Home / AI implementation partner
Working alongside your IT supplier

AI implementation partner

You already have an IT supplier, or an internal IT team, or both. An AI implementation partner is not a replacement for either. It is the specialist you add for one thing your existing supplier does not carry: a system whose output is a judgement rather than a calculation, and which therefore has to be evaluated, monitored and governed differently from every other application in your estate.

By Tom Joseph · Last updated: 24 September 2026

Summarize with AI Prompt copied. Paste it into the chat
Share
In short

Crux Digits is an AI partner, not an ICT partner. We do not take over your infrastructure, your service desk or your systems management, and we do not want to. We arrive next to the supplier you already have, take one process from audit to production, and hand over the code, the models and the evaluation set. What changes in your IT organisation is narrow and specific, and this page sets it out: who owns the model, who owns the evaluation set, who receives the incident, and who signs off a change that alters a model's behaviour without a release.

Adding a specialist to an existing landscape

What changes, and what does not

Five steps. Your existing supplier keeps everything that is already theirs.

  1. An engineer working at a console in a server roomToday

    The landscape you already run

    Your IT supplier or internal team owns infrastructure, the workplace and the service desk. None of that changes.

  2. Two colleagues reviewing documents at a desk in an officeAudit, €2,500

    One process, costed

    One to two weeks: the process measured, the data checked, a return per opportunity and an EU AI Act judgement per idea.

  3. A team reviewing printed charts together around a tableEvaluation set

    The proof it is still right

    Real cases with the answer your own experts agree on. Without that set nobody can see a model drift.

  4. A developer writing code at a workstation in an open officeBuild

    Inside your standards

    On your data, behind your identity provider, connected through the interfaces you already maintain.

  5. Two engineers discussing a diagram at a whiteboardHandover

    Code, models and runbook

    You receive everything needed to run it yourselves or to hand it to a different supplier.

The split

What an AI implementation partner does, and what your IT supplier keeps doing

ResponsibilityYour IT supplier or internal teamThe AI implementation partner
Infrastructure, network, workplace, service deskOwns it, unchangedNo role
Application landscape and integrationsOwns the estate and the standardsBuilds to those standards and to the interfaces you already expose
The model and its behaviourUsually has no basis to judge itOwns it: selection, prompts or training, versioning and rollback
The evaluation setRarely exists before an AI projectBuilds it with your people and hands it over, because it is what proves the system still works
An incident where the output is wrong but nothing is downNot a category in most SLAsTriage sits here, because the fix is a model or data change rather than a restart
Change controlOwns the processSupplies the evidence a change advisory board needs to approve a model change
Security and accessOwns policy and identityWorks inside it: your identity provider, your roles, your retention rules
The deliverable

What a technical implementation consultant leaves behind

Judge the engagement by what survives it, not by who attends the meetings. Four things should be in your hands at the end, and a project that cannot produce them has not been implemented, it has been demonstrated.

  • The working system, in your environment. Running against your data, in your tenant, behind your identity provider, integrated with the systems it needs (AFAS, Exact, an ERP, a WMS or TMS, an EPD, a case management system) through the interfaces you already maintain.
  • The evaluation set. A held-out set of real cases with the answer your own experts agree on. Without it nobody can tell a model that has drifted from a model that is fine, and every future discussion becomes an argument about impressions.
  • The runbook. What to check weekly, what the thresholds are, what to do when one is crossed, and who to call. Short enough that your own second-line reads it.
  • The code, the prompts and the model configuration. In your repository, under your licence, with the documentation. You should be able to hand the whole thing to a different supplier and have them continue.
The seam

Where two suppliers go wrong, and how to prevent it

Nearly every failure we are called in to repair is a boundary problem rather than a technical one. Four of them account for most of it, and all four are settled in an hour before the project starts rather than in a meeting after it breaks.

  • Nobody owns the evaluation set. The AI supplier built it, the IT department never took it over, and six months later nobody can say whether quality has moved. Name an owner inside your organisation on day one, not a supplier.
  • The incident has no queue. Your service desk categorises by availability, and a model that answers confidently and wrongly is available. Add the category before go-live and agree what evidence a ticket must carry: the input, the output, the expected output.
  • A model changes without a release. A provider updates the underlying model, or a prompt is tuned, and behaviour shifts with no change record. Decide whether that is a change under your own process. Usually it is, and then it needs a version, a note and an evaluation run.
  • Exit is not written down. Ask at the start what leaving looks like: which artefacts you receive, in what format, and how long the supplier supports a handover. A partner who cannot answer cleanly is selling dependency.
The test

What to ask an AI implementation partner before you sign

Five questions, and the answers are checkable rather than reassuring.

