A1. Thanks to AI, it is no longer necessary to forecast a business' finances once a month, as financial forecasting can now be done continuously based on many different variables. When companies have moved away from using "flat" (i.e., static) spreadsheets to using AI-based financial forecasting models, companies have reportedly reduced their financial budget variances by 20% on average. However, the greatest change resulting from AI-based forecasting can be seen in AI's ability to continuously ingest external market signals (for example, changing shipping index values or shifts in regional economies) and correlate those signals with internal ERP system data. Another example of the transformation associated with AI-based financial forecasting is its ability to identify emerging or smaller financial trends ("quiet" trends) that would generally be missed by traditional P&L reviews. A2. One of the biggest challenges when attempting to implement AI-based financial forecasting systems is data fragmentation (i.e., large amounts of important financial data residing in different, often outdated, legacy systems). To overcome data fragmentation in legacy systems, organisations should have one centralised (unified) data validation layer to scrub and normalise the entries from all data entry screens into the AI engine. Another major step to overcoming the data fragmentation issue with AI-based forecasting systems is overcoming the scepticism of many senior executives regarding the "black box" nature of AI systems. By establishing clear explanations and visualisation capabilities of the top drivers behind forecasts, companies that have used AI for financial forecasting have been able to overcome the scepticism of their senior leadership and provide reliable insights into financial forecasts as trusted advisors to CFOs. Implementing AI can be seen less as a replacement of an organisation's financial forecasting capabilities and more as providing organisations a clearer, more consistent way to look at their future finances. By cleaning the data that is being used to drive AI-based forecasts and making the logic used by AI to create those forecasts visible, organisations have taken away the question of whether or not the numbers are valid, and instead focused their discussions on making the best strategic decisions based on those forecasted numbers.
1. AI connects Ads to our Warehouse Inventory. Furniture is very sensitive. Let's say I overspend on ads for a product with a 12-week lead time from the factory. If the demand spikes, I will be out of stock. If I underspend, I am left with a warehouse full of wood that costs me a couple of thousand dollars to store. Before AI, we used to set a monthly budget for Meta and Google and review it every Friday. AI changed this because it looks at our SKU-level in real time. If it gets below 15 units, it automatically turns off the ads and pushes it to another product with more than 200 units. We stopped paying to advertise products we don't have and cannot ship. The chance to run a sentiment analysis using AI improved our forecasting too. Furniture is emotional and unfortunately, trends move fast. Sentiment analysis helps us see where people are complaining. We then reallocate our budget to what people are positively talking about. 2. The main challenge we still struggle with is AI hallucinating growth. Last year, we fed it five years of data. Some data from 2020 and 2021 were anomalies because people were stuck at home due to the pandemic and they did buy furniture. AI took that into consideration without context and predicted a 40% YOY growth. If we blindly budgeted as predicted, we would've drowned in inventory costs. We changed the system to weigh pre-2020 data and macro-economic indicators. Attribution also remains a challenge. AI tells me a customer saw an ad, clicked and bought. In our industry, people will see an oak stand on Instagram. They will visit a site more than five times over a month before buying. AI struggles to map this, so we moved to the Marketing Mix Modeling. We run blackout tests where we turn off ads in one city for a week. If sales in the city drop by say 11% despite what the AI predicted, we know the real value of those ads.
Hi, At Best Interest Financial, we integrate Artificial Intelligence (AI) to go beyond number crunching to understanding the market. Our team used to spend days checking trends to forecast rates, and this important, tedious, and time consuming process often lost track of quick changes. After using new AI tools, such as Zest AI and proprietary machine learning systems, we were able to forecast rates 40% faster and improve the accuracy of predicting clients' loan approvals by almost 25%. There are always challenges to overcome, such as the quality of data. Due to the inconsistent data formatting of the old systems, we needed to build a data normalization layer and model retraining on a quarterly basis. There was also real change management a few loan officers were doubtful. We showed them AI recommendations and how it increased loan closure rates, and this was how we built confidence. AI has not replaced our instincts; it has honed them. If you need, I am happy to provide more information. Best regards, John Donikian, Vice President, Best Interest Financial https://bifmortgage.com/ https://www.linkedin.com/in/johndonikian/ I am John, the vice president of Best Interest Financial in Detroit, Michigan. I am a top-producing home financing loan officer and had nearly a decade of success at one of the nation's largest lenders before joining Best Interest Financial. At Best Interest Financial, we make home financing easier with personalized mortgage solutions from experienced professionals
1. The ways in which AI has transformed your forecasting process (how it has made it easier and more accurate) We have evaluated AI, but we are not yet confident in incorporating it into our forecasting process. We also work with clients going through an AI readiness journey for their financial forecasting processes. While some clients are going through the evaluation process and using some AI-native tools for their FP&A processes, it's still early. 2. The challenges of using AI for budgeting and forecasting (and the working solutions you implemented). It could be things like data quality concerns or change management. The current challenge is that AI agents require even more governance and readiness than traditional software development for them to be commercially viable. Customers are struggling to see ROI from the required calibration effort. They are choosing other solutions on the market that are already commercially available and have been implemented multiple times with other businesses.
