How to Calculate Your Real AI Spend (And Why You're Probably Wrong)

Most businesses know they're spending money on AI. Almost none of them know how much.

It's not because they're careless. It's because AI spend doesn't behave like normal software costs. There's no single line item. No single invoice. No single person who owns it.

It's scattered across departments, buried in expense reports, and split between subscriptions that nobody audits and API usage that nobody tracks.

Here's how to find the real number.

Why AI spend is invisible

Traditional software is easy to track. You pay Salesforce £X per seat per month. It shows up on one invoice. One person approves it. Done.

AI doesn't work like that. It arrives in your business through three completely different channels, and most companies are only tracking one of them — if that.

The first channel is subscriptions. ChatGPT Plus, Claude Pro, Copilot, Jasper, Grammarly — individual seats that people sign up for, often on their own credit cards. These show up as expenses but rarely get categorised as "AI spend." They're buried under "software" or "tools" or just expensed as miscellaneous.

The second channel is API usage. Your developers are calling OpenAI, Anthropic, or Google APIs directly. This is pay-per-use — billed by the token, by the minute, by the call. It shows up on a cloud bill or a separate provider invoice, and unless someone is actively watching it, the number creeps up without anyone noticing.

The third channel is embedded AI. Tools your team already uses that have quietly added AI features — Notion, Canva, HubSpot, Slack. You're paying for AI without realising it because it's bundled into existing subscriptions. The cost is hidden inside a price increase you accepted without reading the changelog.

When you add all three together, the number is always bigger than anyone expected.

Step one: audit your subscriptions

Start with what people are paying for individually. This is the easiest to find and usually the most wasteful.

Check expense reports for the last three months. Search for ChatGPT, Claude, OpenAI, Anthropic, Copilot, Jasper, Grammarly, Midjourney, ElevenLabs, Perplexity. Check personal credit card claims. Check company card statements.

Ask every department: "What AI tools are you paying for?" Don't just ask the tech team — marketing, sales, HR, and operations are all using AI tools now.

What you'll typically find: multiple people paying for the same tool individually when a team plan would be cheaper. People paying for tools they signed up for three months ago and haven't used since. Overlapping tools that do the same thing — two different writing assistants, three different image generators.

It's not uncommon to find four people each paying for ChatGPT Plus at £20 per month — £80 total. A Team plan would cost £48. Same access, 40% savings, and nobody had thought to consolidate.

Step two: find your API usage

This is where it gets harder. API costs are invisible to anyone who isn't a developer, and often invisible to the developers too.

If your team uses AI APIs, someone has an account with OpenAI, Anthropic, Google, or similar. Log into each provider's dashboard and check the usage page. Look at the last 90 days. Note the total spend, which models are being used, and how many tokens are flowing through.

Key things to look for: are your developers using GPT-4o for everything when GPT-4o-mini would work for most tasks? Are there API keys that haven't been used in weeks but are still active? Is anyone running expensive models for simple tasks like formatting or summarisation?

The model choice matters more than most people realise. GPT-4o costs 15 times more than GPT-4o-mini. For email drafts, FAQ responses, data extraction, and summarisation — the cheap model works just as well. But developers tend to pick a model once, confirm it works, and never revisit the decision.

A team running 100 API calls per day through GPT-4o is spending roughly £480 per month. The same calls through GPT-4o-mini would cost about £30. That's £450 per month in savings for zero quality loss on most tasks.

Step three: check your embedded AI

This one is subtle. Go through your existing software subscriptions and check which ones have added AI features recently. Many tools have introduced AI tiers or AI add-ons in the last 12 months — sometimes as a free trial that silently converts to paid.

Look at your Notion subscription — are you on a plan that includes Notion AI? Your Canva subscription — does it include Magic Write and AI image generation? Your CRM — has it added AI features to a higher tier you've been upgraded to?

These aren't usually large individual costs, but they add up. And they represent AI usage that nobody is tracking, auditing, or assessing for compliance.

Step four: calculate the real total

Add it all up. Subscriptions plus API usage plus embedded AI equals your actual monthly AI spend.

In most SMBs, the real number is two to three times what the business owner thought it was. Not because anyone is being reckless — but because the spend is distributed across so many channels that nobody has ever seen the full picture.

Common traps

The wrong model trap. Developers default to the most capable (and expensive) model because it works. Nobody checks whether a cheaper model would produce the same result for that specific task. For 70% of typical business AI tasks, it would.

The unused seat trap. Someone signs up for a Pro plan during a busy week, uses it intensively for a few days, then never touches it again. The subscription renews quietly every month.

The duplicate tool trap. Marketing uses one AI writing tool, sales uses another, and customer support uses a third. All three do essentially the same thing. A single tool with a team plan would cost less than one of the individual subscriptions.

The personal account trap. Employees using their personal AI accounts for work. They expense it, or they don't. Either way, company data is flowing through an account that IT doesn't know about, finance can't track, and nobody is auditing for compliance.

The "it's only £20" trap. Every AI subscription sounds cheap in isolation. £20 here, £15 there. Multiply by 15 employees and suddenly it's £300 to £500 per month that nobody approved as a category of spend.

Your AI spend audit checklist

One — Search expense reports for AI tool subscriptions across all departments.

Two — Ask every team what AI tools they use, including free trials and personal accounts.

Three — Log into every AI provider dashboard and check 90-day API usage.

Four — Review existing software subscriptions for embedded AI features and tier upgrades.

Five — List every AI tool, its monthly cost, who uses it, and whether it overlaps with another tool.

Six — Identify model downgrades that wouldn't affect output quality.

Seven — Consolidate individual subscriptions into team plans where possible.

Eight — Calculate total monthly AI spend across all three channels.

What to do with the number

Once you have it, two things happen. First, you'll probably want to cut it — and the audit will have already shown you where. Consolidate subscriptions, downgrade models, cancel unused seats.

Second, you'll realise you need to keep tracking it. AI spend doesn't stay static. New tools get adopted, usage patterns change, developers switch models, employees sign up for the next shiny thing. A one-off audit helps, but ongoing visibility is what actually controls the cost.

That's what SpendLil does. For API-based AI, it tracks every request automatically — per key, per provider, per model, in real time. No spreadsheets, no chasing developers for usage reports, no monthly audits. Connect once and the spend data builds itself.

It won't track your SaaS subscriptions yet — that's on the roadmap. But for the API channel, which is usually the fastest-growing and least visible part of AI spend, it's handled from the moment you connect.

And with the EU AI Act landing in August 2026, that audit trail isn't just about cost control anymore. It's about compliance.

Start the audit this week. You'll be surprised what you find.