Claude Partner Network · Sustained Adoption & Scaled User Support

An AI center of excellence is a rota and an on-call, not a steering committee.

Two halves of one function: the charter that makes decisions binding, and the support tier that answers users inside a working day.

Pexon stands up and runs an AI center of excellence as an operating function rather than a committee: a named rota, a support tier with an on-call, published decision records, and a monthly cost report per team. Sustained adoption gets measured in work that changed shape, not in logins.

Our position

A function has a queue. A forum has an agenda.

The thing sold as a centre of excellence is usually a forum: interested people, a recurring invitation, and guidance nobody is obliged to follow. It fails in a recognisable way. Nobody is on call, so questions go to whoever answered last time. Nothing is written down, so the same entitlement question is re-decided three times with three answers. Attendance decays, the meeting is quietly dropped, and the work returns to whoever shouts loudest.

A function is the same people with three things added: a queue with a response target, a rota naming who is on it this week, and a decision log that is binding until somebody overturns it in writing. None of the three is sophisticated. All three are administrative, which is exactly why they get left out of a programme designed to look strategic.

We build this way ourselves, and the part we got wrong first was the decision log. Our own early rollouts had a rota before they had records, so the on-call person answered well and the answer evaporated. The second time the same question arrived it was answered differently, by the same team, in good faith. A log is not bureaucracy; it is the only thing that makes a rota compound instead of reset.

Three ways to staff it

Who holds the queue

The choice is who carries the rota, and it is expected to change. Almost every engagement starts on the first model and ends on the third; the useful conversation is how fast, not which one forever.

We run it

A forward deployed engineer holds the rota and the queue, chairs the decision review, and writes the records. Your people are on the rota as participants from the first week, which is what makes the handover later a change of ownership rather than a restart.

Ours: queue, on-call, decision records, monthly report.

We run it with you

The rota is shared and the decisions are joint. This is where most engagements sit after the first quarter: your practitioners take first line because they know the business context, and the architecture calls stay joint because that is where the expensive mistakes are.

Shared: your first line, our second line and architecture.

You run it, we are on call

Your function, your rota, your records. We hold a monitoring tier and are reachable when something is either architectural or on fire. The cheapest tier, and the right one once the triage categories have stopped changing every month.

Yours: everything except escalation and architecture review.

What we report every month, in this order

Which named workflows now run differently from how they ran last month.
How long a question waited before it was answered, and the worst case rather than the mean.
Decisions taken, and decisions overturned — the second number matters more.
Spend per team against what that team expected to spend.
What is stuck, and with whom, including when it is stuck with us.

How far to go

Three sizes, and the lightest is a real answer

A channel and a callA named control function

A guild
Lightest
What it is
A channel and a fortnightly call. No rota, no queue, no budget.
Who it suits
Under a few hundred employees, one or two teams using Claude, no regulated data in scope.
What it cannot do
Make a decision that binds anyone, or answer a question inside a working day.
Honest verdict
A legitimate permanent answer for a smaller company. Not a failed centre of excellence.
A funded function
Middle
What it is
A named rota with released time, a queue with a response target, and published decision records.
Who it suits
Most companies between 500 and 5,000 people once more than three teams are using Claude in earnest.
What it cannot do
Carry a formal control obligation on its own, or operate production systems overnight.
Honest verdict
Where the large majority of engagements belong, and where they should stop.
A control function
Heaviest
What it is
The function above plus named accountability for an obligation: model inventory, eval evidence, approval gates with an audit trail.
Who it suits
Companies with a supervisor to answer to, or a deployment that falls inside the EU AI Act's higher-risk obligations.
What it cannot do
Be stood up quickly. The evidence trail has to exist before the obligation bites, not after.
Honest verdict
Buy this only if an obligation requires it. Adopted voluntarily it slows everything down for no gain.

The charter

Eight lines, and it fits on one page

What it decides. The specific question types this function rules on — entitlement, tool approval, model choice — and what it explicitly does not rule on.
Who is on it. Named people with a stated share of their week. A percentage below about fifteen is zero the first week something urgent happens.
The queue and the target. One inbox, one response target, and what happens when the target is missed. Two inboxes is no inbox.
Who is on call this week. A rota that is visible to the people who need it, not a name somebody remembers.
Where decisions are written. One log, in a system your people already open, binding until overturned in writing.
How something gets overturned. A named route to reverse a decision. Without one, the log becomes a thing people route around rather than argue with.
What gets reported, and to whom. The monthly list above, to the budget holder. Reporting to nobody in particular is how a function stops being funded.
When it stops. The condition under which this becomes ordinary line work and the function is dissolved. A programme with no end state defends itself instead of working.

