How to audit AI tool spend in 30 minutes without integrations

Most companies do not decide to build an AI stack. They buy one tool at a time.

A developer adds a coding assistant. Marketing buys a writing tool. Someone puts a research subscription on a company card. Then a product already in the stack adds a similar AI feature. Each decision can make sense on its own. Six months later, the stack as a whole is hard to explain and even harder to change.

I think teams overcomplicate the first review. You do not need a six-week integration project to ask the basic questions. You need one list, the strongest evidence you already have, and somebody who owns what happens next.

Why start with a manual inventory?

AI buying is increasingly decentralized. Zylo's 2026 SaaS Management Index says business units control 81% of SaaS spend in its dataset, while IT directly manages 15%. It also reports that expense-based SaaS spend increased 267% year over year and that ChatGPT became the most-expensed application.^[1]

Those figures come from Zylo's enterprise customer dataset, not every company, so I would not treat them as a universal benchmark. The operating problem is still easy to recognize: an IT-approved application list can miss employee-card subscriptions, reimbursed plans, and AI features bundled into software already under contract.

Start with what can show an actual charge:

• corporate-card and expense exports;
• accounts-payable vendor records;
• invoices and renewal notices;
• team-level purchase lists;
• known API or usage-credit accounts;
• the software list maintained by IT or Security.

Do not wait for a perfect source. Start with what you have, mark what is unknown, and keep moving.

The 30-minute first-pass audit

Minutes 0–5: Put every known tool in one list

Use one row per product, not one row per charge. Normalize obvious naming variants such as "OpenAI", "ChatGPT", and "ChatGPT Business" only after you confirm whether they refer to the same workspace or separate accounts.

Capture: tool name; team or cost center; base monthly or annual charge; variable credits or usage charges; renewal date, if known; current owner, if known.

Do not estimate savings yet. An unknown cost should stay unknown.

Minutes 5–10: Write down the job each tool performs

Categories are too broad. Two products can both be called "AI assistants" while one is used for code review and the other for customer research.

Write the actual job in plain language: drafts outbound email; searches internal documents; reviews pull requests; creates product images; summarizes sales calls; automates support tickets; provides a general-purpose model workspace.

Overlap becomes visible when two tools serve the same job for the same people.

Minutes 10–17: Add evidence of use

You do not need perfect telemetry for the first pass. Use the strongest evidence available:

• active seats in the last 30 or 90 days;
• usage or credit reports from the vendor;
• a named workflow the team depends on;
• a recent output that reached a customer or internal process;
• an owner willing to defend the tool at renewal;
• a contract, security, or data requirement that makes replacement difficult.

Separate used from valuable. A tool can be opened often without changing an outcome. It can also be used rarely for a high-value task. Record both facts.

Minutes 17–23: Test overlap and bundling

Ask three questions for each tool:

1. Does another paid product already perform the same job?
2. Did a product we already own add this capability to its current plan?
3. If we removed this tool tomorrow, which workflow would stop?

Do not cut a tool just because two feature lists look similar. Product overlap is a lead. Evidence from the people doing the work decides whether it is a safe move.

Minutes 23–28: Give every tool one provisional call

Keep — the tool has a clear job, owner, evidence of value, and acceptable cost.

Cut — the tool has no defensible job or is inactive, and removing it does not create an unresolved operational risk.

Review — the decision is blocked by missing cost, usage, owner, contract, security, or renewal evidence.

Consolidate — two or more tools serve substantially the same job, but the team still needs a replacement or migration decision before anything is canceled.

The word provisional matters. A first-pass score should create the next question, not pretend the decision is already complete.

Minutes 28–30: Assign the next move

Every non-keep call needs one owner, one missing piece of evidence, one due date tied to the renewal or notice deadline, and one next action. Examples:

• Marketing lead: confirm Jasper usage before July 31.
• Finance: verify the annual Notion AI charge and cancellation deadline.
• Engineering: compare active Cursor and Copilot seats before renewal.
• Security: confirm whether the replacement can handle the same data.

Without an owner and date, a cut list becomes another document nobody acts on. That is the part that matters most: what happens next?

The minimum audit table

What not to count as savings

A flagged charge is not captured savings. This is where software audits can lose credibility fast.

If a $300 monthly tool looks redundant, that is $300 of reviewable spend. It becomes captured savings only after the tool is canceled, downgraded, or consolidated and the outcome is confirmed. If the cost is unknown, no savings amount should be counted at all.

That difference matters. The goal is a defensible decision, not the biggest number the audit can put on a screen.

Where AlignCube fits

I built AlignCube around the same evidence this manual first pass uses. You can paste a tool list or invoice into the browser-only preview to see categories and likely overlap without creating an account. A free account includes one full AI-scored audit that gives each tool a keep, cut, review, or consolidate call with reasoning.

The product does not log into your vendors or cancel software for you. It turns the inventory into a decision-ready cut list, then keeps reviewable spend separate from captured savings.

Start with the browser-only preview, then run the free audit if the stack is worth a full review.