How to find the AI subscription waste hiding in your company
The fastest way to find AI subscription waste at a mid-market company is to pull a 12-month export of every SaaS charge over $200, filter for anything that mentions AI, ML, GPT, copilot, assistant, or generation, and then ask one question of every line: who owns this and what specific outcome are we measuring. About half the line items will not have an answer. That is your waste pile. In every audit I have run for companies between $25M and $500M in revenue, the waste comes in at 30-60% of total AI spend. The dollar range I see most often is $40K to $200K per year sitting in tools nobody is accountable for, nobody is measuring, and in many cases nobody is actively using. This paper is the framework I use, the six patterns I find every time, and the consolidation playbook that gets the money back without losing a single productive workflow.
Why mid-market AI spend is bloated by 30-60% by default
Mid-market companies are uniquely exposed to AI subscription bloat for four structural reasons, and none of them are about anyone being careless. They are about how procurement actually happens.
The first reason is department-level procurement. Marketing buys a writing assistant. Sales buys a meeting note-taker. Customer Success buys a sentiment tool. RevOps buys a forecasting copilot. Each charge is below the threshold that triggers central review, usually $5K or $10K annually, so each one slides through on a department card or a manager-level approval. By month nine the company has 14 AI tools, and no single person has ever seen the full list.
The second reason is no central inventory. At companies under 500 people, there is rarely a SaaS management platform like Vendr, Zylo, or Torii in place. The system of record is a spreadsheet maintained by whoever in finance had time last quarter, and it is always behind. When I ask a CFO to send me their AI tool list, what I get back is partial. The audit always surfaces 20-40% more line items than the company knew it had.
The third reason is the free trial autorenewal. AI vendors aggressively offer 14-day or 30-day trials. An employee signs up, evaluates, decides not to push for a paid rollout, and forgets. The credit card on file gets charged on day 31. Three months later the charge is buried in a corporate card statement and the original employee has moved teams or left.
The fourth reason is personal subscriptions on company cards. ChatGPT Plus at $20/month, Claude Pro at $20/month, Perplexity Pro, Midjourney, ElevenLabs, dozens of point tools at $15-$50 per seat per month, all expensed individually. At a 100-person company, if even 30 employees have one personal AI tool on their company card, that is $7,200 to $18,000 a year in fragmented spend that nobody is governing. Some of it is high value. A lot of it is duplicate.
None of these are pathologies. They are the default state when AI procurement runs faster than AI governance, which is the default state at every mid-market company I have worked with.
The six waste patterns I find in every audit
I have done enough of these now that the patterns are predictable. If you are reading this and want to do a self-audit before bringing in outside help, these are the six places to look first. They will account for 80-90% of the recoverable spend.
Pattern 1: Multiple seats on multiple tools doing the same thing
The most common waste pattern at mid-market is paying for two or three tools that do the same job. The classic example is writing assistants. I will find Grammarly Business, Jasper, Copy.ai, and ChatGPT Team all active on the same company, with overlapping seat lists. Each was bought by a different department for a different perceived use case. Marketing bought Jasper for campaign copy. Sales bought Copy.ai for outreach. HR bought Grammarly Business for internal comms. IT bought ChatGPT Team for general productivity. Annual cost across all four at a 150-person company: roughly $42,000. Realistic consolidated cost using one tool with broader rollout: $14,000-$18,000. Recovered: $24,000-$28,000 per year.
The same pattern shows up with meeting note-takers (Otter, Fireflies, Fathom, Read, Granola, sometimes Gong on top), with image generators (Midjourney, DALL-E via OpenAI, Adobe Firefly seats nobody activated), and with chat assistants (ChatGPT Team plus Microsoft Copilot plus a separate Claude subscription).
Pattern 2: Underutilized enterprise plans
This one stings the most when I show the CFO. A company will buy a 100-seat enterprise plan because the vendor offered a discount at that tier, the assumption being the rollout will catch up. Six months later, 12 seats are active. The other 88 seats are paid for and unused. Every AI vendor I know of has a usage dashboard accessible to the admin. Pull it. The active-seat number is almost always lower than what was procured, sometimes by an order of magnitude.
