AI Application Maintenance
Ongoing maintenance for AI-touched systems - provider migrations, prompt tuning, dependency upgrades and the quiet engineering work that keeps production systems running as the AI landscape shifts.
Entry packageMaintenance Handover Packfrom EUR 2,6001 system · up to 2 weeks from scopingWhat is includedModel deprecations without a scramble
Provider migrations planned and executed with an eval run guarding the output.
Prompts tuned on a cadence
Quarterly re-eval, regression check and minor tuning as your inputs drift.
A quarterly report you can budget from
What changed, what to expect, what to budget for next quarter, in writing.
23Minds provides ongoing maintenance for AI-touched applications.
Show the full summary
AI systems need maintenance that traditional applications do not: provider migrations when a model is deprecated, prompt tuning as inputs evolve, eval re-baselining, dependency upgrades for the LLM SDK layer, cost monitoring as spend drifts. Retainer model fits teams with continuous AI features; fixed engagements fit specific events like a provider migration. Provider notice before a model retirement differs: at least 60 days at Anthropic for publicly released models, at least six months at OpenAI for generally available ones (checked 25 September 2026).
What makes this service defensible.
- Handover pack opens the retainer
State audit, dependency and model inventory, migration plan, runbooks and an opening report in two weeks.
- Eval-protected migrations
No provider or model switch without a before-and-after score.
- Retainer or fixed engagement
Continuous AI features get monthly hours; a single migration gets a fixed price.
- ISO 9001:2015 quality system
Certified by DNV, no. 10000439701-MSC-RvA-POL; maintenance runs as a documented procedure.
The process, phase by phase.
Pick the way in that fits your budget and timing.
Every practice has a free first step, an entry package with a written deliverable and its price listed here, and a full project or retainer quoted after a first call. Once the scope is agreed, the price is fixed, with no hidden costs.
- 1Step 1 · Free
Intro call or self-assessment
60 minutes · 1 engineer · no obligation
FreeTalk to an engineerShow details
- Intro call with an engineer, or the online self-assessment
- A written next step within one business day
- No obligation and no sales deck
- Recommended2Step 2 · Entry package
Maintenance Handover Pack
1 system · up to 2 weeks from scoping
from EUR 2,600 netThe price is fixed once the scope is agreed on the intro call, no hidden costs; extras only from the Not included list.
Ask for this packageShow details
- State audit of one system
- Dependency and model inventory
- Provider migration plan
- Runbooks and an opening report
- Not included: Ongoing maintenance hours, fixing issues found in the audit, 24/7 on-call, SLA.
- 3Step 3 · Project or retainer
Full scope, quoted after a first call
Full scope · phased delivery · quoted per phase
from EUR 2,800 net per month, quoted after a first callStart the conversationShow details
- Maintenance retainer
- Provider migrations and quarterly prompt tune-up
- Written quarterly report
The situations that bring teams to this service.
- Provider deprecates a model; the team has no migration plan and no time to write one
- Prompt that worked at launch produces worse outputs six months later - drift
- LLM SDK version is stuck at launch-day version; security patches piling up
Show the other problems (1)
- Cost trends are rising slowly enough that nobody noticed until the quarterly review
- Engineering teams with AI features in production who need part-time senior support
- CTOs whose team built the AI feature and moved on; nobody owns it now
- Product teams whose AI providers deprecate models faster than the team can react
- Mid-stage companies extending or maintaining AI features without growing the team
Where this goes wrong, and what we do about it.
1. The model your feature stands on gets retiredWhat stays with you: regression set with results for old and new model side by side, and the decision recorded per difference
What happensProviders deprecate on notice, and the notice is measured in weeks. A feature built on a specific model works until the day it does not, and the migration then happens under time pressure with no way to tell whether quality moved, because there was never a baseline.
What we do about itWe keep a regression set for the behaviours that matter, run it against the replacement before the switch, and report the differences rather than a pass or fail. Where a difference is acceptable we record why.
2. Prompts drift with no history and no testWhat stays with you: versioned prompts with review, version stamped on responses, and behaviour tests that fail for the right reason
What happensA prompt gets edited in place because it is text, not code. There is no diff anyone reviews, no version attached to the output, and no test that fails. Six edits later the behaviour has moved and nobody can say which change did it or return to what worked.
What we do about itWe move prompts under version control with the same review as code, attach the version to every response, and add cases that fail when the behaviour changes rather than when the text changes.
3. The unglamorous work has no ownerWhat stays with you: maintenance schedule with named owners and the automation that removed the repetitive part
What happensDependency updates, provider migrations, expiring keys and quota changes are invisible until they are an incident. Invisible work does not get assigned, so it waits for the person who notices, and that person is usually whoever is on call when it breaks.
What we do about itWe name the recurring work, put it on a cadence with an owner, and automate the parts that can be automated so the rest is small enough to actually happen.
Deliverables, not deliverable-shaped slides.
- Monthly maintenance hours covering the issues above (retainer)
- Provider migration plans + execution when models deprecate (eval-protected)
- Quarterly prompt tune-up: re-eval, regression check, minor tuning
- Dependency upgrade cadence: SDKs, observability libraries, eval frameworks
- Quarterly written report: what changed, what to expect, what to budget for next quarter
Questions procurement teams ask.
Where the dates and numbers come from.
Talk to an engineer.
Describe where you are with AI application maintenance. A reply within one business day.