Claude Consulting Partner: 12 Selection Criteria (2026)

TL;DR

A good Claude consulting partner proves three things: certified people who use Claude, customers running in production, and a processor contract that survives a GDPR audit. Check the Anthropic directory tier, the cloud route you use, the subprocessor list, the access model and the exit clause. Score every candidate on the same 12 criteria before you talk price.

What does a Claude consulting partner actually deliver?

A Claude consulting partner designs, builds or operates Claude for you: use case selection, cloud setup on Bedrock, Vertex or Foundry, integrations, training, and in some cases ongoing operations. The partner does not replace your contract with Anthropic or your cloud provider. It adds a third party to your data flow, and that is where most selection mistakes start.

We split the market into two jobs because the selection criteria differ:

Job	Typical deliverable	What matters most
Implementation	Landing zone, EU inference profile, first agents, integrations, training	Certified engineers, cloud route experience, references in production
Operations (managed service)	Monitoring, model upgrades, cost control, user support, incident handling	SLA, access model, DPA and subprocessors, exit clause

Many firms do both. Few are equally strong at both. A firm that builds excellent prototypes may have no on-call rota, and a cloud operations shop may never have shipped an agent. Decide which job you are buying first, then weigh the criteria below accordingly.

Anthropic runs a formal program for these firms, the Claude Partner Network. The launch announcement from 12 March 2026 put 100 million USD behind training, technical support and joint market development, and membership is free. A free program with a public directory is useful, but it also means the logo alone says little. The tier says more, as the next section shows.

What do Claude partner tiers and certifications prove?

Partner tiers prove scale and delivery volume, not quality on your project. Anthropic’s Services Track has three tiers: Select, Preferred and Global Premier. Each requires a minimum number of certified people, production customers and public customer stories. Standing is verified quarterly, so a tier is a current signal, not a one-time badge.

The thresholds come from Anthropic’s Services Track announcement of 3 June 2026:

Tier	Active certified individuals	Joint customers in production	Public customer stories	Extra
Select	at least 10	at least 2 (trailing 12 months)	at least 1	none
Preferred	at least 100	at least 15	at least 3	none
Global Premier	at least 1,000	at least 100, in three or more regions	at least 15	joint business plan, named executive sponsors

“Active certified” has a precise meaning: a person holds a current Anthropic certification and has used Claude in the past 90 days. Promotions happen twice a year, on 1 January and 1 July. Certifications belong to individuals, not firms. Since July 2026 there are four role-based certifications: Associate Foundations, Developer Foundations, Architect Foundations and Architect Professional. Exams are proctored and identity-verified through Pearson, and passing candidates receive a Credly badge you can check.

The public service partner directory listed 88 partners when we checked on 1 October 2026. You can filter by region (DACH, France & Benelux, Nordics, Southern Europe, United Kingdom & Ireland and others) and by deployment surface: Claude API, Amazon Bedrock, Google Vertex and Microsoft Foundry.

What we take from this: Select is a meaningful floor for a mid-sized EU company. Ten certified people and two production customers is enough to staff your project. Global Premier tells you the firm is large, not that your project gets its best people. Always ask who exactly is certified and staffed on your account.

A cloud-side credential is a separate signal. AWS runs a Generative AI Competency for which services firms must hold Advanced Tier status first. That says something about Bedrock depth, nothing about Claude specifically.

Which 12 criteria should you score?

Score every candidate on the same 12 criteria, weighted for the job you are buying. The four that eliminate most firms in practice are the data access model, the DPA with its subprocessor list, pricing transparency on tokens, and the exit clause. Tier and references come next. Training capacity and scope fit decide between the finalists.

#	Criterion	What good looks like	How to verify
1	Partner tier and certifications	Listed in the Anthropic directory; named certified people on your team	Directory entry, Credly badges
2	Cloud route experience	Has shipped Claude on your cloud with EU inference (Bedrock EU profile, Vertex eu, Foundry)	Architecture sample, reference call
3	DPA and subprocessors	Article 28 contract, current subprocessor list with locations	Draft DPA before the second meeting
4	Data access model	Works in your cloud account through a role you control, no copies of prompts	Written access concept
5	EU staff and locations	Engineers with access to your data sit in the EU or EEA	Staffing plan with locations
6	SLA	Response and resolution times by severity, named escalation path	SLA annex
7	Pricing transparency	Day rates or fixed fees; tokens billed by your cloud or Anthropic, not marked up	Sample invoice
8	Exit clause	Handover of code, prompts, configs and runbooks; data deletion confirmed	Contract clause
9	References in production	At least two customers live, ideally one in your sector	Reference calls you choose
10	Security certifications	ISO 27001 or equivalent for the partner itself	Certificate and scope statement
11	Training capacity	Can train your users and your engineers, in your language	Course outline, trainer profiles
12	Scope fit	Size and focus match your project, not their sales target	Proposed team and timeline

Two criteria deserve detail.

