A CIO signs the contract, seats are provisioned, and Claude, powered by Anthropic, is now technically “in” the organisation. Eight months later the pilot that impressed the board is still running in one team’s sandbox, security has frozen a wider rollout pending a review nobody scoped, and half the seats are unused. None of that is a product failure. It is what happens when a licence is treated as the whole adoption plan.
The thesis of this piece is simple: a Claude licence gives an enterprise access to the model; it does not give the enterprise the governance, integration, evaluation, and change management needed to run that model reliably in production, and closing that gap is precisely the work a Preferred Services Partner in the Claude Partner Network is built to do.
This is not a hypothetical gap. Anthropic’s own framing of its partner programme makes the distinction explicit: as the company put it when it introduced the Services Track, “a successful pilot is not the same as a system a business can run on.” That line is the whole argument in miniature, and it is worth taking seriously before signing off on a Claude rollout as a licensing decision alone.
Why licences alone stall at the pilot stage
The industry data on enterprise AI adoption is unusually consistent, and unusually blunt. A S&P Global survey of more than 1,000 enterprises found that 42% of organisations abandoned most of their AI initiatives in 2025, up sharply from 17% in 2024, and that 46% of AI proofs of concept failed before they ever reached production. The same analysis puts the overall AI project failure rate at over 80%, roughly double the failure rate of non-AI technology projects. The stated cause is rarely the model. It is that “model accuracy does not solve access, workflow, data, and accountability problems,” and teams discover the missing controls too late to fix them without a restart.
Separate research into what practitioners call “pilot purgatory” quantifies the cost of that gap. Nearly half of enterprise AI licences go unused, costing large organisations an average of $80.6 million a year, according to Incedo CEO Nitin Seth. The same source reports that 86% of organisations lack visibility into how AI is actually moving through their systems, that one in five has already had a security breach linked to unauthorised AI use, and that compliance overhead alone can add roughly 17% to total AI system costs when it is bolted on after the fact rather than designed in from the start.
That last point matters most for CIOs and COOs weighing a Claude rollout. Compliance and governance are cheaper when they are part of the deployment plan than when they are retrofitted onto shadow usage that has already spread across the business. Unmanaged deployments, sometimes called shadow AI, create duplicate data stores, orphaned integrations, and parallel review paths that fragment accountability well before anyone tries to formalise a rollout.
What “enterprise adoption” actually involves
A licence answers one question: does the organisation have contractual access to Claude models. Running Claude reliably at enterprise scale requires answering several more, none of which come bundled with the API key:
- Security and governance review. Data handling, access controls, and audit trails need to satisfy the organisation’s existing compliance regime (SOC 2, ISO 27001, sector-specific rules such as those in financial services or healthcare) before Claude touches production data, not after.
- Integration into the existing SDLC and data platform. Claude has to sit inside version control, CI/CD, ticketing, and the data infrastructure teams already use, rather than becoming a parallel tool that duplicates work or bypasses review.
- Prompt and context engineering standards. Without agreed standards for how prompts, system instructions, and retrieved context are built and versioned, quality varies team to team and nobody can debug a regression.
- Evaluations and quality assurance. Production use needs a repeatable way to measure whether outputs are accurate and safe for the task, not a one-off demo that impressed a steering committee.
- Cost and usage governance. Token spend needs the same forecasting and chargeback discipline as any other infrastructure cost, or the licence itself becomes an unmanaged budget line.
- Change management and training. Teams need to understand what the model is reliable for and where it is not, and that understanding has to be maintained as usage spreads beyond the first pilot group.
- Ongoing model upgrades. Anthropic ships new model generations on its own schedule (for example, moving from Claude Sonnet 4.5 to Claude Opus 4.1 for a given workload). Each move requires re-testing prompts, evaluations, and integrations rather than assuming behaviour carries over unchanged.
Every one of these is an operating capability, not a licensing term. An enterprise can buy all the access it wants and still lack the muscle to do any of the above at scale.
What a Preferred Services Partner adds
This is the specific role Anthropic built the Claude Partner Network to fill. Anthropic’s Services Track groups partner firms into three tiers based on demonstrated production work, not sales volume: Select, Preferred, and Global Premier. Anthropic has described the role of these partners as handling “the integration, the evaluation, and the way people’s work evolves,” which is a fair summary of everything a licence does not include.
Zartis is a Preferred Services Partner in the Claude Partner Network. In practice that means the governance review, SDLC integration, evaluation framework, and change management work described above are not left for the enterprise to invent from scratch or improvise mid-rollout. They are a scoped, delivered engagement that runs alongside the licence, informed by having done this work across other production deployments rather than treating each client’s rollout as the first one.
It is worth being precise about what this is. This sits alongside the staff augmentation work Zartis also delivers, but it is a distinct engagement: an AI transformation partnership that advises on the operating model, the governance structure, and the evaluation approach, and then delivers the integration and enablement work required to put that plan into production. The advisory and delivery functions are one engagement, not two separate relationships.
The decision in front of the CIO
A Claude licence is the start line, not the finish line. The data on pilot purgatory, shadow AI, and abandoned initiatives all point to the same operational gap: the organisations that stall are the ones that scoped the licence and never scoped the governance, integration, and evaluation work that determines whether the licence pays for itself.
The practical question for a CIO or CTO evaluating a Claude rollout is not “do we have access to the model.” It is “who owns the security review, who owns the evaluation framework, who retrains the teams when Anthropic ships the next model generation, and who is accountable when usage spreads beyond the pilot group.” If those questions do not have answers attached to names and a delivery plan, the rollout is still at the licensing stage, whatever the seat count says.
If your organisation is scoping a Claude rollout and those ownership questions are still open, that is the conversation worth having before the next model generation ships and the pilot has to be re-tested from scratch.