Software sprawl is an old problem wearing a new badge. Ten years ago it was project management tools. Five years ago it was video conferencing. In 2026 it is artificial intelligence, and it is spreading faster than either of those did, because the buying decision has moved out of procurement and onto individual expense cards.
The pattern is predictable once you have seen it. A writer expenses ChatGPT Plus. A developer adds Claude Pro because it handles long files better. Someone in research signs up for Perplexity. A designer needs Midjourney. Marketing wants Gemini for the Workspace integration. None of these purchases is wrong on its own. Together they form a line item nobody owns and nobody reviews.
This article is about measuring that line item honestly, understanding why the waste is structural rather than careless, and deciding what to do about it without banning the tools your team actually depends on.
What c actually costs
Start with arithmetic rather than anecdote. The four mainstream assistants each sit at roughly the same price point:
| Tool | Standard plan | Typical buyer |
|---|---|---|
| ChatGPT Plus | $19.99/mo | Everyone |
| Claude Pro | $20/mo | Writers, developers |
| Gemini AI Pro | $19.99/mo | Google Workspace teams |
| Perplexity Pro | $20/mo | Research, competitive analysis |
Four subscriptions is roughly $80 per person per month, or about $960 a year. Add a specialist tool or two — an image generator at $30, an AI code editor at $20 — and a single knowledge worker can carry $130 a month in AI licences. For a fifteen-person team that is well past $20,000 annually, spread across enough individual expense reports that no single approval ever looked large enough to question.
The number itself is not the problem. Plenty of teams would happily pay $130 per head for tools that genuinely produce $130 of value. The problem is utilisation. Surveys of AI-heavy knowledge workers consistently find that people use a small fraction of the features they pay for, and that most of the actual work concentrates in one or two tools while the rest sit idle as insurance against the day the primary tool gets something wrong.
That is the real shape of AI sprawl: you are not paying for six tools, you are paying for one tool and five fallbacks.
Why the sprawl happens, and why blaming people does not fix it
It is tempting to treat this as a discipline problem. It is not. Three structural forces produce it, and none of them responds to a policy memo.
1. Models genuinely differ, and the differences are task-shaped
If every assistant were equally good at everything, nobody would keep a spare. But they are not. One model handles long documents without losing the thread. Another writes cleaner code. A third is better at recent information because it searches the live web more aggressively. A fourth refuses fewer edge-case requests. These are real capability gaps, and the person keeping two subscriptions is responding rationally to them.
2. Vendor pricing is designed to be unbundleable
Every major assistant is sold as a flat monthly seat with a generous-sounding allowance. That structure punishes exactly the behaviour that would save you money — using each tool a little, for the thing it is best at. A user who sends thirty messages a month to Claude and thirty to ChatGPT pays full price twice. There is no partial seat, no shared pool, no way to pay for the slice you use.
3. Nobody owns the category
Design software has an owner. Cloud infrastructure has an owner. AI assistants, in most organisations under a few hundred people, have no owner at all, because the spend arrived through the back door one $20 charge at a time. Unowned categories do not get audited.
How to measure your own sprawl in one afternoon
Before deciding anything, get a real number. This takes less time than most people expect.
Step one — pull the charges. Export the last three months from your card provider or expense tool and filter for the obvious vendors: OpenAI, Anthropic, Google, Perplexity, xAI, