The real cost of production AI (and why the Big Four overcharge)

You have budget for a production AI deployment. You need to know what it will actually cost and how long it will take. This article gives you a transparent breakdown of what drives pricing and what a realistic engagement looks like for a mid-market business deploying AI across three to five departments.

What drives the cost

Production AI pricing depends on four main variables. Understanding these will help you evaluate quotes from any provider.

Scope of adoption

Deploying AI to one department is different from deploying across five. Each department needs use case mapping, workflow analysis, prompt library development, and hands-on training. The first department takes longest because it includes platform configuration, governance setup, and initial workspace structure. Subsequent departments are faster because they build on that foundation.

Number of custom integrations

Native connectors (Slack, Gmail, Jira) are configured, not built. Custom MCP servers for internal systems are engineering work. A Salesforce-to-internal-database connector is one workstream. Adding your legacy ERP is another. The first custom integration takes longest because it includes the authentication framework and audit trail infrastructure.

Security and compliance requirements

A technology company deploying AI across engineering needs sensible access controls. An FCA-regulated wealth manager deploying to advisory teams needs per-user scoping, seven-year audit retention, data residency guarantees, and compliance documentation. Both are valid. The security engineering for the second scenario is substantially more involved.

Knowledge base complexity

Curating 50 documents into a governed knowledge base is a few days of work. Building an external RAG pipeline over 50,000 documents with continuous sync, hybrid search, and per-document access controls is weeks. Most mid-market clients start with built-in retrieval and add external RAG later when they outgrow it.

What the Big Four charge

Accenture, Deloitte, PwC, and EY have all invested heavily in AI capabilities. Their day rates for AI deployment work typically range from £1,500 to £3,800 per day. At those rates, a modest deployment (three departments, basic integrations, no external RAG) can reach six figures before the first workshop happens.

For large enterprises with complex governance requirements and multi-year transformation roadmaps, the Big Four model may make sense. For mid-market businesses that need their AI platform configured, three custom integrations built, five departments trained, and governance established, it is dramatically overscoped.

What a realistic engagement looks like

Launch: 2 to 3 weeks, fixed fee

Platform configuration (SSO, RBAC, workspaces, audit logging). Integration audit mapping which systems to connect. Department-by-department adoption roadmap. You get a fully configured AI environment and a prioritised deployment plan.

Build: 4 to 8 weeks, project fee

Custom MCP servers for priority systems. Knowledge curation into your governed knowledge base. Hands-on workshops with each department. Prompt libraries built for actual workflows. Native connector configuration and permissions setup. This is where the transformation happens.

Run: monthly retainer, ongoing

New integrations as the business evolves. MCP maintenance. New team onboarding. Compliance reviews. Performance optimisation. Adoption measurement and refinement. Your deployment stays current and keeps delivering value.

Fixed-fee vs hourly billing

Hourly billing creates a perverse incentive: the longer the project takes, the more the provider earns. We use fixed-fee pricing after the Launch phase. Once we understand your systems, data quality, and requirements, we quote a fixed price for the Build phase. If we underestimate the effort, that is our problem. If we overestimate, we refund the difference.

How to evaluate a quote

• Ask for a fixed price after discovery. If a provider cannot commit, they either have not done proper discovery or they are hedging against their own uncertainty.
• Check the team structure. You should know exactly who is doing the work. A programme manager, solution architect, technical lead, and three developers for a three-department deployment? Question it.
• Separate strategy from execution. If 40% of the quote is for strategy and assessment documents, you are paying for output that does not train your team or configure your AI platform.
• Ask about adoption, not just integration. If the quote covers MCP servers but does not include workshops, prompt libraries, and champion programmes, you will end up with connected systems nobody uses.
• Request references from similar-sized organisations. A provider who primarily serves FTSE 100 clients may not be the right fit for a 200-person firm.

The bottom line

Production AI is real work that requires real expertise. It is not cheap and anyone quoting a trivially low number is probably underestimating the adoption and governance requirements. But it also does not require Big Four rates and six-month timelines for mid-market scope. A well-run deployment across three to five departments with custom integrations and proper governance should take six to ten weeks from kickoff to full adoption, at a fraction of Big Four pricing, and it leaves you owning a model-agnostic capability rather than a stack of consultant slideware.