AI tool sprawl: how to tell which of your AI tools to kill

To decide which AI tools to kill, score each one 0–2 on five tests: 28-day active use, a measured outcome, a job no other tool does, a named owner, and company control. Out of 10, kill at 0–4, consolidate at 5–7, keep at 8–10. Zylo’s 2026 index puts unused SaaS licences at 36%, measured against industry-recommended utilisation.

Sources last checked: 23 September 2026. Every external figure on this page links to the publisher that produced it, and was re-read at that source before publication.

• The rubric: the Keep-or-Kill Score, five tests scored 0, 1 or 2, total out of 10 (table below).
• The override: the Double-Zero Rule. Zero on active use and zero on measured outcome means kill at the next notice date, whatever the total.
• The first number to pull: the 28-day active rate, active users divided by paid seats. Copilot’s admin report gives it directly.
• Where sprawl comes from: expense-based SaaS spend rose 267% in a year and ChatGPT is now the most expensed application (Zylo, January 2026).
• The worked example: 400 Copilot seats at a 35% active rate cost US$85.71 per active user a month. Right-sizing to 168 seats saves US$83,520 a year.

What is AI tool sprawl, and how do I know if I have it?

AI tool sprawl is paying for more AI tools than your organisation uses, with two or more of them doing the same job and at least some bought outside procurement. The test is not the count. A 50-person team can be sprawling with four tools.

The Zylo 2026 SaaS Management Index (40 million+ licences) shows how it builds. AI-native application spend rose 108% in a year overall and 393% in organisations with more than 10,000 employees. Large enterprises add an average of 21 applications a month, and business units now control 81% of SaaS spend while IT directly manages 15%.

You have AI tool sprawl if any one of these is true: nobody can produce a list of every AI tool being paid for, two teams bought different tools for meeting notes or email drafting, or AI subscriptions show up on expense claims. AI tool sprawl is a visibility problem before it is a cost problem: you cannot kill a tool you do not know you pay for.

The four causes of AI tool sprawl, and the test that tells them apart

No study ranks these causes by frequency, so they are ordered by how much published evidence exists for each. Each needs its own fix: cutting seats does nothing for a duplicated job.

Evidence for the first two rows: Zylo’s 36% unused-licence average (measured against industry-recommended utilisation) and 267% growth in expense-based spend. Both cover all SaaS, not AI alone.

Which of my AI tools should I kill? The Keep-or-Kill Score

The Keep-or-Kill Score rates every AI tool you pay for on five tests, 0, 1 or 2 each, for a total out of 10. A 1 means the evidence exists but is partial. Score with the tool’s owner, from records, not opinions.

The verdict bands (0–4, 5–7, 8–10), the Double-Zero Rule and the 30% and 60% cut-offs are our own decision rule, not an industry benchmark. Set your own before scoring, never after. Test B uses the same numerator-and-denominator discipline as our guide to which AI efficiency metrics move in weeks and which never move. The Keep-or-Kill Score kills on evidence of non-use, never on a hunch that a tool is not worth it.

How do I tell whether anyone actually uses an AI tool?

Use the vendor’s admin report, or SSO sign-in logs where there is none. Never a staff survey: people over-report tools they feel they should use.

The Microsoft 365 Copilot usage report shows the pattern to look for in any tool. It gives enabled users, active users and an active users rate (active divided by enabled) over 7, 28, 90 or 180 days, plus a per-user last-activity date you can export to CSV. Microsoft counts a user as active only after an intentional action such as submitting a prompt. Opening the Copilot pane does not count. Use 28 days: seven swings with leave, and 180 counts people who tried it once.

For tools without an admin report, count unique SSO sign-ins in the last 28 days and divide by paid seats. For tools that are not behind SSO, the only evidence is the invoice and the expense claim. That is why they score 0 on test E. An AI tool’s active rate is active users over paid seats in a fixed 28-day window, and a survey cannot produce it.

Worked example: a 400-seat Copilot licence at a 35% active rate

This example uses the published Microsoft 365 Copilot enterprise price of US$30 per user a month, paid yearly. The seat and usage counts are illustrative placeholders, so substitute your own.

