Monitoring and repairs
We set up logs and alerts for errors, slow replies and failed tasks. When a connected tool such as your CRM or help desk changes and breaks a workflow, we fix the link. Each fix is written up in the change log.
AI support and monitoring
AI system maintenance for the AI you rely on
We watch, test and update your AI systems so they keep doing the job you built them for.
AI system maintenance is the ongoing work of keeping a live AI system accurate, safe and affordable. It starts the day the system goes live and lasts as long as people use it.
AI rarely breaks all at once. It drifts. The model changes, your information goes out of date, and answers get worse until a customer spots it first.
Who this is for
For companies whose AI already does real work
This service suits a business that runs AI every day but has no one whose job is to look after it.
The system might answer customers, qualify leads or sort documents. What matters is that people depend on it and would notice if it got worse.
A support or sales assistant that answers customers in your name
Workflows that read or write documents with no person checking each one
An internal AI tool your team now uses every day
An AI step inside your sales process, such as lead research or scoring
It is not the right fit if you are still exploring AI, or if your system is a pilot nobody relies on yet. If you need new features rather than upkeep, that is a build project. We will tell you which one you need.
Signs it needs attention
How a working AI system slowly stops working
Most AI problems never show up as an outage. They show up as small changes that nobody is paid to notice.
The model moved
Providers retire older model versions on their own schedule. The replacement behaves differently, and no one tested your use case against it.
The facts went stale
Your prices, products and policies changed, but the documents your AI reads did not. It now answers with confidence using last quarter's facts.
The bill crept up
Usage grew, prompts (the written instructions your AI follows) got longer, and no one reads the invoice line by line. Each month costs a little more, and nobody can say why.
The builder left
The freelancer or employee who set it up has moved on. The prompts, keys and settings now live in their head, or in an account you cannot open.
None of these set off an alarm, which is why they can run for months before anyone looks.
Scope of work
What AI system maintenance covers
The work falls into six areas. Each runs on a schedule, and each shows up in your monthly report.
Quality testing
We build an evaluation set: a fixed list of real questions, each with an agreed good answer. It runs after every change and on a regular schedule. When quality drops, you see the exact question that failed.
Model and provider updates
Providers release new models and retire old ones on their own timetable. We track those notices, test the new version against your evaluation set, and switch only when it holds up.
Prompt and knowledge upkeep
Your AI is only as current as the instructions and documents behind it. We update prompts when your team wants new behavior, and refresh the knowledge base when your offer or policies change.
Cost control
Providers bill by usage, so costs move with every longer prompt and busier month. We review spend monthly, cut waste, and send simple tasks to cheaper models where quality holds.
Security and access
We review who and what can reach the system, and rotate access keys on a schedule. We also test for prompt injection (text written to trick the AI into ignoring its rules). Anything that touches customer data is checked against your own data policy.
Send us one conversation your AI got wrong. We will tell you where we think it broke.
Book a Strategy Call →The process
From handover to a monthly rhythm
The first month is spent learning your system and setting a baseline. After that, the work settles into a rhythm you can plan around.
Week 1
Access and inventory
We get access to the model accounts, prompts, logs and connected tools. Then we write down what runs where, who owns it and what it costs.
Weeks 2 to 3
Baseline tests
We pull real questions from your logs and agree with you what a good answer looks like. The first run of the evaluation set shows where the system stands today.
Week 4
Fix list
You get a written list of issues, ranked by risk to your customers and your budget. We agree together which fixes come first.
Monthly
Checks, changes and a report
Monitoring runs all the time. Every change passes the evaluation set before release, and new failures from live use become new tests. You get a report on what changed, what broke and what it cost.
Quarterly
Look ahead
We review upcoming provider changes, usage trends and new risks. You decide which changes are worth making in the quarter ahead.
A change that fixes one answer can quietly break another, so nothing ships until every test that passed before still passes.
Deliverables
What you receive
Good AI system maintenance leaves a paper trail. These are the files and tools you hold, and they stay yours whether or not we keep working together.
- A system inventory listing every model, prompt, key, connection and owner
- Your evaluation set of real test questions and agreed answers, in a file your team can edit
- A live dashboard of errors, usage and spend, with access for your whole team
- A change log recording what changed, why, and how it tested before release
- A monthly report on incidents, fixes, costs and what we recommend next
- A runbook explaining how to pause, roll back or restart the system without us
If you ever bring maintenance in house, this set is your handover. Nothing about your system lives only in our heads.
Examples of past work are available on request, so get in touch and we will send work relevant to your market.
Our commitments
What we promise in writing, and what we won't
Our outreach services carry the Assured Outreach Model. If agreed KPIs are not met in time, work continues at no additional service charge until they are. Maintenance does not fit that shape.
Most AI runs on a provider's servers, outside your control and ours. Anyone who promises it will never fail is guessing. These are the commitments we can keep, and each one is written into your agreement.
The systems we cover and the checks we run are listed before work starts
Response times are agreed for each level of problem, from urgent to routine
No change reaches your live system until it passes the evaluation set
Every release can be rolled back to the last working version
Accounts, keys, prompts and logs stay in your name, not ours
What we will not promise: zero errors, or uptime on servers we do not run. When a provider has an outage, we tell you, keep you posted and rerun the evaluation set once it clears.
Common questions
Frequently asked questions
Our AI was built by someone else. Can you still maintain it?
Usually, yes. We start with an audit of the model, the prompts, the connected tools and the accounts behind them. If the system sits inside a closed platform with no admin access, we will tell you there is little to maintain.
Can you guarantee our AI will never make mistakes?
No, and be wary of anyone who does. Every AI model makes errors some of the time, and outages at the provider are outside our control. We can promise that every change is tested first, problems get answered within agreed times, and you see the results monthly.
How is AI system maintenance priced?
It depends on how many systems you run, how complex they are and how often they change. We quote once we have seen the system, not from a rate card. Model usage fees stay on your own account, billed by your provider directly.
What access do you need to our data and accounts?
Only what the work requires, agreed in writing before we start. That usually means the model provider account, the system's logs and the tools it connects to. Where your setup allows, we review logs with customer details masked.
What if we want new features, not just upkeep?
Small changes, like a new instruction or an updated document, are part of maintenance. New capabilities, such as a second agent or a new integration, are a separate build project. We scope those on their own, so build work never eats into your maintenance time.
The next step
Walk us through the AI you already run
On a Strategy Call, we ask what the system does, who relies on it and what has changed since launch. You will hear what we would check first, and whether you need outside help at all.