  • What have you put into production, not piloted? Ask for the number the client measured, and ask what it was before.
  • How do you evaluate the system, and who holds that evidence afterwards? If evaluation is a demo rather than a set, there is nothing to monitor against.
  • What does this cost to run per month, and what happens if we stop paying you? If the honest answer is that it degrades, the build price was never the price.
  • How will you work with our existing supplier? A partner who wants to route around your IT department is creating a second estate for you to manage.
  • What would make you say AI is not the answer here? A supplier who has never said it is selling technology rather than an outcome.

Our audit is built to be judged on exactly these. See what to ask before you hire an AI consultant and our thirteen published cases.

How it runs

Three fixed steps, fixed prices, your code

AI audit and strategy, EUR 2,500: one to two weeks. Your processes and data checked against what you actually run, use cases ranked by return, an EU AI Act and GDPR judgement per idea, and a plan another supplier could execute. Production-ready MVP, EUR 20,000: four to six weeks, working on your own data. Production, from EUR 50,000: integrated, secured, monitored and handed over with the evaluation set and all source code. Prices exclude VAT, and the audit is credited in full if you continue within 60 days. For the service itself see AI implementation, and for cost and duration see AI implementation in the Netherlands. Once a system we built is live, managed AI support covers monitoring and improvements to that system at a published monthly price. It covers what we built, not your wider IT estate.

What you can verify
One real engagement

AI Predictive Maintenance for Industrial Plants

What we found. Run-to-failure maintenance meant recurring, unplanned stoppages that cost output and forced expensive emergency repairs. The producer needed early warning of failures so work could be scheduled into planned windows instead of triggered by a breakdown.

<2%Unplanned downtime (down from recurring)
−25%On-site maintenance time

Real result from a Crux Digits engagement; client name withheld pending permission.

Read the full case study →

FAQ

Frequently asked questions

What does an AI implementation partner do?

It takes an AI use case from decision to a system in production and stays accountable for the parts a general IT supplier does not carry: model selection and versioning, the evaluation set that proves the output is still right, monitoring for drift, and the EU AI Act and GDPR duties that attach to the specific application. It works inside your existing standards, identity and change process rather than beside them.

How is that different from our ICT partner?

Your ICT partner owns infrastructure, the workplace, the service desk and the application estate, and that does not change. An AI implementation partner owns the behaviour of one system whose output is a judgement. The distinction matters most during an incident: a normal incident is something being unavailable, and an AI incident is usually something being available and wrong, which needs a different triage and a different fix.

Does Crux Digits replace our IT supplier?

No, and we do not offer to. We do not sell infrastructure management, service desk, hosting or 24/7 support. We are a specialist supplier next to the one you already have. In practice that means we build to your standards, use the interfaces you already maintain, and hand over the code and models so your own team or your existing supplier can carry it.

What does an AI implementation project cost?

Crux Digits works in three fixed steps: an AI audit and strategy at EUR 2,500, a Production-ready MVP at EUR 20,000 and production from EUR 50,000, excluding VAT, each with the scope and the lead time fixed in advance. The audit is credited in full against the next step if you continue within 60 days.

Will you work with our existing supplier?

Yes, and it is usually the faster route. Your supplier knows the estate, the standards and the constraints, and duplicating that knowledge wastes budget. We agree the split in writing before the build starts: who owns the model, who owns the evaluation set, who receives which incident, and what a model change means under your change process.

Who fixes it when the system gets something wrong?

Triage sits with us for the model and the data, because that is where the fix is. Your existing supplier keeps everything that is genuinely an availability or access problem. That split is agreed before go-live and written into the runbook you receive, so a ticket does not spend two days moving between two suppliers.

Bring us in next to the supplier you already have

A free 30 minute call: you describe the process and your current landscape, we tell you whether AI pays off there and what the first step costs.

Book a free consultation →