Maintaining custom software and AI, what it covers and how you pay for it
The items in a year of maintenance, the ones an AI assistant adds, what to write in the contract and what it costs, estimated on the work and not as a percentage of the price.
What does maintenance of custom software cover?
The work that keeps the software secure, running and the same as when you accepted it. A management system nobody touches does not stay as it is: the libraries it rests on receive security fixes, the services it connects to change their interfaces, the data grows, and the people who use it find cases that were not there during acceptance. Maintenance is the set of these items, written one by one in the contract.
| Item | What it covers | What happens without it |
|---|---|---|
| Fixes | the defects that show up with use, after the month of warranty | they stay, and the people who use the software work around them |
| Updates and security | libraries, frameworks, operating system and certificates brought to the patched versions | a known vulnerability stays open |
| Connected services | the changes needed when e-invoicing, payments, couriers or the accounting system change | the connection breaks because of a change made by someone else |
| Backup and restore | the copies of the data and a restore test | you find out the copy is useless on the day you need it |
| Monitoring | alerts on errors, slowness and disk space running out | you notice when a customer calls |
| Support | answers to the people who use the software and the small changes you expect to ask for | every question becomes a request to be quoted |
New features stay out: one more report, a module, a connection to a system that does not exist today. That is evolutive maintenance, and it is quoted separately like a small project. The line between a fix and a new feature is written before you sign, because without it that is where the arguments start.
What changes if the software has an AI assistant?
Three items are added, because one part of the software is controlled neither by whoever wrote it nor by whoever uses it: the model. The model provider retires it after a while, and from the retirement date requests to the old model fail, so the assistant stops answering. Anthropic says it gives at least 60 days' notice for publicly released models, OpenAI at least six months for generally available models; Claude Sonnet 4, from May 2025, was retired on 15 June 2026. Moving to the new model is not a matter of changing a name in the configuration: the same instructions can give different answers, and they have to be tested before they reach your customers.
Answers change even without a new model, because your price lists, your opening hours and your customers' questions change. A periodic test on real questions, compared with the answers you expect, tells you whether the assistant still answers well. The model, then, is paid by usage, and the bill follows the volumes: the more customers write, the more it costs. Usage is your cost, declared in the quote, and it should be looked at month by month.
| AI item | What it covers | With us |
|---|---|---|
| Model change | the move to the new model when the provider retires the one in use, with tests before release | included in the fee |
| Answer test | every three months, your customers' real questions compared with the expected answers | included in the fee |
| Usage | a periodic report with usage month by month and its trend | report included, model usage on top |
With us, the move to a new model, when the provider retires the one in use, is included in the fee. We do not tie the software to a single provider, and a new model is a component change with a verification phase, not a rewrite. The usage report shows you in a chart whether the cost follows your customers or grows on its own.
The pages the engines cite say little about these tests. Of the 140 we read, cited on five questions about maintenance, 38 give the 15-25% percentage and 33 talk about response times, but only 5 mention a backup restore test and 3 a periodic test of the AI's answers.
What the pages cited by the engines on maintenance mention
pages out of 140One page can mention more than one item; the count is made on the text of the pages. Source: Volcanic Minds, reading of the 140 pages cited by Google, AI Overviews, AI Mode, ChatGPT and Perplexity on five questions about maintenance, 29 October 2026.
Would you rather talk it through?
Book a call with people who work on these topics every day: we will look at your case together.
How much does maintenance cost per year?
With us about 2,500 € a year for a simple management system, like the one of the garage in the guide to costs, and about 5,000 € for a management system with integrations. A customer-facing AI assistant on its own costs on average 2,500 € a year; a simple management system with the assistant, together, about 3,500 €. These are our estimates on typical cases, without hosting and without model usage: the figure for your case comes from the items of your software.
| Software | Maintenance per year, estimate | On top |
|---|---|---|
| Simple management system | about 2,500 € | hosting |
| Management system with integrations, for example with the accounting system and e-invoicing | about 5,000 € | hosting |
| Customer-facing AI assistant | on average 2,500 € | hosting and model usage |
| Simple management system with AI assistant | about 3,500 € | hosting and model usage |
The figures do not add up: management system and assistant together cost less than the two maintenance contracts on their own, because updates, backups, monitoring and support are done once for the whole system.
Hosting is on top: in November 2026 database, backend and frontend start from 950 € a year. If the software runs on a server of yours, in the company or owned by you, this item does not exist. The first month after release is under warranty, included in every fixed-price project of ours, for blocking bugs and minor tasks; maintenance, if you want it, starts after that.
Is 15-25% of the development cost a rule?
No. It is the figure that many pages and almost all AI engines repeat, but none of the answers we collected says where it comes from. Of the 140 pages read, 38 give the range and 14 name Gartner. Asked whether 20% a year is normal, Google's AI Overview and AI Mode answered yes.
The percentage is applied to a development cost, and when the engines do not know the cost, they make it up. To the same question, for the same 30-person company with a custom management system and an AI assistant, the fee ranges from 1,500 to 20,000 € a year.
| Who answers | Fee per year | How it is calculated |
|---|---|---|
| Google AI Overview | 7,500-10,000 €, hosting included | 15-20% of an assumed development cost of 50,000 € |
| Google AI Mode | 1,500-4,000 €, on the second request 3,000-5,000 € | 15-25% of a development cost it estimates at 5,000-15,000 €, on the second request 15,000-40,000 € |
| ChatGPT | 6,000-10,000 €, on the second request 12,000 € | item by item, in hours, with no percentage |
| Perplexity | 16,000-20,000 € | 20-25% of an assumed development cost of 80,000 € |
| Volcanic Minds | about 3,500 € for a simple management system with the assistant, estimate | item by item, with hosting and usage on top |
Answers collected on 29 October 2026 to the question of a 30-person company with a custom management system and an AI assistant that answers customers. Google AI Mode and ChatGPT, asked twice, gave different figures.
