A platform engineer hired three years ago to look after a company’s Azure subscriptions would now be expected to understand token-based pricing, GPU quota requests and content filtering policies, on top of the networking, identity and backup work the job always involved. Very few job descriptions have been rewritten to reflect that. Most teams have simply absorbed the extra load, and some are starting to admit they can’t keep doing so.
The AI services are the most visible source of the strain. Microsoft’s model-hosting offering has been reorganised and renamed more than once since the Azure OpenAI Service first appeared, and each change has brought new portals, new resource types and new ways of doing things that the documentation doesn’t always catch up with straight away. Teams have to decide between pay-as-you-go consumption and provisioned throughput, choose which regions actually have capacity for the model they want, and work out how private networking, key management and logging apply to a service that behaves quite differently from a virtual machine or a database. None of it is impossibly hard. It’s just a lot, and it changes often.
Cost is the second pressure, and it’s arguably the one finance directors notice first. Cloud cost management was already a discipline in its own right before generative AI arrived, with its own practitioners, its own community and a growing body of practice around tagging, reservations, savings plans and chargeback. AI workloads make it harder. Token consumption varies with how people use a product, not just with how many people use it, so a feature change that adds a longer system prompt or an extra retrieval step can raise the monthly bill without anyone touching the infrastructure. GPU-backed compute is expensive enough that an idle cluster left running over a bank holiday weekend becomes a line item somebody has to explain.
Governance is the third, and the least glamorous. Azure Policy, management groups, Defender for Cloud, role-based access and the wider Entra ID estate all need someone who understands how they fit together, and AI services add questions of their own about where data is processed, what gets logged and who can deploy a model into production. For a regulated business, or one selling software to regulated customers, getting those answers wrong has consequences that go well beyond a surprise invoice.
The obvious response is to hire, and some organisations do. But recruiting a specialist for each of these areas is expensive, slow and often unnecessary. A mid-sized software company might need deep FinOps expertise for a few months while it sorts out commitments and tagging, and then only occasional attention afterwards. It might need an AI platform architect to design its model deployment pattern once, not to sit in a permanent role reviewing it. Full-time hires for part-time problems tend to leave, or drift into other work, taking the knowledge with them.
That’s the gap outside Azure help is filling. Some teams bring in contractors for specific projects. Others are turning to Microsoft-focused consultancies for a mix of design work, periodic reviews and ongoing support, on the basis that a firm working across dozens of tenants will have seen the same quota problem or policy conflict before. There’s something to that argument. Breadth of exposure is hard to build inside one company, however talented its engineers are.
It’s still worth being sceptical about what you’re buying, whether that’s a contractor’s day rate or a retainer with one of the consultancies that focus purely on Microsoft’s cloud. Specialist firms vary widely, and those tied closely to one vendor have an obvious incentive to keep you on that vendor’s services, including in cases where a cheaper third-party tool or a simpler architecture would do. Engagements that start as a short review can quietly become a standing dependency, particularly where the consultancy ends up holding the only clear picture of how the environment is configured. The better arrangements build in documentation, infrastructure as code that your own team can read and change, and a deliberate handover of knowledge, so that outside help reduces your reliance on outside help over time rather than increasing it.
There’s also a question the industry hasn’t settled: whether this complexity is a phase or a permanent feature. Microsoft has every reason to simplify the experience, and it has made changes in that direction, folding more AI tooling into shared portals and consolidating how models are deployed. But each simplification tends to arrive alongside new capabilities, new agent frameworks and new pricing models, which bring their own learning curve. If that pattern holds, the skills gap will keep shifting rather than closing. Companies deciding now whether to build or buy their Azure expertise are really deciding which of those moving parts they want to understand in-house, and which they’re content to rent from someone who might, one day, put the price up.