AI helps us create more specific financial forecasts on a month-to-month basis. We use it to analyze past performance looking back a few months, and comparing it with the same period in previous years. AI removes bias and focuses on specific data, which saves us tons of time since the data is already there. The biggest challenge with using AI for forecasting is that it doesn't consider external factors that are hard to account for just by looking at data. For example, we had a poor July 2024 because two of our sales reps left that month. By default, it assumed that July-Aug 2025 would be bad months too. So, you have to give it context and a real human being needs to oversee the output.
I lead the data science and AI efforts at a legal services company. We have four sales channels and three primary products, so forecasting and budgeting each of these is critical for growth and customer retention. We use traditional AI to create most of our forecasts. These are machine learning models that analyze historical data for patterns, seasonal changes, and other factors that impact forecasts, then use those insights to predict the future. In addition, generative AI has fundamentally changed how we fine-tune and present these models. It allows us to add context and depth to what we see in raw numerical data. We're now building AI infrastructure to analyze customer feedback daily (call transcripts, survey responses, reviews, etc.) and incorporate it into our forecasts to improve model accuracy. We're also experimenting with incorporating macroeconomic indicators (consumer stress, inflation, etc.) into our models using AI. This helps us fill blind spots that come from relying only on historical internal data. We're testing AI to identify anomalies, highlight trends, and generate curated sales reports for executives, saving analysts hours of manual work. AI doesn't just create and update forecasts—it also sends daily email summaries on what's changed, along with trends and anomalies. This helps analysts investigate issues, adjust forecasts when needed, and focus on higher-level strategy. Challenges: Traditional AI (machine learning) gives us the foundation; generative AI adds nuance. But if the data is bad, we're doomed. As we say in the data world: "Garbage in, garbage out." There's also the risk of overreliance or complacency. All AI models are probabilistic. No matter how good they are, they'll be wrong at times—because of bad data, hallucinations, missing context, or misinterpreted facts. Recently, our forecasts were very off. We didn't notice until it was too late. A postmortem showed the issue: a couple of large client deals from a few years ago were forgotten, and the models assumed that high volume was the new normal. The evaluation metrics looked fine, so we didn't double-check. We only realized the mistake after the fact. Even though AI improves accuracy and helps fill blind spots, it will never be perfect. Human oversight is critical. An analyst hearing about a major issue in a deal from a sales leader is not something easily captured in data or modeled. We're currently testing generative AI to help bridge that gap.
We use AI to track ad performance now, which makes our spending predictable and lets us shift budget to what works. Last fall, our cost-per-acquisition guess for a new launch was only 3% off, a huge improvement over Excel. But our old tools fed in messy data, so we made a dashboard to flag spikes first. Just make sure your data is clean, otherwise you'll second-guess the AI like we did.
We started using AI at CashbackHQ to predict ad spend and rewards, which cut down on the manual work. The real headache was cleaning up our tracking data. We worked with engineers to fix the outliers, and our budget errors dropped almost 10% in nine months. Here's what I learned: getting the team on board is the hardest part. But once they saw the numbers, they came around.
AI helped us stop guessing on our forecasts at Apps Plus, though it wasn't an immediate fix. Once we worked machine learning into our automation, it got much better at spotting subscription churn patterns. That let us put our retention budget exactly where it needed to go. Our data was a problem at first, the early models made bad assumptions until we added stricter validation. I'd recommend that step to anyone starting out.
