AI agents · 11 min
AI agent maintenance: how to define ongoing support
Use this scope template to agree who keeps your agent useful after launch and what evidence you receive each month.

What an AI agent maintenance service should cover
An AI agent maintenance plan should specify who monitors operation, who checks the usefulness of results, who fixes failures and who approves changes or additional spending. A monthly quote that promises only “support and optimisation” leaves these responsibilities unresolved. Ask for recurring activities, evidence of completion and a clear boundary between maintenance and new development.
This guide is for a business already using an agent, or completing a pilot, that needs to choose its ongoing support arrangement. Another impressive demonstration will not settle that decision. You need someone to notice when the assistant uses an outdated price list, a connection stops working or the cost of completing useful work rises. You also need to know who acts on that information.
Consider an illustrative B2B sales assistant that turns incoming enquiries into draft quotations. A salesperson reviews each draft before sending it. After a catalogue change, the agent starts overlooking a new product group. The application remains available and the platform bill is paid, but salespeople increasingly rewrite its output. A support service that watches only technical outages may never recognise this deterioration as part of its remit.
Start with a support handover inventory
Map one real journey from incoming request to completed work. Identify the enquiry source, approved documentation, business system, model provider, connections and person who accepts the result. For each component, record who controls access and who can approve a change. Describe the production system as it works today, rather than the original project presentation.
Include the current instructions, permitted tasks, representative test cases and a procedure for continuing manually. If a supplier is taking over another team's implementation, ask them to reproduce a normal task and identify a failure using the materials provided. Without this check, they cannot credibly estimate the ongoing support workload.
Missing documentation creates an initial piece of work. Separate it from the recurring fee, otherwise the first month can disappear into discovering what was built. Distinguish an initial assessment and handover from steady operation. The assessment should produce a list of gaps, responsibilities and conditions that must be met before the support arrangement can begin.
Separate five kinds of work in the proposal
The first is technical operation: availability, functioning connections and failed jobs. The second is knowledge maintenance: which approved document or record the agent uses. The third is quality review against agreed cases. The fourth is change management. The fifth is reviewing consumption and costs. Each needs a visible deliverable, even if the same supplier performs several activities.
For a quotation assistant, a broken catalogue connection is a technical problem. An incorrect minimum-order rule is a business-content problem. A request to support an additional export market expands the task. Distinguishing these cases helps assign responsibility and avoids treating every issue as an argument about what the monthly fee includes.
The comparison matrix below is a working template, not a description of a standard NUMEDIA package. Request a concrete answer for your system in each row. When an activity is excluded, identify who will perform it instead. An empty cell represents unowned work, even when the overall quote appears attractive.
| Activity | Evidence to request | Clarify in the proposal |
|---|---|---|
| Technical operation | Failed-job record and resolution | Who monitors, when they respond, how work continues |
| Sources and documentation | Approved sources and change history | Who approves content and who updates it |
| Quality review | Reviewed sample and correction reasons | Who judges usefulness and accepts the fix |
| Changes | Change record and test results | What is included and what needs new scope |
| Consumption | Costs by type and accepted task | Who receives alerts and approves spending |
Specify response and recovery separately
An initial response time tells you when someone acknowledges and takes ownership of a report. It does not tell you when the problem will be resolved. Define intake, initial assessment, the next update and restoration of a usable process separately. When an external platform fails, your supplier may not control its repair time, but can prepare an alternative way to continue working.
Priority should follow business impact. An incorrect write to a business system calls for a different response from an awkwardly phrased draft. For each priority, agree how to report the issue and who decides whether the agent's activity should be restricted. Also specify what happens outside support hours and who checks the outstanding queue on the next working day.
Google's SRE guidance distinguishes service indicators, target levels and service agreements. For a support buyer, the practical point is to connect a target to a useful task. “The agent is reachable” is a weak indicator if enquiries never reach the sales team. Choose a small number of measures that actually change the decision to intervene.
Give quality review a business owner
A technical supplier can check that an output has the required structure and that an integration returned a response. Someone in the business must judge whether the result is useful. Include a recurring review of selected cases with a person who understands the work. Cover ordinary tasks, reported problems and cases where handing the task to a person was the correct outcome.
For the illustrative quotation assistant, the reviewer might classify outputs as usable, needing minor edits, containing a substantive error or outside the agreed scope. Record a reason with each classification. An outdated document, an unclear commercial rule, the wrong product selection and missing input information require different remedies.
OpenAI recommends evaluating changes and expanding test sets with newly observed cases. Anthropic describes evaluation, production monitoring and human review as complementary practices. Turn this into a purchasing requirement: request evidence of a repeat test after a fix, rather than a message saying that the instructions were updated. Our linked guide to AI quality covers the design of evaluation criteria in more detail.
Compare spending with accepted work
The model provider's invoice is only one cost component. Add hosting, integrations, monitoring tools, maintenance labour and employees' time correcting results. Keep usage charges separate from recurring support and new development. This makes it possible to see which costs change when task volume increases and which remain fixed.
Define a completed, useful task. For quotation preparation, it could be a draft accepted by a salesperson within an agreed editing allowance. Ten retries of the same task are not ten successful quotations. Counting only calls or conversations can disguise repeated failures as increased adoption.