Take a provider asking for 20% a year. On a management system that cost 40,000 € that is 8,000 € a year, and in five years the fee pays back the development without one new feature; ChatGPT and Perplexity point this out too. The AI items, then, do not grow with the price of the project: a model change takes the same work on a 15,000 € management system as on an 80,000 € one. A fair fee is read item by item: if you do not know what you pay for, you do not know whether you pay too much. Ask the provider for the items in the first table with the work they estimate for each; why the percentage makes you pay for the size of the project is in the guide to costs.
What should a maintenance contract say?
The items included, the response times by severity, who owns code and accounts, how you leave. If there is an AI assistant, an annex on usage, model and data.
| Clause | What to write |
|---|---|
| Scope | the items included, one by one, and that new features are quoted separately |
| Response times | for each severity, from total outage to minor defect, within how long the provider takes charge and within how long it restores service, in working hours |
| Backup | how often, where the copies are, when the restore is tested, how much data you can lose and how long you can stay down |
| Ownership and access | code, repositories, service accounts and AI model keys in your name |
| Exit | the handover of code, data and documentation, and the transition to another provider |
| Term and price | how long it lasts, with what notice you cancel, how the fee can change each year |
The right times depend on what an hour of downtime costs you. For a total outage the engines' answers range from 30 minutes to 2 hours to take charge and from 4 to 8 working hours to restore service, but for recovery times there is no standard value. Code and accounts in your name matter on the day you change provider: the code is already yours, and once the month of warranty is over you have no obligation to stay with us.
What to add if there is an AI assistant?
| Clause | What to write |
|---|---|
| Usage | on which account model usage is paid, with what spending limit, how often the report arrives |
| Model change | who does it when the provider retires the model, with what tests before release, how to roll back |
| Answer quality | how often the assistant is tested, on which questions, what happens if the answers get worse |
| Handover to a person | when the assistant hands the conversation to one of your people |
| Data | the contract with the provider as data processor (article 28 of the GDPR): data only on your instructions, returned or deleted at the end, never used to train models |
| Transparency | that the assistant tells whoever writes that they are talking to an AI (article 50 of the AI Act, from 2 August 2026) |
What the AI Act asks of a company that uses AI, beyond the assistant's transparency, is in the guide AI Act for deployers.
A technical contribution
The rules are summarized from the text of the regulations, read on the date at the top of the page. Like the other guides, this is a technical contribution, not legal advice.
If you are still choosing who to entrust the software to, what to ask a software house is in the guide How to choose a software house, and what to write in the development contract in the guide Risks of a custom software project.
Sources and further reading
- Anthropic, «Model deprecations», documentationplatform.claude.com, accessed 30 October 2026
- OpenAI, «Deprecations», documentationdevelopers.openai.com, accessed 30 October 2026
- AI Act, article 50, transparency obligations for providers and deployersai-act-service-desk.ec.europa.eu, accessed 30 October 2026
- Regulation (EU) 2016/679 on data protection (GDPR), article 28eur-lex.europa.eu, accessed 30 October 2026
What does maintenance of custom software cover?
Bug fixes, security updates, adjustments when the connected services change, backups with a restore test, monitoring and support for the people who use it. New features stay out and are quoted separately, and hosting is a separate cost.
How much does maintenance of a custom management system cost per year?
With us about 2,500 € a year for a simple management system and about 5,000 € with integrations, hosting excluded. A simple management system with a customer-facing AI assistant costs about 3,500 € a year, model usage excluded. These are estimates on typical cases: the real figure comes from the items of your software.
Is 20% a year for maintenance normal?
It is the most repeated figure, not an estimate of the work. On a management system that cost 40,000 € it is 8,000 € a year, and in five years the fee pays back the development without one new feature. Ask the provider for the list of items included and the work they estimate for each.
What changes in maintenance if there is an AI assistant?
The move to a new model when the provider retires the one in use, a periodic test of the answers on real questions and the control of usage are added. With us the model change, the test every three months and the usage report are included in the fee; you pay for model usage.
What happens when the provider retires an AI model?
From the retirement date requests to the old model fail and the assistant stops answering. Anthropic says it gives at least 60 days' notice, OpenAI at least six months for generally available models; in that time you move to the new model and test it on real questions before release.
Is hosting included in maintenance?
No, it is on top. In November 2026 database, backend and frontend start from 950 € a year; if the software runs on a server of yours, in the company or owned by you, this item does not exist.
Do you have software to maintain?
Tell us what it does and which services it connects to: we tell you item by item what it needs each year, before you sign.
- You won't talk to a salesperson
- NDA before we get into the details
- The code stays yours
Articles on the same topics.
Web Bot Auth: visitors with no eyes
AI agents read your site and convert better than organic, yet they never show up in analytics. How to measure them and verify them with Web Bot Auth.
Read
Generative UI: for humans and machines
Explore how Generative UI and Dual-Interaction UX transform enterprise SaaS by creating adaptive interfaces for human operators and AI agents.
Read
Multi-tenant: SaaS data segregation
What multi-tenant means and how to pick data isolation in a SaaS: dedicated database, separate schema or shared row, and what changes with AI and RAG.
Read