Running Lakeshore Home Buyer got easier with AI looking at our past deals. Our first automated budgets were a mess since data came from different places and didn't match. We fixed it by getting everyone on the same spreadsheet format. Now we see money problems weeks before they happen and know exactly what each rehab should cost.
Hi, I'm Max Avery, CBDO & Principal at Digital Ascension Group, where we provide advice to both private clients and operating companies in regard to capital planning and growth strategies. The introduction of AI has fundamentally transformed how we predict cash flow and revenue timing. The greatest transformation we have seen with AI is the speed and accuracy with which we can conduct scenario modeling. By using AI-based forecasting tools to pull in historical revenue; pipeline data; and macroeconomic indicators, we can now produce dozens of stress tests within minutes rather than days. For example, we identified lagging indicators in deal cycles that had previously been viewed as linear, which lowered our forecast variance from approximately +-18% to below +-7%. This allowed leadership to slow down hiring earlier and maintain runway without sacrificing growth initiatives. The most difficult challenge we faced was the quality and trustworthiness of the data. Early on, the only way to get accurate forecasts was to provide high-quality data and there was skepticism among teams regarding "black box" numbers. To address this issue, we applied a straightforward, yet critical, solution; we used AI solely for pattern forecasting rather than decision-making; and required every projection produced by AI to present an understanding of the sustainability of the inputs used. In addition, we established standardised data inputs prior to automation. One thing is clear: AI does not replace the need for financial judgement; it enhances financial judgement. The major transformational benefit comes from seeing AI-generated forecasts as dynamic tools and not static spreadsheets. As a result, budgeting becomes proactively managed instead of reactively managed; and this shift in thinking creates the greatest value proposition.
We use AI to track our health programs in the US and Australia, cutting our budgeting time from weeks to just a few days. We now predict expansion costs with less than a 10% error rate, down from over 20%. The tricky part was connecting this to our old financial systems. Some people needed help reading the new dashboards, so it took a lot of hands-on training and consistent check-ins to get everyone comfortable.
1. How has AI changed your forecasting process (how has it made it simpler and more accurate)? To begin with, AI helped us identify areas where we were losing money: 1. We operate in the international market, and often our holidays and weekends do not coincide with those of our clients. AI helped identify these overpayments, and we restructured the entire system. In addition, AI forecasts for the entire year how much compensation for overtime is expected for the company as a whole and for each engineer, depending on the project. 2. One of the company benefits is paying for employees' sports activities. HR collected data on all employee benefits, and AI indicated that only about 50% of employees actually use this benefit; for the rest, we were simply paying money without any benefit. We changed the benefit structure and solved the problem. 3. After AI analyzed our planned salary review system, we prepared a document on a new Salary Review system and working with grades. AI gave us suggestions that we ultimately discussed and used to form a new concept. It describes how grade, results, assessment, and salary decisions are now linked. 2. What problems did you encounter when using AI for budgeting and forecasting (and what solutions did you implement)? These could include, for example, problems with data quality or change management. Indeed, much depends on the completeness and accuracy of the data. Because the result is based on their analysis, small inaccuracies can ultimately show, for example, 16% instead of the actual 7%. In the first stage, we double-checked everything manually, identified bottlenecks, and corrected them. For example, in engineer payments, it wasn't entirely correct to simply take the payment amount; we needed to take into account sick leave, vacations, and unpaid leave days. This required further development, and as a result, we now have an almost perfect AI planner.
Ways AI has made our forecasting more accurate and easier: Integrating Qualitative Factors In contrast to traditional quantitative forecasting models based solely on numbers (i.e. past sales records, revenue trend history, and market research data), we developed predictive models that include both structured and unstructured qualitative input data. Specifically, we used machine learning to develop natural language processing (NLP) algorithms to analyze unstructured qualitative data from two primary sources; i.e. customer support ticket entries and community forums. NLP algorithms analyzed the unstructured qualitative data to discover trends in customer sentiment, topic frequency, and urgency and then translated the qualitative results into quantitative inputs to be processed with other data to predict future demand. For us, this resulted in increased forecast accuracy, which resulted in savings of approximately $275,000 through reduced inventory levels ($180,000) and lower overage of capacity costs ($95,000). Additionally, we have saved considerable time; what would have taken 40 hours of manual sentiment analysis and integration can now be completed in less than 2 hours of computational processing time. Main challenges we've faced in using AI for budgeting and forecasting, and our solutions: Dynamic Market Rendering Models Obsolete Because machine learning-based models are built around historical data patterns, they can be obsolete in markets where fundamentals (e.g. competitor pricing, customer behavior, economic conditions) can change in just a few months. For example, if you train your model with data collected from 6 months ago and the pattern it was trained upon changes significantly since then (a condition called "model drift"), your model will likely produce less and less accurate predictions until eventually producing very poor predictions. So we transitioned from an ad-hoc model maintenance process to a disciplined and schedule-driven process for maintaining our models to prevent model drift. Process: We conduct a quarterly review of all models used in the company. Action 1: We perform an assumption review. We assemble a cross-functional team (data scientists, finance and product managers) who manually evaluate the model's top assumptions (e.g. "Customer sentiment is the second strongest driver of demand") compared to recent real world events and trends in the new data.
