Artificial Intelligence (AI) has been tested by most organisations over the last couple of years as a potential tool to improve productivity through automation of repetitive activities, generation of content or improvement of organisational processes.
However, the use of AI in organisations is changing. What began as an experimental productivity layer is now transforming into a much more structural layer, a layer which will be able to influence the cost structure and economic efficiency of organisations as a whole.
This transformation is especially relevant to organisations which are based upon subscription-type business models. The long-term profit of subscription-type businesses often depends on small improvements, compounded over time, in areas such as customer retention, prediction of customer churn, success rates of payment, and operational efficiency. Even small inefficiencies in these areas can lead to the loss of millions of dollars of revenue.
Stay Ahead of the Curve!
Don’t miss out on the latest insights, trends, and analysis in the world of data, technology, and startups. Subscribe to our newsletter and get exclusive content delivered straight to your inbox.
Subscribe Now
Although general-purpose AI systems such as ChatGPT have shown impressive abilities, they were developed to understand the economic and behavioural characteristics of no one industry specifically.
Therefore, it is likely that the next major wave of AI innovation will derive from highly specialised models: systems which are trained to understand the deep operational mechanisms of a particular domain. For organisations using subscription-type business models, the development of such domain-specific intelligence could provide a great deal of cost-efficiency compared to the analytical tools used previously by these organisations to identify the same types of inefficiencies.
Artur Zinnurov, Software Engineer, provides his perspective on why specialised AI models may be the next major area of improvement for efficiency in the subscription economy in this interview.
1. AI has become one of the leading business technologies today. Artur, you’ve worked closely on applying specialised AI to subscription businesses. Could you please guide us through the specific product or system you developed, explaining the problem it addresses and why existing tools were unable to resolve it effectively?
Subscriptions now sit across dozens of tools and teams, and we are also entering a period where AI token usage can materially change spending behaviour. This creates overspending and hidden financial risk for businesses. Unlike one-time purchases, subscription signals are usually scattered across emails, receipts, and fragmented financial systems.
I developed a system named RenewlyAI to identify and oversee these “invisible” financial obligations.
RenewlyAI combines:
- OCR (optical character recognition) and email ingestion (Gmail sync, forwarding)
- AI classification of financial documents
- Detection of subscription signals such as renewals, plans, and billing cycles
- A notification system and smart categorisation via a message broker
- AI token usage summaries
Existing tools failed for several reasons:
- Analytics tools require structured input.
- Banks do not capture the context or intent behind transactions.
- Accounting tools are retrospective.
RenewlyAI builds structure from chaos, which is where most real-world subscription data exists.
2. Most people talk about churn or retention at a high level. What non-obvious inefficiency did you discover in subscription businesses that others typically overlook, and how did you quantify its impact?
There is an interesting discovery I have found: passive churn leakage combined with AI usage token costs.
This refers to users continuing to pay for subscriptions they no longer actively use.
With current AI systems and workflows, such as Claude and OpenClaw, there is a second layer, which I call “invisible usage-based spending,” driven by token consumption.
It could have potential problems in the future, such as the following:
- Delays and aggravation escalate costs.
- Teams don’t map usage, which leads to financial impact
- Spend scales with behaviour
For instance:
- Extensive API LLM calls to compute extensive tasks
- An initial feature using an LLM may appear cheap at the start
- Background tasks or retries silently increase cost
According to our internal research, small inefficiencies in token usage (for example, retries, background jobs, and inefficient prompts) can increase costs by 30-140%.
3. Companies that use a subscription-based business model follow a specific economic model. Based upon your experience, what are some of the most common inefficiencies you find in subscription companies today?
From my experience, the most common inefficiencies in subscription businesses today are the following:
- Unnecessary force of obtaining a free trialб the user is obligated to obtain a free trial, which leads to frustration and not looking at the product at all. Some companies offer a demo app for users to try, but this is rare.
- Passive churn leakage, this is when users forget about services, or there is no connection between usage and cost.
With the rise of AI agents that can write code, more and more SaaS and small businesses rely on API endpoints, which incorporates more management of their accounting and tracking of which services they are paying for. However, this could lead to inefficiency due to the additional time needed to understand which subscriptions they are currently paying for.
My system, RenewlyAI, addresses these inefficiencies by structuring raw invoice data for analysis. For instance, general-purpose models will see an AI tool invoice showing $9 without structural context. My system will see it as properly categorised data – category: AI tool, next renewal: date, and service name.
4. As you mentioned, many subscription companies are relying on analytical and data platforms. What unconventional or innovative approaches did you use when designing your model or system (e.g., training strategy, feedback loops, or real-time adaptation)? Why did these choices matter?
