The short answer
With traditional software, a one-off purchase made sense because the product was finished once delivered. An AI agent is never finished: vendors retire language models after roughly 12-18 months, document formats and processes change, and quality drifts if nobody measures it. So the real question is not “buy or rent” but who is responsible for the agent still working in eighteen months, and what that costs under each model. Buying wins when you have an in-house team that can maintain the agent. A subscription wins when you do not.
Why an AI agent is not “bought and done”
Companies carry habits from buying traditional software into AI agent projects: specification, quote, acceptance, invoice. That worked when an accounting system deployed in 2015 still did exactly the same job in 2020. An AI agent depends on four things that change without your decision:
- The language model. The agent is built on a specific model version (OpenAI, Anthropic, Google). That version has an expiry date.
- The input data. A supplier changes its invoice layout, an authority changes a form, a partner starts sending scans instead of PDFs.
- The process. An approval threshold changes, a new department joins, the ERP is replaced.
- The quality. The success rate measured on acceptance day is not constant. Without monitoring, the drop is noticed only after complaints.
Every one of these changes requires technical work. The only question is who does it, and on what terms.
A technical fact rarely mentioned at signing: models get retired
Model providers publish retirement policies and stick to them. The specific model version an agent was built on stops working after a while, and requests to it return an error.
| Provider | Retirement policy | Example |
|---|---|---|
| Anthropic (Claude) | min. 60 days notice before a publicly available model is retired | Claude 3.5 Sonnet (22 Oct 2024 version): retirement announced 13 Aug 2025, model shut down 28 Oct 2025, about 12 months after launch |
| OpenAI | min. 6 months notice for generally available models, less for preview versions | Successive GPT-4o and o-series generations retired during 2026 |
| Google (Vertex AI) | min. 12 months availability of a stable model from launch | Gemini 2.5 Pro and 2.5 Flash (stable since mid-2025): retirement from Vertex AI scheduled for October 2026 |
| Amazon Bedrock | min. 12 months availability, then at least 6 months in “Legacy” status | Claude Sonnet 4: shut down on the Anthropic API on 15 Jun 2026, end of life on Bedrock on 14 Oct 2026 |
Planning rule: an AI agent deployed today will go through at least one model migration within its first 18 months.
Sources: Anthropic · OpenAI · Google Cloud · AWS. Retirement dates may change; status as of 30 September 2026.
A simple planning rule follows: an agent deployed today will go through at least one model migration within its first 18 months. A migration is not a one-line configuration change. A new model interprets the same instructions differently, so the agent has to be tested on real data, its prompts and rules adjusted, and often the thresholds at which a case goes to a human recalibrated.
A company that bought a one-off deployment without a plan for this finds out when the agent stops working. It then negotiates a new contract from a position of necessity, under time pressure, with a process that has come to a halt.
Four pricing models for an AI agent
| Criterion | 1One-off purchase | 2Purchase + service contract | 3Subscription (setup fee + monthly fee) | 4Pay per transaction |
|---|---|---|---|---|
| What you buy | The deployment, the code and the documentation | The deployment plus a pool of support hours or an SLA | The deployment plus the running of the agent: hosting, monitoring, updates, model migrations, development | A processed case, document or conversation |
| Who is responsible for it working in 18 months | You | Depends on the scope of the service contract | The implementation partner | The platform vendor |
| Ownership | Usually full, on your side | Usually on your side | To be agreed in the contract: configuration, prompts, data, exit terms | None, you use a product |
| Cost in year 1 | Highest, one-off | High one-off + fixed fee | Lower entry barrier, spread over time | Low, grows with volume |
| Cost in year 2 | Unpredictable: migration, changes, failures | Predictable, if the contract covers migrations | Predictable | Predictable per unit, hard to control as you scale |
| Fit to your process | Full | Full | Full | Limited to what the product offers |
| Main risk | An agent with nobody looking after it | The service contract covers “fixes”, not migrations and development | Dependence on the partner without clear exit terms | Cost grows faster than value, no control over the logic |
Highlighted in blue: the two criteria that most often decide the choice. Model 1-4 columns are numbered as in the article. For a way to put numbers on the other side of the equation, see the AI automation ROI blueprint.
Indicative price ranges for each model are a separate topic. Here we focus on what decides the choice; if you are comparing offers, the AI automation ROI blueprint shows how to put numbers on the other side of the equation.
Why the maths changes in year two
In year one the comparison looks simple: a one-off purchase costs more upfront, a subscription costs less upfront and more in total after a few years. Most spreadsheets stop there.