Pricing transparency. Anthropic’s Commercial Terms (effective 17 June 2025) forbid customers from reselling the Services “except as expressly approved by Anthropic”. A partner that wants to bill you for tokens itself either has that approval or is in breach. Ask which one. Our preference: tokens run through your own AWS, Google or Anthropic account at list price, and the partner bills only for work. Then the EU route premium stays visible. On Bedrock, Anthropic’s own documentation says regional endpoints cost 10% more than global ones for Sonnet 4.5 and later models. Our Claude API pricing comparison has the token tables.

Security certifications. Anthropic itself holds ISO 27001:2022, ISO/IEC 42001:2023 and SOC 2 Type II. That covers Anthropic, not the partner. A partner with admin access to your cloud account is a new attack surface. Ask for its own certificate and read the scope: a certificate that covers only the head office is worth little.

Which red flags should end the conversation?

End the conversation when a candidate wants your data in its own tenant without a clear reason, bills tokens with a markup it cannot explain, has no draft DPA, or cannot name production customers. Each of these signals either a contractual gap or inexperience. None of them gets fixed after signature, so treat them as knock-out criteria.

Our list of knock-outs:

• Tokens resold with markup and no Anthropic approval. The Commercial Terms reserve resale for parties Anthropic approves.
• “We are an Anthropic partner”, but no directory entry. The directory is public. If the firm is not listed, ask why. A pending application is a fair answer. Silence is not.
• Permanent admin access by shared user. AWS explicitly tells customers not to give third parties an IAM user with long-term credentials and to use a role with temporary credentials instead.
• No subprocessor list. Article 28(2) GDPR requires your prior written authorisation before a processor engages another processor.
• Global endpoints by default. A partner who leaves global. profiles in production for personal data has not understood your residency requirement.
• References only as logos. Ask for a call with an operations contact, not a marketing quote.
• Lock-in by tooling. Prompts, agents and evaluation sets stored only in the partner’s proprietary platform.
• One certified person for the whole firm. Certification counts are per person. One badge does not staff a project.

What should you ask in the vendor interview?

Ask questions that force a concrete answer: names, account IDs, clauses, numbers. A good partner answers most of them in the first call and sends the documents within a week. Use the same question set for every candidate and record answers directly into the scoring sheet so the comparison stays honest.

The questions we use, in order:

1. Which of the people on our project hold a current Claude certification? Please send the Credly links.
2. Which two customers run Claude in production with you today, on which cloud, and may we call them?
3. Which EU inference route would you use for our data, and how do you prevent fallback to a global endpoint?
4. How do you access our environment: own account, our account through a role, or a shared user?
5. Please send your draft DPA and current subprocessor list, with locations.
6. Where are the engineers located who can see our prompts or logs?
7. Who bills tokens, and at what price relative to the cloud provider’s list price?
8. What are your response and resolution times for a production outage?
9. Which artefacts do we receive at contract end, and how do you confirm deletion?
10. Which certificate covers your own operations, and what is its scope?
11. Who trains our users and in which language?
12. What part of our scope would you not take on?

Question 4 has a correct answer on AWS. The IAM documentation on third-party access describes it: you create a role in your account, the partner assumes it, and an external ID in the trust policy prevents the confused deputy problem. The trust policy looks like this:

{
  "Version": "2012-10-17",
  "Statement": [{
    "Effect": "Allow",
    "Principal": { "AWS": "arn:aws:iam::PARTNER_ACCOUNT_ID:root" },
    "Action": "sts:AssumeRole",
    "Condition": { "StringEquals": { "sts:ExternalId": "ID_ISSUED_BY_PARTNER" } }
  }]
}

AWS states that the partner’s use of your resources is billed to you, so the permission policy behind this role should be as narrow as the scope. Before handover, check yourself whether prompt logging is active and where it writes:

aws bedrock get-model-invocation-logging-configuration --region eu-central-1

Anthropic recommends logging activity on at least a 30-day rolling basis. If the partner configured logging, the log group or bucket should sit in your account, not in theirs.

For the scoring itself, weight each criterion 1 to 3 and score each candidate 0 to 2. Multiply, sum, compare.

