5% target.

• Expenses: Marketing spend month-to-date is $1.2M, which is 10% over the run-rate budget. Mainly due to higher “Dreaming of Integrations” campaign.

• Forecast vs Plan: We are tracking 3% below plan on revenue for Q3. At this pace, end-of-quarter revenue is projected to be $X, which is $Y below target.

• Today’s Insight: An anomaly was detected in the Europe sales pipeline – conversion rates dropped to 15% this week vs. 25% average. This might need sales leadership attention.”

• Actions:

  • 2 big payments are now 15 days past payment. You should escalate and check in on the team.

  • Usage is down 10% due to the lack of user uploads across two companies. Follow up with CS.

  • These 15 customers have still never had any significant usage. 

Now, we aren’t there yet. We still need analysts. But, with current trends this is highly likely over the coming year. For now, the best way is to have an AI agent running this across your system with an analyst checking and then distributing to the team.

Tip: Ensure AI agents actually add value. The last thing you want is 50 Clippies running around your financial systems.

5. Memory 

Over time, the right setup helps you tell a consistent story. You’ll be able to quickly look at how reports appeared in the past, track the decisions made, and analyze the outcomes. This can be intensely time-consuming for individuals, but when systematized, that knowledge becomes instantly accessible.

Goal: All current decisions are informed by past decisions and their outcomes without wading through endless tables that no one reads.

Ways AI can help now:

1. Troubleshoot past reports: You can review historical reports and compare them with re-runs using the same logic. This is especially useful in due diligence.

2. Construct a narrative: Since the system understands past data, you can ask it questions like: “How has sales commission changed over the past five years? What results did it drive? How does it compare to today?”

3. Hold teams accountable: Agents can generate performance summaries quickly. For example, they could review the past six months of a sales rep’s activity and produce a data-backed performance review.

You know you’re doing it well when this is automated:

Every decision you make has a clear, updated narrative including previous decisions.

In Conclusion

AI is good at some things, not great at others. Lean into where it is genuinely useful for its ability to interpret words quickly, process thousands of transactions instantly, and interact through text with your team.

Yes, it takes some effort to set up. But like any good system implementation, once it’s in place, the payoff is faster, smarter decisions.