Most analytics and data platforms assume that the data will be structured and normalised. In reality, most data is fragmented. Subscription evidence is scattered across emails, invoices, bank transactions, and dashboards, with no single source of truth. Traditional tools were not designed to reconstruct intent over time from such heterogeneous sources.
There are a couple of approaches I have used:
AI-first ingestion
Unlike most tools that expect users to connect a bank account or manually enter data, our system treats raw unstructured inputs — emails, receipt images, and PDFs — as the primary data source and uses AI to impose structure. In my case it was something like this:
“Receipt”,
“invoiceUrl”: string
This approach allows us to reuse the data for further manipulations and analytics, and eliminates manual data entry. The user is not obligated to connect a bank account, which preserves privacy. This matters because it removes the dependency on structured input, which is the main bottleneck for traditional analytics tools.
Feedback loops
Our system adapts to the user’s behaviour, which helps adjust future predictions. When a user confirms or dismisses a detected subscription, the confidence score is updated and classification thresholds are adjusted accordingly. Over time, this process reduces false positives and makes the system increasingly personalised for the end user. Our target is to achieve ≤5% false positives on a labelled evaluation set.
Event-driven architecture
Major parts of our system are built around events, including an upcoming renewal, a spike in AI usage costs, and new subscription detection. These events trigger particular actions – notification of renewals using the Inngest framework, alerts for unusual spending patterns, and recommendations to cancel or optimise. This matters because, unlike traditional batch processing or scheduled reports, an event-driven approach allows the system to react in real time rather than generating retrospective analysis.
Cross-validation from multiple sources
Rather than relying on a single data channel, our system cross-validates subscription signals across email content, OCR-extracted invoice data, and transaction patterns. By using multiple independent indicators instead of single-source inference, subscriptions are confirmed with greater accuracy. This multi-source verification is what makes the system more reliable than tools that only look at bank feeds.
By combining these system design choices, we are able to operate on real-world messy data and detect inefficiencies before they occur.
This allows the system to switch from retrospective analytics to proactive cost optimisation.
5. The emergence of general-purpose AI models like ChatGPT has greatly altered the way individuals perform tasks and gather information. Why are general-purpose AI models insufficient in solving business problems related to specific industries? Can you share a concrete example of how your system translated insights into measurable cost savings or revenue impact for a subscription business?
There are several reasons why general AI is insufficient. First of all, the context window. General-purpose AI does not maintain persistent state by default, and as a result, it cannot reliably take into account a previous invoice. In other words, general-purpose AI functions as a reporting layer, not as a continuous financial cognition system. It processes one request at a time, but it cannot track evolving financial obligations over weeks or months. Secondly, a generic model can summarise an invoice, but it cannot answer how much token usage will cost in the next month if I use a specific AI feature.
UX is also a main part. A user will lose sight of the data if they operate with a lot of invoice operations. When a user operates with dozens of subscriptions across multiple services, a chat-based general AI interface is simply unable to maintain visibility over that data. The user loses sight of the full picture. What is needed is a structured, persistent system, not a conversation.
For a concrete example, our system integrates with the Gmail API to ingest user emails and uses AI classification to automatically detect invoices and subscription-related emails. Through our internal testing, we found that users typically have at least one forgotten recurring charge per month. By providing automated ingestion and a daily digest feature, RenewlyAI surfaces these charges before the next billing date, giving users a window to cancel or downgrade.
Another example relates to AI token usage. A team using LLM-based features across three tools could silently accumulate token costs that exceed the base subscription fees. Background tasks, retries, and API calls all contribute. Our system maps this invisible spend layer alongside traditional subscriptions, giving the user a complete financial picture rather than a fragmented one.
The measurable impacts include:
- Detection of upcoming charges before they are billed
- Identification and cancellation of recurring payments the user no longer needs
- Discovery of inefficient AI usage patterns that silently inflate costs
6. In the subscription business, would it be possible to have proactive actions apart from data analysis?
Apart from the analytics, we have been testing whether we can provide the user with a temporary email. After some time of development and focusing on this idea, we realised that most solutions nowadays can detect such emails. Since we already have an analytics solution, we have started to test whether users or organisations wish to cancel plans on their behalf. This involves many GDPR rules, one in particular being the right to be forgotten, where, when we receive a request to delete an account from a trial system, we will perform the action. We are currently in the alpha version with a small number of users, where this feature is available. However, the results show that around 12% of people successfully got their account removed based on the records that they provided.
7. A key area of focus for your research is developing specialised AI to solve problems related to subscription businesses. How does your approach to developing domain-specific AI differ from a general-purpose AI model with respect to model training? What type of data signals are critical in this context?