The support half

What first and second line actually handle

Scaled user support for Claude Code and Cowork is mostly triage, and the categories are stable enough to publish. The value is in deciding which of four things a ticket is before anyone engineers anything.

Access and entitlement: the user cannot reach something, and the question is whether they should be able to. Resolved against the identity model, not by widening a scope.
Refusals that are correct: the tool declined and the user reads it as a fault. The answer is an explanation and, sometimes, a decision record that already covers it.
Refusals that are wrong: the permission model is too narrow for a legitimate task. This is the queue's most valuable output, because it is the only reliable signal that the entitlement design needs changing.
Unexpected behaviour: the agent did something surprising. Triaged into prompt, tool definition, model change or genuine defect — and the model is the least likely of the four.
Cost surprises: a team's spend moved. Answerable in minutes where per-call attribution exists, and not answerable at all where it does not.
Anything third line and genuinely a defect in something we built goes to the engineers who built it, under the operating tier you are on — from €3,000 per month for monitoring only, net of VAT.

When we are the wrong choice

If you have fewer than about three teams using Claude seriously, you do not need this and a guild is the honest answer; funding a function to govern two teams produces process for its own sake and we would rather tell you that than sell it. If what you actually want is a training programme with a defined curriculum and an end date, that is a different purchase and a cheaper one. And if nobody can be released onto a rota, buying an operating engagement does not fix it — we end up running a function your own people are not in, which works until we leave and then stops working, which is the one outcome worth refusing.

Claude Partner Network — Certified Services Partner

Pexon Group is a Certified partner in Anthropic's Claude Partner Network (Services Track), granted on 1 October 2026. This site is published by PEXON ROMANIA S.R.L., a separate legal entity in the same group, and the party you contract with here.

Where this goes next

The Claude Partner Network catalogue — All sixteen Services Track categories and which page each one lands on.
Claude Code training — The curriculum with an end date, if that is what you actually need.
Pricing — The published monthly operating ladder this engagement is priced from.
Claude cost optimisation — Where the per-team spend number in the monthly report comes from.

What the sponsor of the programme asks

What makes a center of excellence fail?

Having no queue. The failure is reliably the same: a group of interested people with no rota, no inbox and no authority, who meet fortnightly and produce guidance nobody is obliged to follow. Within two quarters the meeting is cancelled and the work has moved back to whoever shouted loudest. A function with a named on-call and a published decision log survives; a forum does not.

How many people does this need from our side?

Fewer than most proposals claim, and they have to be partly released rather than fully committed. One person with real authority over the decisions, and two or three practitioners with a defined share of their week on the rota. Three people at a genuine twenty per cent each beats ten at a notional five, because five per cent is zero the first week something urgent happens.

What does the support tier actually cover?

First and second line for your Claude Code and Cowork users: access and entitlement questions, why a tool call was refused, why an agent did something unexpected, and the triage that decides whether something is a user misunderstanding, a policy decision or a defect. Third line, where it is a defect in something we built, goes to the engineers who built it.

How is adoption measured, if not by licence usage?

By work that changed shape. Logins and message counts rise during any rollout and tell you almost nothing, because a curious user and a dependent one look identical in them. What we report is which named workflows now run differently, how many people are on the rota, how long a question waits before it is answered, and the spend per team against what the team expected.

What does it cost?

This is an operating engagement rather than a sprint, so it is priced per month from the published ladder: monitoring only from €3,000, one forward deployed engineer at €7,500, up to three at €21,000. All net of VAT. Model consumption is billed to you directly by Anthropic, AWS or Microsoft — we do not resell tokens.

Can we run this ourselves after a while?

That is the intended ending and the charter is written to make it possible: the decision records, the runbooks and the triage categories are yours and live in your systems. Most teams take the rota over first and keep us on the architecture decisions for a while longer, which is the right order — the queue is easier to own than the judgement calls.

Next step

Name the three people who would be on the rota

Thirty minutes with the architect who would run the engagement. If you cannot name three people with released time, that is the finding, and it is worth more than a proposal.