Mid-market firms are increasingly required to make procurement decisions about artificial intelligence capabilities. The choice is no longer binary between building and buying. Three distinct procurement paths have emerged: API-based models from major providers, fine-tuned open-source alternatives hosted internally or via managed services, and embedded automation modules sold as part of existing software platforms. Each path carries a different total cost of ownership (TCO) profile that shifts with usage volume, latency requirements, data privacy constraints, and internal technical capability.
This analysis provides a structured comparison of the three approaches, drawing on observable pricing data, infrastructure cost estimates, and operational considerations relevant to firms with 50 to 500 employees and annual revenues between £10 million and £200 million.
The Three Procurement Paths
API-based models include services such as OpenAI’s GPT-4o, Anthropic’s Claude, and Google’s Gemini. Pricing is per token or per API call, with no upfront infrastructure investment. Providers handle model updates, security patches, and scaling. The buyer pays for convenience and ongoing access.
Fine-tuned open-source alternatives involve taking a base model such as Llama 3, Mistral, or Falcon, adapting it on proprietary data, and hosting it on infrastructure the firm controls or rents. Costs include compute for training and inference, storage, engineering time, and ongoing maintenance.
Embedded automation modules are AI features sold as part of existing SaaS platforms such as Salesforce Einstein, HubSpot Breeze, or Microsoft Copilot. Pricing is typically a per-user monthly add-on or a tiered subscription. The buyer gains integration with existing workflows but cedes control over model choice and data handling.
Cost Comparison by Usage Volume
At low usage volumes — fewer than 100,000 API calls or 10 million tokens per month — API-based models are the cheapest option. The per-call cost of a model such as GPT-4o is approximately £2.50 per million input tokens and £10 per million output tokens. For a firm processing 5 million tokens per month, the monthly API cost is roughly £30 to £60. No infrastructure or engineering overhead is required beyond integration.
At medium usage volumes — 100,000 to 1 million API calls or 50 million to 500 million tokens per month — the cost advantage shifts. API costs scale linearly. At 500 million tokens per month, the API bill reaches £3,000 to £6,000. A fine-tuned open-source model hosted on a single GPU instance from a cloud provider such as AWS or Azure costs approximately £1,500 to £3,000 per month for compute, plus £500 to £1,500 for engineering maintenance. The open-source path becomes cheaper at this scale, provided the firm has the technical capability to manage the deployment.
At high usage volumes — above 1 billion tokens per month — the open-source path is significantly cheaper. API costs exceed £10,000 per month. A dedicated inference cluster with multiple GPUs costs £5,000 to £10,000 per month. The breakeven point varies by model size and optimisation, but the general pattern is clear: volume favours self-hosted open-source models.
Embedded automation modules follow a different cost structure. A per-user add-on such as Microsoft Copilot costs £25 per user per month. For a firm with 200 users, that is £5,000 per month regardless of usage volume. This model is cost-effective for firms with low per-user usage but becomes expensive if usage is concentrated among a small number of power users.
Hidden Costs and Operational Overhead
API-based models carry minimal hidden costs. The main risk is vendor lock-in and price changes. OpenAI and Anthropic have both adjusted pricing over the past 18 months. Firms that build workflows around a specific API face switching costs if pricing becomes unfavourable.
Fine-tuned open-source models carry significant hidden costs. Engineering time for fine-tuning, evaluation, and ongoing monitoring is often underestimated. A typical mid-market firm may need one full-time engineer or a fractional consultant to manage the pipeline. Data preparation and labelling add further cost. Security patching and model updates require ongoing attention. The total operational overhead can add 30% to 50% to the direct infrastructure cost.
Embedded automation modules reduce operational overhead but introduce integration dependency. The buyer cannot easily switch providers without changing the underlying SaaS platform. Data governance is also constrained: the SaaS provider may use customer data for model improvement unless explicitly opted out, and the buyer has limited visibility into how the model is trained or updated.
Why It Matters
Procurement decisions made today will shape mid-market firms’ AI capabilities for the next three to five years. Choosing the wrong path can lock a firm into an unfavourable cost structure, limit flexibility, or expose it to vendor risk. The TCO comparison is not static: as open-source models improve and API prices fall, the breakeven points will shift. Firms that understand the cost drivers can negotiate better terms, time their investments, and avoid overpaying for capacity they do not use.
Commercial Impact
For AI vendors, the mid-market segment represents a large and growing revenue opportunity. API providers benefit from low-volume customers who may never reach the breakeven point for self-hosting. Open-source infrastructure providers such as Together AI, Fireworks AI, and Modal compete on inference cost and ease of deployment. SaaS platforms that embed AI modules gain stickiness and can raise average revenue per user. The commercial battle is over who captures the mid-market procurement budget.
For mid-market firms, the commercial impact is direct. A firm that chooses the wrong procurement path may overspend by 40% to 60% compared to the optimal alternative. The difference between a £5,000 monthly API bill and a £3,000 self-hosted solution is £24,000 per year — material for a firm with a £50,000 AI budget.
Risks / Unknowns
Several factors could alter the TCO comparison. API pricing may fall faster than expected, narrowing the cost advantage of self-hosting. Open-source model quality may improve to the point where fine-tuning is unnecessary for many use cases, reducing the engineering overhead. Regulatory changes, particularly around data sovereignty and AI safety, could impose compliance costs that favour one procurement path over another. The EU AI Act and UK AI regulation are still evolving, and their impact on mid-market procurement is uncertain.
Another unknown is the pace of commoditisation. If inference costs continue to decline at the current rate — roughly 10% per quarter for some providers — the breakeven point for self-hosting may shift to much higher volumes, making API-based models the default choice for all but the largest mid-market firms.
FY Outlook
Over the next 12 to 18 months, we expect mid-market firms to adopt a hybrid procurement strategy. Low-volume, non-sensitive workloads will use API-based models. High-volume, data-sensitive workloads will use fine-tuned open-source models hosted on managed infrastructure. Embedded automation modules will be adopted for specific use cases where integration convenience outweighs cost and control considerations.
The key strategic decision for mid-market firms is not which model to use today, but how to build procurement flexibility. Firms that negotiate volume discounts with API providers, invest in internal capability to manage open-source deployments, and maintain the ability to switch between procurement paths will be best positioned as the market evolves.
Conclusion
Mid-market firms face a genuine procurement choice with material cost implications. API-based models offer low upfront cost and simplicity but scale linearly. Fine-tuned open-source models offer lower per-token cost at scale but require engineering investment. Embedded automation modules offer convenience and integration but limit flexibility and may be expensive for power users. The optimal choice depends on usage volume, data sensitivity, and internal capability. Firms that benchmark their TCO across all three paths will make better procurement decisions and avoid costly lock-in.