Developing specific AI for subscription businesses requires a fundamentally different approach. General AI models are trained to understand language and perform broad tasks. However, they are not optimised for the specific use case, particularly when working with large amounts of unstructured finance data.
RenewlyAI focuses on combining domain-specific signals, structured representations, and behavioural patterns, not on training larger models.
The most important data signals in this domain are:
- Classification of invoices and receipts, this includes vendor name, plan name, recurring amount, and category (productivity, AI tools, finance, and news).
- Email-level signals: To validate subscriptions, we filter invoice and receipt data from PDFs, images, or HTML email bodies. We use fuzzy matching and regex to extract relevant keywords like “trial ending” or “invoice”.
- Behavioural signals, whether users ignore or confirm alerts and which subscriptions they cancel.
- AI usage signals, cost per API call, background jobs, retries, spikes by team or tool, and estimated usage predictions.
This combination makes the model domain-specific: it learns not just the language of the industry, but also the structure and mechanisms of the subscription economy.
8. Based on your experience, where are most subscription companies misapplying AI today, and how should they rethink their approach?
From my experience, many organisations adopt AI by adding a chatbot or using an LLM for customer support, but they don’t rethink how AI could restructure their cost base or improve mechanics at a systemic level. They can use general-purpose models to summarise support tickets, but these won’t detect particular customer usage pattern signals or upcoming cancellations.
There are several improvements organisations could expect:
Improved user retention insights – Identifying users who are paying but no longer engaging and achieving customisable token efficiency for each specific user.
Faster time-to-insight – Traditional analytical workflows require data cleaning, normalisation, and manual configuration. A domain-specific AI system that understands the structure of subscription data can bypass much of this overhead, delivering actionable insights in real time through event-driven architecture rather than batch reporting.
Cost savings for small and medium businesses – By performing additional accounting analysis, SaaS startups, freelancers, and organisations can expect clearer visibility into their spending and reduce dependency on other productivity or service SaaS tools.
Organisations that invest in understanding their domain deeply enough to build or adopt tailored AI systems will benefit most, rather than relying on general-purpose models that lack the necessary context.
9. Do you think we are headed towards a world where industries will be increasingly dependent on vertical AI models versus general-purpose AI systems?
In my point of view, I believe that we are moving towards a world where vertical, domain-specific AI will play an increasingly central role, but it does not mean general-purpose systems will disappear. We will move towards a hybrid approach.
According to Bessemer Venture Partners, vertical AI companies are already reaching 80% of the average contract value of traditional SaaS, growing at roughly 400% year-over-year while maintaining healthy margins. Although general-purpose models like GPT-4 or Claude are powerful, they are often over-engineered for specific business tasks. As a result, the market values specialised solutions.
Another area to consider is privacy and data-sensitive industries. In sectors like healthcare, reliance on general-purpose AI introduces compliance risks. These organisations are moving towards vertically trained models where they can control the data pipeline.
The most effective AI architectures will likely use an orchestration layer that routes queries to the right model based on complexity and domain requirements.
From my experience building RenewlyAI, the real competitive advantage comes not from using the largest model available but from understanding the domain deeply enough to build structure from unstructured data. General-purpose models cannot do this out of the box, and this is where vertical AI creates the most value.
10. What changes do you see occurring in the economics of subscription companies over the next 5-10 years due to the adoption of AI technology?
I believe that AI will cause significant changes in the way subscription-based businesses generate revenue within the next five to ten years, and these pricing changes will be both at the pricing layer and in the way value is created and recognised by customers.
The first large shift in this space will be an increase in emphasis on building and maintaining trust with customers. With AI integrated into all types of products and services, SaaS companies will begin experimenting with different plan options (i.e., lifetime access to a product/service; more flexible commitment options; etc.) to build a sense of reliability and longevity in what has traditionally been a relatively unstable/complex marketplace.
At the same time, AI agents will begin to take a much more proactive approach on behalf of users. While AI agents have assisted users with completing tasks, they will now begin monitoring users’ spending habits, optimising users’ tool usage, and making autonomous decisions regarding which subscriptions should continue to receive payment, whether those subscriptions should be upgraded, or whether those subscriptions should be cancelled. As a result of these developments, companies will no longer compete primarily for customer attention but will instead also compete for algorithmic approval from the AI agents acting on behalf of customers.
Pricing models for subscription-based businesses will undergo substantial development. Hybrid pricing models that include both subscription-based elements combined with usage-based pricing elements will emerge. Additionally, there will be a movement towards “real-time” pricing models that adjust prices based on usage patterns, value creation, and other factors. While potentially improving the efficiency of pricing models, it will also create additional complexity.
Featured image credit