Criterion	Weight (implementation)	Weight (operations)	Candidate A (0-2)	Candidate B (0-2)
1 Tier and certifications	3	2
2 Cloud route experience	3	2
3 DPA and subprocessors	2	3
4 Data access model	2	3
5 EU staff	2	3
6 SLA	1	3
7 Pricing transparency	2	3
8 Exit clause	2	3
9 References in production	3	2
10 Security certifications	1	3
11 Training capacity	2	1
12 Scope fit	3	2

A zero on criterion 3, 4 or 8 disqualifies regardless of the total.

Run it yourself or outsource?

Run it yourself when you already have a cloud platform team that can own IAM, logging, model upgrades and cost alerts, and when Claude serves a handful of use cases. Outsource implementation when speed matters and the skills are missing. Outsource operations only when you can define the SLA, the access role and the exit before the contract.

What operating Claude in-house means, by recurring task:

Task	Role	Frequency
Model and profile upgrades, retirement dates	Cloud or platform engineer	Several times a year
Cost monitoring and budget alerts	FinOps or platform engineer	Weekly
Access reviews, IAM and logging checks	Security	Quarterly
Processor documentation, records of processing	Data protection officer	On every change
User training and prompt guidelines	Product owner or enablement	Ongoing

We do not publish effort figures for these tasks because we have no sourced benchmark. Our estimate: for a single Bedrock or Vertex account with a few internal apps, this is part-time work for an existing platform team, not a new department. It grows quickly with agents that touch production systems.

Outsourcing makes sense when nobody internally can be on call for Claude, when you need certified architects for a first build, or when the rollout covers several countries and languages. It does not make sense when the partner would become the only party that understands your prompts and agents. Then you have bought dependency, not capacity.

What to demand in the contract for this topic: an Article 28 DPA with the processor duties the EDPB lists in its Guidelines 07/2020 (documented instructions, confidentiality, security measures, subprocessor conditions, support for data subject requests, deletion or return at the end, audits). Add a current subprocessor list with locations, a role-based access model in your own account, severity-based response times, and an exit clause that lists the artefacts you receive. The EDPB is explicit that the duty to use processors with sufficient guarantees is continuous. Review the partner at least once a year, not only at signature. Our GDPR provider comparison covers the cloud-side processor chain, and the Claude data privacy guide covers retention.

Disclosure: this site is operated by Pexon Consulting, which offers a managed Claude service and states that it is a member of the Claude Partner Network, although it did not appear in the public directory when we checked on 1 October 2026.

FAQ

What does a Claude consulting partner cost?

There is no published price list. Partners bill day rates or fixed fees for work. Tokens should run through your own cloud or Anthropic account at list price. On Bedrock, EU regional endpoints cost 10% more than global ones for Sonnet 4.5 and later, and that premium belongs on your bill, not inside a partner markup.

Is every Claude partner listed in the Anthropic directory?

Anthropic describes the Partner Hub as a public directory, and listing is one of the benefits partners receive. A firm that claims membership but is not listed may have a pending application or a different arrangement. Ask, and verify on the directory yourself before you rely on the claim.

Select vs Global Premier: which tier should a mid-sized company prefer?

Select requires at least 10 active certified people and 2 production customers, which is enough to staff a typical mid-sized project. Global Premier requires 1,000 certified people and 100 customers across three regions. It signals scale, not better fit. Prefer the tier whose typical project matches yours.

Can a partner resell Claude tokens to us?

Only with Anthropic’s approval. The Commercial Terms state that customers may not resell the Services except as expressly approved by Anthropic. If a partner bills tokens, ask for evidence of that approval and for the price relative to list price.

Does a partner need a DPA with us?

Yes, if it can access personal data in prompts, logs or outputs. Article 28 GDPR requires a written contract with the processor and your prior authorisation before it engages subprocessors. The EDPB adds that the controller should verify the processor’s guarantees at appropriate intervals.

Do partner certifications prove anything about EU data protection?

No. Claude certifications test product and architecture skills. They say nothing about GDPR processes. For that, check the partner’s own ISO 27001 scope, its DPA and its subprocessor list. Anthropic’s ISO and SOC 2 reports cover Anthropic, not the partner.

Sources

1. Anthropic: Claude Partner Network Services Track and Partner Hub (1 October 2026)
2. Anthropic: $100 million for the Claude Partner Network (1 October 2026)
3. Claude: Four role-based Claude certifications (1 October 2026)
4. Claude Marketplace: Service partners directory (1 October 2026)
5. Anthropic Commercial Terms of Service (1 October 2026)
6. EDPB Guidelines 07/2020 on controller and processor (1 October 2026)
7. Claude Help Center: Anthropic certifications (1 October 2026)
8. AWS IAM: Access to AWS accounts owned by third parties (1 October 2026)
9. Anthropic: Claude on Amazon Bedrock (1 October 2026)
10. AWS: Generative AI Competency Partners (1 October 2026)