Hi Hubspot Team, Ryan McCallister here from F5 Mortgage. I'm a Certified Mortgage Advisor with 12 years of experience dealing with mortgage strategy and financial forecasting. Accurate forecasting in mortgage finance is super important because it directly impacts when clients lock rates and how much they will pay over 30 years, so getting predictions right can mean the difference between a 6.5% rate and a 6.8% rate. That is why the change from manual tracking to AI-based forecasting has been massive for us at F5 Mortgage. Before using AI in this process, we were spending every Monday morning pulling up rate data from different sources and building out spreadsheets to identify trends. We monitored the 10-year treasury rate, Fed announcements and lender pricing sheets, then tried to build a prognosis of where rates were headed based on rate patterns we had seen in the past. It worked sometimes, but we were always playing catch up, at least a week or two behind the market. Now we have machine learning algorithms available to us in platforms such as Optimal Blue that scan thousands of data points at once and detect shifts in patterns 48 hours before they occur. The system takes real-time information from economic indicators and past movements in rates and alerts us when the conditions are similar to those that caused rates to change in the past. AI forecasting models have accuracy rates of approximately 91.5% as compared to 84.8% with traditional ones. That window of early warning saves thousands of those clients because we're able to do a better job of timing rate locks. Last month alone, three clients were able to lock rates two days prior to a jump because the model picked up Fed signaling patterns that we would've missed manually. That difference amounted to $340 less per month for one borrower, which amounts to more than $122,000 over a 30-year loan. If you do have questions, let me know. I attached my email below. Best, Ryan McCallister President & Founder at F5 Mortgage Web: www.f5mortgage.com Email: Ryan.M@f5mortgage.com Headshot: bit.ly/3ZZPspr Address: 1283 Farmington Dr. Traverse City, MI 49696, NMLS# 1938115 Author profile: https://f5mortgage. com/ryan-mccallister/
Hi there, I own a mortgage brokerage and have been using AI to help improve our cash flow forecasting. In the mortgage industry, although we charge a small fee directly to our clients, the majority of our income comes from the bank through which we arrange a mortgage. The difficulty this causes for us is that it can often take 3-6 months for a house purchase to go through - particularly if there is a property chain where multiple sales need to happen simultaneously- and even then, the banks can take further weeks or even months to pay out our procurement fee. So in some cases, we could onboard a client in January, do 90% of the work in February, and then not see payment until November / December, for example. This makes it really difficult to forecast where we'll be financially in 3-6 months, and it always feels like we're playing catch up - particularly with new hires and balancing increased workload from growth with significantly delayed payments. What's been helpful here is letting AI analyse our historic and current pipelines and payment timescales, learning the trends, and then modelling 3-6-12-month scenarios based on our current trajectory, and this has massively helped us plan the other aspects of the business more effectively around our income. So for us, it looks at average completion with each lender, each lender's typical payment times, average completion by property type and location, seasonal fluctuations, and, as we work with many international clients, even trends among which countries complete faster then others, etc. Many of these things we have a gut instinct for already, but there's no way we can really apply it to our data in a way that feels reliable - the results we get from AI are by no means perfect, but they really help give us a sounding board for what our finances are going to look like in 3-6 months which we can then apply more confidently to other areas of the business.