The fix here is rarely complicated. Drop to the actual tier. The vendor will fight, but most will negotiate down at renewal, especially if you bring a usage report and a polite alternative-vendor note to the call. I have seen single-line savings of $24,000-$60,000 from one renegotiation on one tool.
Pattern 3: AI features bundled into existing SaaS that nobody uses
This is the silent one. Salesforce Einstein, HubSpot AI, Notion AI, Microsoft Copilot in M365, Google Gemini in Workspace, Slack AI. Most of these are add-ons that got toggled on during a renewal, often pushed by the vendor as a free trial that converted to paid. The company is paying $5-$30 per seat per month for AI capabilities inside a tool the employees use, and zero employees know the AI is there or how to invoke it. At a 200-person company, a $10/seat/month bundled AI add-on is $24,000 a year for capability that is not getting touched.
The audit move is simple. For every SaaS line item, ask the vendor or your account rep: is there an AI add-on attached to this contract, when was it added, and what does the usage telemetry show. If usage is below 20% of seats, kill it at renewal.
Pattern 4: Trial-to-paid escalations on tools with no clear owner
This is the credit card cleanup pattern. Pull every AI-related charge under $500/month from the corporate card statements for the last 12 months. For each one, ask: who initiated this, are they still here, is the tool still being used, and is there any business outcome attached. About a third of these charges will fail at least one of those questions. Kill them.
I have seen companies with $8,000-$15,000 a year in this category alone, spread across 25-40 different micro-charges. Each individual line is too small to flag in finance review. The aggregate is real money.
Pattern 5: API spend with no monitoring
If your company is using OpenAI, Anthropic, or any other LLM API directly, this pattern is almost always present. Engineering or a vendor stood up an integration, often a chatbot, an internal search tool, or a document summarizer. The token spend started low. Six months later, it is $4,000-$12,000 a month. Nobody is watching the cost dashboard. Nobody can articulate the business value the integration is producing. The classic case is the internal vanity dashboard that summarizes Slack channels for an executive who looks at it twice a week.
The fix is to require, for every API integration, a quarterly cost-to-outcome review. Tokens consumed, dollars spent, decisions made or hours saved. If a team cannot defend the ratio, the integration gets paused, not killed, and the team has 30 days to justify resumption with a new use case. Pausing is more politically palatable than killing, and 60% of the time the team never asks to resume.
Pattern 6: Vendor pivot tax
The AI vendor landscape has been moving in 12-month cycles. Whatever framework or pricing model you bought into last year is probably not what the vendor is selling now. Pricing has shifted from per-seat to consumption-based at several vendors. Tools that were premium in 2024 are commodity now. Frameworks that were hot in early 2025 (RAG-as-a-service, autonomous agent platforms) have either consolidated, pivoted, or quietly raised prices on existing customers to subsidize their next round.
The audit move is to recheck the public price page on every AI vendor in your stack against what you are paying. If the vendor has launched a new tier and is no longer actively selling the SKU you are on, you have leverage. Either they move you to the new tier with equivalent or better terms, or you leave. Companies leave more often than I expected when I started doing this work. The switching cost on most AI tools is low because the underlying capability is increasingly commoditized.
The five-question framework for every AI line item
This is the audit framework. For every line item in your AI stack, get an answer to all five. If you cannot get answers, that line item is a candidate for cancellation.
Question 1: Who owns this. A specific person, not a department. If the answer is "marketing uses it" or "sales bought it," the answer is no one. Tools without owners do not get measured, do not get optimized, and do not get cancelled when they should be.
Question 2: What is the use case. One sentence, specific. "We use Jasper to generate first drafts of campaign emails, which our marketing manager then edits." Not "productivity" or "writing." If the use case cannot be stated in one sentence with a verb and an output, the use case is not real.
