You signed the engagement six months ago. The new chatbot is live, the document classifier is in production, and the team is using ChatGPT Enterprise daily. You ask your finance lead to summarise AI-related spend year-to-date. The number is twice what you remembered approving. The consultant’s invoice is the smallest line in the file.
This is the year-two surprise that catches almost every SME running its first AI engagement. The headline fee got the attention. The stack of subscriptions sitting underneath the deployed capability quietly built up to something larger than the consulting fee, and most buyers do not see it coming.
Why does the tooling stack grow faster than you notice?
Each piece of AI tooling is small in isolation. £30 a month for ChatGPT Plus, £14 for Power BI Pro, a few hundred for a vector database, a hundred for an automation platform. None of it looks material when approved one at a time, and most of it is approved one at a time. By month four, eight or ten subscriptions sit on the company card, and the aggregate has crossed the consulting fee.
Calling this a procurement failure misses the point. The shape of the spend follows from how AI tooling gets bought, one tool at a time, in response to live problems.
The six categories that make up an SME AI stack
Six categories sit underneath any productive AI capability at SME scale. Large language model licences. Integration and automation platforms. Data infrastructure. MLOps and model management. Monitoring and observability. Security and audit. Each one has its own monthly run rate, and together they typically total £1,500 to £5,000 a month at SME scale before you have added a single in-house engineer.
Large language model licences are the largest category for most SMEs. ChatGPT Enterprise, Microsoft Copilot, Anthropic Claude, or Google Gemini, usually multiple at once because different teams favour different tools. Per-seat pricing of £20 to £75 a month adds up quickly across a 20-person team using two tools each.
Integration and automation platforms are the second. n8n, Zapier, Make, or a similar workflow tool is usually required to wire AI tools into your operational systems. SME-tier pricing runs £50 to £500 a month depending on volume, and the pricing tier creeps as adoption deepens.
Data infrastructure is the third. Cloud storage, a vector database for retrieval-augmented generation, and any data warehousing required to feed the AI capability run £500 to £5,000 a month depending on volume. Most SMEs underestimate the storage and compute costs because the consultant scopes them at pilot volumes, then production volumes are five to ten times larger.
MLOps and model management is the fourth. If you are running custom or fine-tuned models, you need versioning, deployment, and rollback infrastructure. Hosted MLOps platforms run £1,000 to £5,000 a year at SME scale.
Monitoring and observability is the fifth. AI systems behave non-deterministically and need monitoring for accuracy, latency, drift, and cost. A modest monitoring stack runs £200 to £1,000 a month.
Security and audit is the sixth. AI systems handling sensitive data need access controls, audit logging, and ideally cyber insurance uplift. The line is £2,000 to £10,000 a year at SME scale, and it is rarely on the consultant’s proposal.
What the year-two surprise actually looks like
Year one tooling spend is typically smaller than the consulting fee, because the stack is still being built and many tools are on free or trial tiers. Year two is when the surprise lands. Free tiers expire, usage hits paid bands, and the consultant is no longer in the room to keep the stack disciplined.
A typical SME running two or three AI capabilities ends year two paying £1,500 to £5,000 a month aggregate across the stack. That is £18,000 to £60,000 a year, and the consulting fee that selected the stack was probably £15,000 to £25,000 once. The relationship inverts. The advice cost less than what the advice told you to buy.
The Tech Insider 2026 comparison of Tableau Creator at $75 a month per user against Power BI Pro at $14 a month is illustrative. The order-of-magnitude variation in BI tooling alone, before AI is added, is wide enough that tooling choice can move the year-two run rate by a factor of five.
How to budget for the stack on day one
You can avoid the year-two surprise by treating tooling as a budget line at proposal stage, not a discovery surprise at deployment. Three numbers belong in the budget before you sign, and a good consultant will populate them without resistance. A consultant who pushes back on giving you these numbers has either not designed the architecture or is hoping the cost stays invisible.
The first is the year-one tooling cost, broken down by the six categories above. The consultant should be able to give you a per-month estimate within £500 a month accuracy. If they cannot, the architecture is not yet designed.
The second is the year-two run rate forecast. Year two is typically 1.5 to 2 times year one because free tiers expire and usage compounds. Forecast it explicitly so the board approval covers the real shape of the spend.
The third is a quarterly review cadence to drop tools that are not being used. A typical SME six months in has 30 to 40% of its AI subscriptions sitting unused, and a 30-minute quarterly review usually surfaces them.
How to keep the stack from sprawling
Tool sprawl is the dominant pattern in SME AI adoption, and it has a simple cause. New tools get added when they solve a problem. They rarely get removed when the problem stops mattering. Three habits keep the stack in check.
Centralise the procurement decision. Anyone in the team can request a tool, but one person on the leadership team decides whether to add it. That decision-maker holds the run-rate forecast and can see the trend.
Time-box every new subscription. New tools get a three-month review by default, with explicit success criteria for renewal. If the tool is not delivering at the review, it goes.
Rationalise the per-seat licences. Most SMEs end up paying for ChatGPT Enterprise, Copilot, and Claude in parallel because different team members prefer different tools. By month nine, two tools are usually doing the work of three. Drop one.
The AI subscription stack is real spend on real capability. Treating it as a budget line from day one keeps it productive. Treating it as background noise lets it grow into a number that surprises the board in year two.
If you would like to talk about how to scope an AI engagement that builds the stack on a sustainable footing, book a conversation.
