In the past couple of weeks, we’ve had a number of conversations about the pricing of AI accounting software. There appears to be some confusion (and a lot of disagreement) about the costs incurred when building or using AI tools, so here’s our best attempt at clearing it up.

[Read time: 6 minutes]

💰 A short history of software pricing

Last time I checked, I was personally paying for around 8 monthly software subscriptions (and I share as many subscriptions as possible, like Netflix and Spotify, with my family; the average US citizen has 12 monthly subscriptions), and the average company easily has 10 times that. We are a pre-seed startup and already have 24 software vendors, most of which bill us monthly. But it wasn’t always like this.

The Desktop Era 💿

In the early days of personal computing, software pricing was simple. You’d buy a physical copy of the software, usually on a floppy disk or CD, for a one-time fee*. This was the era of boxed software, and once you bought it, it was yours to use indefinitely. Think of classics like Microsoft Office or Adobe Photoshop, where you made a single payment and enjoyed the product until technology moved on or you needed an upgrade.

At the time, this made sense because software would use your own device’s compute power and storage. The same remains true for many video games, which, for the most part, still charge a one-time fee.

However, the desktop era had significant limitations. First, app performance would lag if your computer had low processing power. You also had limited storage. Moreover, if something went wrong with the internal storage, file recovery was a challenge, and if you hadn’t manually backed up your data, it was sometimes impossible.

*In recent years, desktop vendors have started charging monthly or yearly subscriptions, which they enforce via a ‘license key’.

The Cloud Era ☁️

In the mid-2000s, the cloud was born. Compute and storage moved to faster offsite servers with more storage and built-in redundancy/data recovery. Other benefits of cloud computing include access to greater processing power, improved connectivity and collaboration tools (G Suite > MS Office).

However, to provide this better customer experience, software vendors had to pay for servers, which cost money to run indefinitely (unlike code, which only costs money temporarily while engineers are getting paid to write it). As a result, many vendors shifted to a monthly subscription model to cover their costs.

Although it is the topic of many complaints, monthly subscriptions do have benefits over one-time payments: the barrier to entry is often much lower, and it can facilitate a much better product experience. For example, I can pay £12 for Spotify and immediately get access to nearly every song in the world. It would cost a lot more than that to buy all the relevant CDs.

The AI Era 🤖 

We are now entering the third era of software pricing, dictated by AI-powered products. AI tools bring unprecedented power and potential, but they also come with a price tag that can vary wildly. Some companies charge per user, others per usage, and some by the amount of data processed. AI software can range from free (with a catch, like your data being used to train their models) to thousands of dollars per month for enterprise solutions.

Most AI-powered tools do not use their own models, they use OpenAI (famous for ChatGPT), Anthropic (Claude) or Google (Gemini). This is similar to how most SaaS companies do not use their own servers, they use Amazon (AWS) or Microsoft (Azure). In the case of AI, you are not buying server capacity directly. It is one step more removed: you are paying to access an API, and when you call that API, an AI model is being run on a server, and the output is provided back through the same API.

🏷️ How should AI accounting software be priced?

The cost of using AI models as a software vendor 👨‍💻

AI model providers charge per token. A token is a known item to AI vocabulary, analogous to a word in natural language vocabulary. In fact, a single word is often equal to a single token (long words are sometimes split into multiple tokens). Generative AI predicts the next token one at a time when generating output.

In order to get a response from an AI model, you need to provide it with a prompt. This prompt is split into input tokens and the response that is provided by the AI is split into output tokens. Most AI model providers charge for the total number of input and output tokens generated every time the API is called. Gpt-4o (OpenAI’s leading AI model) charges $5.00 / 1M input tokens and $15.00 / 1M output tokens.

In summary, the more data you put into the AI model and the more data it generates in response, the more expensive it is. As we build our AI accounting tool, every time we process a transaction, we use an AI agent architecture which involves detailed input data from multiple sources and many chained prompts, resulting in up to tens of thousands of total tokens. The cost is not prohibitive, but also isn’t trivial.

A real-life example: invoice & receipt capture 🧾

Invoice and receipt capture is a prime example of a task that can become 10 times more efficient with AI. By leveraging multimodal LLMs instead of outdated OCR models, the process can be completed faster, more accurately, and with greater end-to-end automation. Imagine not having to check every receipt to confirm the accuracy of extracted information. Imagine automatically applying the correct account code and VAT rate to every processed invoice. This is the reality we are building.

When considering pricing strategies for an AI-powered invoice capture tool, there are several models to explore.

❌ Per Client:

Many existing OCR tools charge per client, which can lead to frequent price increases and reduced transparency because the model does not align with the cost to provide the service (one client may process 10 invoices per month while another processes 100, but the cost is the same).

This pricing model will also likely restrict the ability of existing software providers to adopt new AI features without raising prices, leading to potential dissatisfaction and trust issues among users.

❌ Per Hour:

Some new AI accounting tools charge per hour of tool runtime. This is impractical for AI services since it can misalign incentives. Providers may benefit from slower processing times, which isn’t in the best interest of users seeking efficiency and speed. The cost is also unpredictable to the end-user: What will one hour of ‘tool runtime’ get them?

Moreover, this model does not reflect the actual cost incurred by the software vendor. AI model providers do not charge per hour but on a usage basis (per number of tokens processed).

❌ Usage-based:

This model charges based on the number of invoices processed per month. Usage-based pricing aligns incentives and correlates with the value provided and the costs incurred by the software vendor, but it’s unpredictable to the end user.

✅ Usage-based with Tiers:

This model charges a predictable monthly fee based on the number of invoices processed that month, which will fall within a specific tier. For example, a starter plan might cover up to 500 invoices per month with a predictable monthly fee, and higher tiers will be available for larger volumes. This approach provides predictable monthly costs and aligns incentives by covering the AI processing costs for the software vendor. It’s transparent and scalable, ensuring that users pay for what they actually use without unexpected price hikes.

Pricing for our upcoming AI accounting tool is available upon request, with a discount for early adopters.

Do you agree? Get in touch (by replying to this newsletter or reaching out on LinkedIn) and let me know how you think AI accounting tools should be priced.

Until next time 🫡

Reuben