Year two adds items that were in no offer:
- migration to a new model (see the provider table above),
- adapting to changes in documents and source systems,
- quality monitoring: regularly checking on a sample of cases that the agent still meets the threshold set at proof-of-concept stage,
- development: new case types, a new department, a new channel,
- model cost, which grows with volume whatever the pricing model.
Under a one-off purchase these items land in your budget as unplanned spending, or do not appear at all, and the agent’s quality quietly declines. Under a subscription they are part of the fee. So when comparing offers, put the total cost over 36 months next to this list of five items, not the deployment price next to the monthly fee.
When a one-off purchase is the better choice
A subscription is not the answer for everyone. Buying makes more sense when at least three of these conditions hold:
- You have an in-house team that can maintain the agent: someone who understands language models, can run a migration and can measure quality. Not just someone who “knows IT”.
- The process is stable: the same documents, the same rules, few exceptions, no planned system changes in the next two years.
- Volume is low and the cost of an error is small: the agent supports an internal task and its temporary unavailability stops nothing.
- You require full ownership and control: regulation, security policy or audit requires all code, configuration and infrastructure to sit exclusively on your side.
- The agent is mostly deterministic: the language model handles a small part and the rest is classic integration and rules, so a model migration touches only a small share of the system.
If you choose to buy, put a separate line in the budget for a model migration within 12-18 months and name a person responsible for quality monitoring from day one.
When a subscription makes more sense
- You have no in-house skills to maintain an agent and do not plan to build them.
- The agent runs a critical process: downtime means stalled invoices, requests or quotes.
- The process will keep evolving: new document types, new departments, new integrations.
- You want a predictable operating cost instead of a one-off investment and unplanned invoices in year two.
- You want your partner to care how the agent performs after acceptance, not only on acceptance day.
It is also worth knowing that most AI agent projects stall not at the build stage but at the move to production and maintenance. More on that in why AI agent projects never reach production.
7 questions to ask before signing, whatever the model
- Which language model, and which version, will the agent be built on?
- Who carries out and pays for the migration when the provider retires that version?
- How is quality measured after acceptance, how often, and who receives the report?
- What happens when document formats or a source system change: is that covered by the contract or a new order?
- Who owns the configuration, prompts, training data and logs?
- What are the exit terms: what do you receive, in what format and how quickly?
- Is the language model cost included in the fee or billed separately by usage?
An offer that does not answer questions 2 and 6 directly makes a fair comparison of models impossible.
How we do it
At ITSharkz we use a setup fee plus monthly subscription model, for the reasons above: an AI agent needs care for its whole working life, and we want a stake in it working as well in eighteen months as on acceptance day. This article is meant to help you judge whether that is the right model for your company, including when the answer is “no”. More about how we build AI agents.
FAQ
Is it better to buy or rent an AI agent? It depends on who will maintain the agent after deployment. If you have an in-house team with language model skills, a stable process and a need for full ownership, buying is reasonable. If you have no such team and the process is critical or will change, a subscription gives a predictable cost and puts responsibility for the agent working on the partner.
How often do providers retire language models? In practice, specific model versions are retired after about 12-18 months. Anthropic gives at least 60 days’ notice, OpenAI at least 6 months for generally available models, and Google guarantees stable models on Vertex AI at least 12 months of availability.
What happens to an AI agent when its model is retired? Requests to the retired model version stop working. The agent has to be moved to a newer model, tested on real data and recalibrated, because a new model interprets the same instructions differently.
How much does it cost to maintain an AI agent? Maintenance is made up of the model cost (which grows with volume), hosting and monitoring, model migrations, adapting to changes in documents and systems, and development.
Am I locked in with the provider under a subscription model? You can be, if the contract does not define exit terms. Before signing, agree who owns the configuration, prompts, data and logs, and what you receive, in what format, when the collaboration ends.
How do I compare a purchase offer with a subscription offer? Compare the total cost over 36 months, not the deployment price against the monthly fee. Add to the purchase offer the model migration, quality monitoring, adaptations, development and model cost, because these items will appear in year two either way.
Sources
- Anthropic, Model deprecations: https://docs.anthropic.com/en/docs/about-claude/model-deprecations
- OpenAI, Deprecations: https://platform.openai.com/docs/deprecations
- Google Cloud, Model versions and lifecycle: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions
- AWS, Amazon Bedrock model lifecycle: https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html
- Gartner, forecast on cancellation of agentic AI projects by 2027: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
Not sure which model fits your process? Let’s talk it through on a concrete example. → Book a call with Kacper