The full score for this tool, on the stated assumptions: A = 1, B = 1 (time saved is self-reported), C = 1 (marketing also expenses a chat assistant for drafting), D = 1 (IT owns the licence, nobody owns a metric), E = 2 (SSO and an enterprise contract). The total is 6, which puts it in the consolidate band, not the kill band. The action is fewer seats and a named metric owner, not cancellation. An AI licence that fails only on active use needs fewer seats, not a cancelled contract.

What does it take to audit an AI stack yourself?

Auditing an AI stack yourself needs three data pulls, one scoring session per tool and a renewal calendar, with no new software. As a planning estimate, not a measured figure: allow one to two analyst days for the inventory and about 30 minutes per tool to score with its owner. That is roughly 10 hours of scoring for a 20-tool stack.

1. Inventory. Export the SSO application list, then search 12 months of accounts-payable, corporate-card and expense data for AI vendor names. This finds what IT cannot see.
2. Usage. Pull the 28-day active rate for every tool, from the admin report or from SSO sign-ins.
3. Score. Run the five tests with each owner, then apply the thresholds and the Double-Zero Rule.
4. Calendar. List every tool’s notice date and act 60 days before it. Kill order is set by notice dates, not by scores.

The part that goes wrong is the kill itself. Before cancelling, export its data, revoke its access to connected systems and give users the replacement first. Without a replacement, the work can move to personal accounts; the cost of unapproved AI is set out in the AI costs nobody puts in the business case. When you do replace a tool, score the replacement before you sign with the 20-point AI vendor selection score, which covers contract terms this rubric does not.

How LeadsNow scores on its own rubric

LeadsNow provides AI appointment setting for corporate sales teams, so it is one tool in the stack and faces the same five tests, including the ones it can fail.

• A. Active use: there are no seats to count. The equivalent test is whether it produced held appointments in the last 28 days. If it did not, it scores 0, the same as an unused seat.
• B. Measured outcome: we are paid per booked qualified appointment, or a 5–20% share of the sales we help generate, so the billing unit is an outcome. That is still our count. Score 2 only if your own CRM shows closed revenue against your own pre-launch baseline.
• C. Unique job: if your SDR team or another AI setter works the same list, LeadsNow scores 0. Two systems contacting the same prospects is sprawl, whoever built them.
• D. Named owner: someone in sales operations has to own the calendar and the metric. Without that person it scores 0, like any other tool.
• E. Company control: this depends on your contract giving you back the transcripts, recordings and consent records. Check it against what you own when you leave a lead generation vendor.

An outcome-priced service can still fail this rubric. The pricing model settles part of test B and nothing else.

Frequently asked questions about AI tool sprawl

How many AI tools is too many?

No published study sets a safe number, so count jobs, not tools. Any job with two tools paying for it is sprawl, whatever the total. For scale, the Zylo 2026 SaaS Management Index puts the average organisation at 305 SaaS applications (median 240), with AI the fastest-growing category.

What percentage of AI licences go unused?

Nobody publishes an AI-only figure. Across all SaaS, the Zylo 2026 index found organisations leave an average of 36% of licences unused, measured against industry-recommended utilisation levels and drawn from more than 40 million licences. Measure your own AI seats with a 28-day active rate before you assume either way.

Should we cancel the AI tools staff bought themselves?

Not before a sanctioned tool covers the same job. The 2024 Microsoft and LinkedIn Work Trend Index found 78% of AI users bring their own AI tools to work (31,000 knowledge workers, 31 markets). Cancel a card-paid tool with no replacement and the work can move to personal accounts, where you cannot see it at all.

Does having too many sales tools hurt quota attainment?

It is associated with it. In a Gartner survey of 1,026 B2B sellers (January to March 2024), 50% felt overwhelmed by the amount of technology their job needed, and overwhelmed sellers were 45% less likely to attain quota. Gartner groups skills and technology overload together, so the 45% is not technology alone.

When is the right time to cancel an AI tool?

At the notice date before renewal, not the renewal date itself. Microsoft 365 Copilot, for example, is sold as an annual subscription that auto-renews (Microsoft pricing page), so a seat reduction decided a week late costs a full year. Put every AI tool’s notice date in one calendar and score each tool 60 days before it.