05. June 2025
The AI Calculation Problem
At an event hosted by GRID in April 2025, Hjalmar Gislason (Founder and CEO of GRID) and Donald Farmer (Data Strategist at TreeHive Strategy) addressed a growing concern arising from the rapid and widespread integration of large language models (LLMs) into business workflow automation: the AI Calculation Problem.
Understanding the root of the problem
LLMs excel at generating fluent, articulate and confident-sounding responses. But when it comes to producing numerical answers that require calculation, they frequently get the math wrong. This can be demonstrated through a very simple yet compelling example:
Given how these systems are designed, this should not come as too much of a surprise: LLMs are trained to predict plausible sequences of text, not to perform arithmetic. As Donald explained, when LLMs are asked to solve math problems “they reason about numbers rather than calculate them”. He even suggested that this behavior is not necessarily a bug, but rather a fundamental characteristic of how LLMs operate.
Awareness of this shortfall is essential when LLMs are deployed to answer questions where numerical accuracy is critical. Erroneous outputs from LLMs - known as AI hallucinations - can easily be missed by humans, especially when numbers are involved. Models can generate results that are within range of the correct answer, seeming plausible enough to trust. In some cases, the results appear extremely precise, again leading us to think that the result has been calculated rather than determined heuristically. But in industries like finance, engineering, and logistics, “almost right” can mean very wrong. Real-world examples that illustrate just how subtle but costly these mistakes can be were presented in the discussion.
Trade off fluency for accuracy?
At the heart of the issue is the non-deterministic nature of LLMs. They operate probabilistically, predicting the most likely next word or token based on context rather than calculating to provide the objectively correct answer. Donald argued that this inherent flexibility is what makes LLMs so effective at conversation and interpretation. Their openness allows us to instruct them in countless ways, handle ambiguity, and generate human-like responses. These are not qualities we would want to tone down nor compromise on.
In any case, it seems that we don’t have a choice as AI hallucinations appear to be an inherent limitation of LLMs, one that may not be fully solvable in the long term. So how can we get the best of both worlds: accurate, consistent numerical outputs delivered through a natural, conversational interface?
Bridging the gap
Hjalmar and Donald explored several practical solutions to address the challenge. One is to explicitly instruct the LLM to generate and execute code to perform calculations or run logic-based workflows (many LLMs have access to a built-in Python runtime environment).
Another approach is to have LLMs call external trusted tools for the task such as databases or computation engines. In this context, spreadsheets present a particularly effective option as they can host business-approved datasets, logic and a calculation engine. GRID has developed an API that enables LLMs to interact directly with spreadsheet environments to return reliable and fully auditable calculated results from sources that reflect how organizations already work.
Both Hjalmar and Donald concluded that LLMs are best used for what they do best: language. For workflows that require computation and calculation, developers and AI practitioners must ensure the workload is delegated to purpose-built, and crucially, deterministic systems to guarantee accuracy, consistency and reliability. This hybrid approach preserves the flexibility and accessibility of LLMs while delivering the consistency and accuracy that businesses depend on.
News
Updates and announcements
02.12.2024
Bringing spreadsheets into the AI-first era
Introducing GRID's new mission The current wave of AI is arguably the biggest shift in user interfaces since the advent of the GUI. Meanwhile, spreadsheets remain a cornerstone of the business world — resilient, ubiquitous, and indispensable despite repeated predictions of their demise. The fusion of AI and spreadsheets is poised to be big, but it requires a fundamental rethinking — not bringing AI to traditional spreadsheets, but reimagining spreadsheets and their workflows for the AI-first era. With a unique set of cutting-edge spreadsheet technologies, GRID is uniquely positioned — and determined — to lead this transformation. AI-First The AI-first paradigm is defined by three key characteristics: Language-oriented: We interact with computers in our language, not theirs. Task-centric: Work starts with the task at hand, not the hunt for the right software. Agentic: Computers will act on our behalf, even when we’re not there. Spreadsheets Spreadsheets are not just tools; they are foundational to modern business: Ubiquitous: Over 1 billion users worldwide rely on them. Empowering: As the original low-code solution, they enable business users to solve problems independently. The fabric of business: Spreadsheets likely hold more business logic and data than any formal IT system. Bringing them together With the world’s most advanced independent spreadsheet engine — designed for lightning-fast performance and seamless compatibility with Excel and Google Sheets — and a suite of other powerful spreadsheet technologies, GRID is uniquely positioned to redefine the future of spreadsheets in an AI-first world. We’re bridging the gap between AI and spreadsheets, delivering the reliable and verifiable calculations that AI solutions currently lack. Bringing spreadsheets to ChatGPT Today, we’re taking a major step forward by expanding the Alpha testing of our ChatGPT solution and opening registrations for early access. Sign up now to secure your spot in the Alpha and see GRID’s solution in action!
26.08.2020
GRID closes $12M in Series A funding round led by NEA
We’re thrilled to announce that we have closed a $12M Series A funding round led by New Enterprise Associates (NEA), with participation from our existing investors BlueYard Capital, Slack Fund, Acequia Capital and other strategic partners! This funding will enable us to bring GRID to market and power accelerated product development. For more information see our press release. Additional coverage: Tech Crunch: GRID raises $12M Series A to turn spreadsheets into 'visual narratives' SiliconANGLE: Iceland's Grid lands $12M to help workers make their spreadsheets more visual Tech Target: Analytics startup Grid raises $12 million in funding
27.03.2019
GRID closes $3.5M seed funding
We are thrilled to share some great news with you: We just closed $3.5M in seed funding! The investment is led by BlueYard Capital, with participation from strategic investors such as Slack Fund, Acequia Capital and angel investor Charlie Songhurst. We are happy to work with this group, as they add a lot of value to our mission other than their funding. Needless to say, they deeply believe in our mission to empower people to turn any spreadsheet into a beautiful web report, dashboard or interactive application. After our private Alpha launch a few weeks ago, we are now all heads-down again working on product, strategy, network expansion and go-to-market planning. This investment - on top of our $1M angel round in October - fuels current plans well into 2021. It gives us breathing room to focus on building the initial version of our product, take it to market and grow it from there - by delighting our users. We will be adding a few people to our team in the coming weeks and months. This is a fantastic opportunity to join an exciting startup at an inflection point. Take a look at our open positions, and keep an eye on our tweets.