AI Spend Is Now an Operations Problem. Here Is How to Manage It.
Managing AI costs across your whole organisation: a practical spend taxonomy, 5 control levers, and ROI measurement for 20-50 person European firms.
TL;DR: Managing AI costs across your whole organisation: a practical spend taxonomy, 5 control levers, and ROI measurement for 20-50 person European firms.
Twelve months ago, your firm had one or two AI subscriptions. Today you have seven or eight, spread across three departments, on three different billing cycles, with two on personal employee cards that get expensed monthly.
Why this matters: AI spend has crossed the threshold where a spreadsheet and good intentions are no longer adequate governance. A 35-person professional services firm running Anthropic API, Microsoft 365 Copilot, n8n, Notion AI, and Jasper simultaneously faces a cost control problem, a vendor overlap problem, and an ROI measurement problem, all at the same time. The total monthly bill is often 40 to 60 percent higher than the finance team realises, because usage-based API costs, per-seat overages, and shadow subscriptions are invisible until the card statement arrives. This article gives operations leaders and finance managers a practical framework for taking back visibility and control. It does not require a dedicated procurement team or new software. It requires a taxonomy, five control levers, and a quarterly review habit.
The AI Spend Visibility Problem
Most small and mid-sized businesses discover their true AI spend only when something goes wrong: an unexpectedly large API bill, a duplicate subscription found during an audit, or a renewal auto-processed on a card that nobody monitors.
Three structural factors create this visibility gap.
Shadow subscriptions. Staff sign up for AI tools individually using personal or team cards. The subscription is expensed or absorbed into a team budget line. Finance sees a vague "software" entry. No central register exists. In a 30-person firm, it is common to find six to ten AI subscriptions that the operations or IT lead does not know about until they specifically go looking.
Mixed pricing models. Foundation model APIs (Anthropic, OpenAI, Google) charge on token consumption. Productivity tools (Microsoft 365 Copilot, Notion AI) charge per seat per month. Workflow automation platforms (n8n, Make, Zapier) charge on operation counts or task volume. Each model has different cost drivers, different overage behaviours, and different forecasting requirements. Aggregating these into a coherent budget line is not straightforward.
Invisible API costs. When a developer or a no-code operator builds a workflow that calls an LLM API, the cost does not appear in a standard SaaS subscription line. It appears in an API billing account, often paid on a credit card attached to an individual developer's account. If that workflow scales, or runs more frequently than expected, costs can increase by multiples before anyone notices.
A Three-Category Spend Taxonomy
Before you can control AI spend, you need a consistent way to classify it. This taxonomy covers the full range of AI tools a 20-50 person firm is likely to be running.
Category 1: Foundation Models and APIs. Direct API access to large language models or other AI services, billed by consumption. Examples: Anthropic API, OpenAI API, Google Vertex AI, Azure OpenAI Service. Cost driver: token volume, request count, or compute time. Risk: unbounded costs if no hard spending limits are set at the API account level.
Category 2: AI Productivity Tools. End-user tools with AI features baked in, typically on per-seat subscription pricing. Examples: Microsoft 365 Copilot, Notion AI, Grammarly Business, Otter.ai, Perplexity Pro. Cost driver: seat count, sometimes with usage caps or overage charges above a monthly allowance. Risk: licence creep as seat counts grow without regular audits.
Category 3: AI Workflow Automation. Platforms that orchestrate multi-step workflows involving AI processing, typically billed on task volume, operation counts, or tier-based plans. Examples: n8n, Make (formerly Integromat), Zapier with AI steps, Relevance AI. Cost driver: task or operation volume, which scales with automation usage. Risk: cost spikes when automations are triggered more frequently than modelled, or when new workflows are added without cost review.
This taxonomy gives you three buckets to track separately in your finance system, with different review cadences and different control mechanisms for each.
Five Spend Control Levers
Lever 1: Central Procurement Gate
Any new AI tool above a defined threshold (a reasonable starting point for most operations leaders is 50 euros per month) must go through a central approval step before purchase. This does not need to be a formal procurement committee. It needs to be a named person, a simple request form (tool name, use case, monthly cost, department, alternative considered), and a five-day turnaround commitment. The gate stops accumulation of shadow subscriptions and creates a record that feeds your vendor register.
Lever 2: Usage Dashboards for Category 1 Spend
For every foundation model API account, set hard spending limits at the account level. Most providers (Anthropic, OpenAI, Azure) support monthly spend caps that cut off API access rather than allow unlimited overcharge. Set the cap at 120 percent of your expected monthly budget. In addition, review the API usage dashboard weekly for the first three months after any new workflow goes live, then monthly once spend is stable.