Here is an illustration using invented assumptions. Suppose monthly operation and support total €600 and the team accepts 300 drafts. That is €2 per accepted draft before employee review time. If only 150 drafts are accepted, the ratio becomes €4. These figures are neither a price quote nor an estimate of savings. They show why each report needs the same denominator.
When spending rises, ask the supplier to explain the cause before recommending a bigger budget: more incoming work, longer documents, more retries or a different implementation. Agree who approves a budget change and what happens at an operational spending limit. An alert without an owner and a response procedure does not settle that question.
Distinguish maintenance from new capability
Restoring an agreed connection after a failure differs from adding another CRM. Replacing an approved document differs from designing commercial rules for another country. Ask a prospective supplier to classify three examples from your own operation. Their answers reveal more about the proposed scope than a long list of generic exclusions.
For each change, request a short record of the reason, expected behaviour, affected processes, test procedure and person responsible for acceptance. A model change should be compared on the same tasks for quality, completion time and consumption. A newer model release is not, by itself, a business case for switching.
Decide who updates staff instructions as well. If the agent changes how it flags incomplete enquiries, salespeople need to know where to find them. A technically successful release can create more manual work when its users have not been briefed. Close the change with test evidence and a concise explanation for the responsible user.
Choose a support arrangement your team can operate
Internal maintenance is a reasonable option when someone understands the process, can inspect the system and has time to fix it. Employing a developer does not automatically provide that capacity. Check who covers absence and which other work will be delayed when a support issue arrives. Ownership should survive a holiday or a change of role.
Shared support can work when the business maintains commercial knowledge and judges results while an external supplier handles implementation and change testing. It needs one reporting location and a clearly identified first contact. If sales reports the same problem to three people and nobody coordinates the response, splitting responsibilities has made the process harder to manage.
A broader external service may fit an agent that matters to daily operations when you lack an internal technical team. You still need a business owner. A supplier cannot independently decide which discount is acceptable or which offer is current. You are buying the operation of an agreed process, not transferring every commercial decision to a support desk.
When a monthly package is not the next investment
If the agent is rarely used and a simple form or rule reliably completes the task, compare simplification first. If poor source data causes weak results, buying monitoring alone will not repair it. If nobody can describe an acceptable output, agree the task and acceptance criteria before expanding support.
Continuous coverage is not automatically necessary either. An internal drafting assistant may have a workable manual fallback until the next business day. An agent on the path of incoming orders requires a different assessment. Choose coverage based on the consequence of interruption and the available alternative, rather than the most comprehensive package name.
“We already paid to build it” is a reasonable purchasing objection. Implementation establishes agreed behaviour in a particular environment; catalogues, rules and connected systems subsequently change. However, the supplier should distinguish correcting a defect in the delivered implementation from meeting a later requirement. A maintenance plan should make the work more understandable, rather than introduce an unexplained monthly charge.
A monthly review that ends in a decision
The review should produce a decision, not a dump of every technical log. Use one page to show workload, quality in the reviewed sample, unresolved issues, spending and the next proposed change. For a small sample, show the case count. A percentage without a denominator can give an exaggerated impression of reliability.
For each open issue, record its consequence, owner, next action and next review date. Separate confirmed defects from improvement requests. Then choose a course of action: continue the current scope, make a bounded improvement, reduce the agent's autonomy or temporarily return to a manual process. Someone who missed the meeting should still understand the reason.
Attach a short change record with test results. If no changes were made, the report should still show completed reviews and the state of the information sources. The purpose is not to make maintenance look busy. It is to establish whether the process remains useful and which piece of work deserves funding next.
Prepare a brief that suppliers can price consistently
Send NUMEDIA a description of one agent task, its connected systems, approximate usage and the three problems that concern you most. Explain who currently accepts results and how work continues if the agent is unavailable. A description and anonymised examples are enough for an initial discussion; do not send credentials or entire customer databases.
The first useful output is a proposed division of responsibilities and a defined initial assessment. If the implementation is documented well enough, you can proceed to comparing ongoing support. If it is not, start with a bounded review and an explicit deliverable. This gives you a concrete basis for deciding whether a recurring service is appropriate.
Sources and methodology
- Service Level Objectives (Google SRE, accessed 2026-11-14)
- Evaluation best practices (OpenAI, accessed 2026-11-14)
- Demystifying evals for AI agents (Anthropic, accessed 2026-11-14)
Frequently asked questions
Is paying for the model enough?
No. Model access or usage charges do not assign responsibility for your knowledge sources, output quality or integrations. Those activities need an internal owner or a supplier with an explicit scope.
Can another supplier take over support?
Potentially, if the necessary documentation, authorised access and verification process are available. Start with a bounded handover assessment before committing to the recurring scope or takeover date.
Does ongoing support include new features?
Only where the scope explicitly includes them. Adding a market, business system or permitted action should be distinguished from maintaining agreed behaviour. Use actual change examples to compare proposals.
How often should the business review results?
Set the frequency according to error consequences, usage and changes. A new or changed process needs closer review than a stable internal task. Also identify the events that trigger an additional review and who performs it.