Question 3: What is the alternative cost. What would it cost to do this work without the tool? If the realistic alternative is ChatGPT Team at $25/seat/month, and you are paying $80/seat/month for a specialized tool, the delta needs to be justified by something measurable. Often it cannot be.
Question 4: When do we re-evaluate. Every AI tool should have a calendared re-evaluation date, ideally 60 days before renewal. Without this date, renewals happen by autopilot and you lose the negotiating window. I tell every client to put renewal dates on the same shared calendar finance uses for tax deadlines.
Question 5: Who is measuring outcomes. The owner from question 1 should be reporting, at least quarterly, on what the tool is producing in business terms. Hours saved, output volume, quality scores, deals influenced. If the owner cannot produce a number, the tool is not being managed, it is being subsidized.
The org chart problem nobody wants to talk about
Here is the structural fix that most companies skip and then end up back in the same waste position 18 months later. AI procurement happens at the department level. AI governance has to happen at the executive level. These are not the same place in the org chart.
What I recommend at every audit close-out is the appointment of a single executive accountable for the AI stack as a portfolio. Not the CIO by default, because at most mid-market companies the CIO is not close enough to the line-of-business AI use cases. Often the right owner is the COO or a VP Operations who already owns SaaS spend governance. Sometimes it is the CFO directly, especially at sub-$100M companies where the CFO is deep in operational decisions.
The owner does three things. First, they maintain the central inventory of every AI tool, owner, and use case. Second, they hold the renewal calendar and the kill/keep decision. Third, they enforce the pre-purchase checklist below before any new AI tool gets bought. Without a single accountable executive, the waste rebuilds itself within a year because the structural drivers are unchanged.
Real numbers from real audits
To make this concrete, here are three redacted engagements from the last 14 months. Names, industries, and exact figures are sanitized but the shape of each audit is real.
Audit A: 100-person professional services firm, $87K annual AI spend. Three writing assistants overlapping (Pattern 1, $18K consolidated), one underused enterprise plan dropping from 75 to 25 seats (Pattern 2, $11K saved), bundled AI in their CRM nobody used (Pattern 3, $4K killed at renewal), and one zombie API integration generating a Slack summary nobody read (Pattern 5, $1K/month paused). Total recovered: $34K. Productivity impact at 90-day check-in: zero. Headcount usage of remaining tools went up because everyone now had one tool that worked instead of three tools they were unsure which to use.
Audit B: 220-person SaaS company, $156K annual AI spend. Heavier engineering footprint, so API spend was the big lever. Three internal LLM integrations had no cost ownership (Pattern 5), and one of them was hitting GPT-4 instead of a cheaper model for a task that did not need it. Total API rationalization saved $58K annualized. Add Pattern 1 consolidation on the go-to-market side ($22K) and an enterprise plan rightsizing ($14K). Total recovered: $94K.
Audit C: 380-person manufacturing distributor, $214K annual AI spend. The biggest single line was a forecasting AI tool sold by their ERP vendor. After a usage audit, only the supply chain director was actively running the tool. We negotiated to a single-user tier and saved $48K. Pattern 3 (bundled AI add-ons) was unusually heavy here because the company had stacked AI features on top of HRIS, CRM, and ERP, none of which were used. Total recovered: $140K.
The pattern across all three: somewhere between 30% and 65% of the original AI spend was recoverable, and in every case the productivity impact at 90 days was either neutral or positive because consolidation reduced tool fatigue.
The three-step consolidation playbook
Once the audit is done and you have your kill/keep list, the consolidation itself takes 30-60 days if done right. Here is the sequence.
Step one: Freeze new AI procurement for 60 days. No new AI tools, no new seats on existing tools, no auto-renewals approved without executive sign-off. This is mostly a political move, not an operational one. It signals to the org that AI spend is now governed.
Step two: Execute the cancellation list in order of contract expiration. Do not cancel mid-term unless the vendor has materially changed terms or the line item is under $500/month. For the rest, calendar the cancellation 30 days before auto-renewal and execute. Send the cancellation in writing, get the confirmation, follow the credit card statement for two billing cycles to confirm the charge stopped.