Due to AI, forecasting is now ongoing and contains many forecasts, not quarterly. Instead of making one forecast and maintaining it for three months, we may generate an ongoing flow of expected outcomes based on input. We can now respond to signs early in the process rather than explaining why we didn't after they occur, improving our forecast accuracy by 15-20%. While AI models worked well in terms of functionality, there were concerns with data quality and trust. When we first started utilizing AI-based algorithms, they accelerated initial incorrect predictions faster than spreadsheet-based models. To fix this, we limited the model's inputs to metrics we trusted and gave users the opportunity to see the model's prediction or a human decision. The greatest benefit of AI has been increased speed (i.e., leadership now has days to make decisions instead of weeks to resolve variations in numbers during budget revisions), but I believe that AI models will eventually tighten the feedback loop around human judgment rather than replace it.
In electrical contracting, using general inflation data (CPI) does not meet our budgeting needs, as we need to have the exact inflation rates for copper, steel, and PVC in hand. We now use an AI tool to scrape every incoming vendor invoice so we can track the prices of individual SKUs, allowing us to forecast our actual project margins based on material price volatility rather than using average quarterly prices. On a recent large hospital project, the AI identified a 12% trend in PVC conduit prices three weeks before our major suppliers announced price increases. By purchasing some of the bulk inventory at the lower price, we saved $28,000. However, the AI also helps identify potential problems in the supply chain that can affect our ability to use the information from the AI to create accurate forecasts. For example, we noticed that vendors were starting to ship only 80% of our orders (and the remaining 20 percent would be shipped later), usually just prior to a significant price increase. When the AI sees a jump in the rate of incomplete shipments, it associates it with probable price increases, so we're aggressively purchasing all we can to lock in a better price. One of the biggest challenges with getting quality data for the AI to work properly - maintaining good data hygiene. One vendor may bill a product as "1/2 inch EMT", while another vendor bills the same product as "Conduit, EMT, 0.5". The AI initially treated these as two separate products, which affected our forecasted numbers. To resolve this problem, we have added an additional step to our workflow to manually instruct the AI to recognize the different vendor descriptions for the same product and to normalize them to a single internal part number to keep our data clean.
The AI enabled us to go beyond basic cost-per-click (CPC) metrics to predict the exact cost-per-booked job, based on factors such as the day of the week and weather. Surprisingly, the AI showed that leads generated for longer distance moves on Tuesday have a 30% lower conversion rate than those generated on Friday - with an equal cost per click. Therefore, by lowering our bids on these low-converting days, we reduced ad spend waste by approximately $1500 each month - without losing a single lead. Because the weather significantly affects our business, we are now including long-range weather forecasting in our budget model. In late January, for example, the AI automatically predicted a reduction in booking intent (searching during boredom, booking when intent is higher), and suggested cutting our ad budget by 60% for those three days due to the forecasted heavy snows. That resulted in saving around $2000 in window-shopping clicks and reallocating that money to the sunny days later that week, when booking intent was high. However, there are challenges associated with using an AI model. One of the main challenges is that the AI can be overly sensitive or "jittery" to short-term data fluctuations, which may result in large swings in the budget recommendations. Additionally, when ad platforms such as Google or Meta make changes to their algorithms, the predictive value of historical data is impacted, so we changed from quarterly forecasting to a 14-day rolling forecast. The shorter time frame allows the AI to respond more quickly to the changing conditions of the current market. The system isn't without fault completely. The AI once tried to cut our budget to zero because of a predicted thunderstorm that turned out to be light rain. So, we created floor and ceiling rules. Now, no matter what the AI predicts, it cannot adjust the budget by more than 25% in a single day without human approval. This prevents the algorithm from accidentally shutting off our pipeline due to bad data input.
Turning Forecasting From a Quarterly Guess Into a Living Model AI is most useful to organisations when they can use it to improve speed in completing tasks without sacrificing quality or accuracy. Organisations will be able to make their forecasts more accurate because AI-driven forecasting analyses real-time inputs, seasonality and historical performance. The ability to update a forecast continuously (throughout the month) as opposed to quarterly enables organisations to create tighter budgets, correct courses quickly, and minimise the number of unexpected results at the end of each forecasting period. Organisations that utilise clean data and combine it with human oversight to drive AI can derive the largest ROI. As long as an organisation has invested the necessary time to standardise input variables into an AI system and review output variables generated by an AI system, then the AI system will become a consistent, dependable decision support tool and not a "black box". Organisations using AI systems in this manner will be able to increase their confidence in the accuracy of the forecasts generated by the AI system and maintain ultimate responsibility for all budgetary decisions related to the forecast.