Lever 3: Per-Department Budget Lines
Assign AI spend to the department that owns it, not to a central IT or software budget. When the marketing team's AI writing tool subscription is on the marketing budget, the marketing lead has a direct incentive to manage seat counts and usage. This is more effective than central oversight alone, because the department lead has context on whether the tool is actually being used.
Lever 4: Quarterly Review Cadence
Schedule a 60-minute AI spend review each quarter. The agenda: (1) review current spend by category against the prior quarter; (2) identify any tool with zero or near-zero usage in the past 90 days; (3) identify any tool where spend has grown more than 30 percent without a corresponding business case; (4) assess vendor overlap (are two tools doing the same job?); (5) confirm that all active tools are on the approved vendor register. This review takes less time than a single contract negotiation and prevents the compound drift that makes AI spend unmanageable.
Lever 5: Vendor Consolidation Criteria
Define in advance what triggers a consolidation review. Useful criteria: two or more tools in the same category serving the same use case; any single vendor relationship above 500 euros per month that has not been re-evaluated in 12 months; any Category 3 tool where actual task volume is less than 40 percent of the plan ceiling. Consolidation is not always the right answer, but having explicit criteria prevents consolidation decisions from being driven by contract renewal pressure rather than operational logic.
ROI Measurement for Non-Technical AI Investments
API cost measurement is relatively straightforward: input plus output tokens times price per thousand. ROI measurement for productivity tools and workflow automation is harder, and most operations leaders either skip it or rely on vendor-supplied benchmarks that overstate real-world impact.
A practical approach for growing businesses and founder-led companies uses three measurement types.
Time displacement. For each AI tool, identify the specific task it replaces or accelerates. Estimate the time saved per week per active user. Multiply by the average fully-loaded cost per hour for that role. If Notion AI saves a project manager two hours per week on meeting notes and status updates, and that manager costs 60 euros per hour fully loaded, the weekly value is 120 euros against a monthly seat cost of 10 to 15 euros. This calculation is rough but defensible and prevents subjective "we feel more productive" assessments.
Error rate or rework reduction. For compliance-adjacent tasks (contract drafting, financial reporting, client communication), track whether AI-assisted outputs require less rework than manual outputs. A reduction in revision cycles has a measurable time cost attached to it.
Revenue contribution. For sales-adjacent tools (AI-assisted proposal writing, outreach personalisation), track whether conversion rates or proposal volumes change after tool adoption. This is the hardest measurement but the most credible one for justifying Category 2 spend to a CFO or board.
When AI Spend Warrants a Dedicated Operations Process
A spreadsheet and a quarterly review is sufficient governance for most finance teams and operations leaders up to around 500 euros per month in total AI spend. Above that threshold, or when any of the following conditions apply, a more structured operations process is warranted.
- More than five active AI vendor relationships, each with separate billing accounts and renewal dates.
- Any Category 1 (API) spend that varies by more than 25 percent month-to-month without a clear operational explanation.
- AI tools integrated into client-facing workflows where a billing or access failure would interrupt service delivery.
- Staff count growing faster than your ability to audit seat-count accuracy on per-seat tools.
In these cases, the operations process should include a dedicated cost centre or budget code for AI spend, a named AI spend owner (this is typically the COO, operations director, or IT manager, not the CFO), and a vendor register that is reviewed and updated at least quarterly.
Frequently Asked Questions
How do we handle AI tools that staff are already using on personal accounts?
Issue a structured disclosure request as an amnesty process: list the tools you use for work, including personal accounts, with an estimate of monthly cost. Offer to migrate active tools to company accounts where the business case is clear. This surfaces the real spend picture without a punitive dynamic. Pair the amnesty with a fast-track approval process so staff have an easy path to legitimise tools they rely on.
What is a reasonable total AI spend for a 35-person professional services firm?
A reasonable working range for a 30-40 person professional services firm actively using AI in core workflows is 800 to 2,000 euros per month across all three categories. Firms significantly above this range without documented ROI for each tool are likely carrying redundant or underused subscriptions. Firms significantly below this range in 2026 may have a competitive disadvantage in throughput relative to peers who have invested.
Should we centralise all AI procurement in IT or keep it distributed by department?
A hybrid model works best at this scale. Central procurement gate for approval and vendor registration. Department ownership for budget and active management. Central quarterly review for consolidation and ROI oversight. Full centralisation in IT creates bottlenecks and disconnects tool ownership from the people who know whether the tool is actually delivering value.