Step three: Roll out the consolidated tools with intent. If you are killing three writing assistants and standardizing on one, you need a 30-minute internal training, a designated power user as in-house support, and a single Slack channel for questions. This is the step companies skip, and it is why consolidation sometimes fails. The consolidated tool needs to feel better to the user than the three separate tools did, or they will go around procurement and put a new $20/month subscription on their company card.
The pre-purchase checklist for every new AI tool
The audit is the one-time fix. The checklist is the recurring discipline. Before any new AI tool gets purchased, the requesting team has to provide answers to seven questions. If they cannot, the request is denied or sent back for more work.
- What specific business outcome will this tool produce. One sentence with a measurable output.
- Do we already have a tool that produces this outcome. Cross-check against the central inventory. If the answer is yes, the burden is on the requester to explain why the existing tool is insufficient.
- What is the cheapest viable alternative. Often the answer is "ChatGPT Team or Claude for Work that we already pay for." If the specialized tool is more than 2x the cost of the generalist, the marginal value has to be defended.
- Who is the named owner. A person, not a team. The owner is on the hook for usage and outcome reporting.
- What is the trial period and what is the success metric. Most AI tools should be trialed for 30-60 days before paid commitment. The success metric should be set in writing before the trial starts.
- What is the renewal date and the kill criterion. Calendared 60 days before renewal. The kill criterion is the threshold below which the tool gets cancelled, not renegotiated, at renewal.
- Has finance and the AI portfolio owner approved. Two signatures required for any AI tool over $5,000 annual spend. Below that threshold, single signature from the portfolio owner.
The five-step audit checklist you can run this quarter
If you want to run the audit yourself before bringing in outside help, this is the sequence I would follow. It will take a senior operations person about 20-30 hours over two weeks.
- Pull the 12-month spend export. Every charge over $200 from corporate cards, ACH, and accounts payable. Filter for AI, ML, GPT, copilot, assistant, generative, and the names of the major vendors (OpenAI, Anthropic, Microsoft Copilot, Google AI, Jasper, Copy.ai, Grammarly, Otter, Fireflies, Midjourney, Perplexity, ElevenLabs, etc.).
- Build the inventory spreadsheet. One row per tool, columns for vendor, monthly cost, annual cost, contract end date, named owner, use case, active seats, total seats, and bundled-in-other-tool flag.
- Run the five-question framework on every row. Mark each tool as keep, renegotiate, or kill. Be ruthless on tools without an owner or a one-sentence use case.
- Group by waste pattern. Use the six patterns above to cluster the kill list. This makes the consolidation conversation easier because you are killing categories, not individual tools, which feels more strategic and less like personal defeat for the original buyers.
- Present to the exec team with a single dollar number. "We have $X in recoverable AI spend, here is the kill list, here is the consolidation plan, here is the new pre-purchase policy." The dollar number is what gets approval. Everything else is implementation detail.
If you run this audit and recover less than 20% of your AI spend, you are unusually well-governed and you can stop reading. If you recover 30% or more, which is the typical outcome at mid-market, the audit pays for itself the first month after consolidation.
Where to go from here
If you want a second set of eyes on your AI stack, my AI Advantage Audit is a structured engagement that runs the framework above against your actual line items, your actual contracts, and your actual usage telemetry. The deliverable is a CFO-defensible memo with the kill list, the consolidation plan, and the pre-purchase policy customized to how your org buys software. I do not sell software, I do not take vendor referral fees, and I do not have a dog in the fight when I tell you to cancel something. The math is the math.
If you want to run the audit yourself first, the framework in this paper is everything you need. The hardest part is not finding the waste. The hardest part is the political work of cancelling tools that someone in the org championed nine months ago. That is the conversation the executive owner has to be willing to have, and it is the difference between a one-time savings and a sustained governance discipline.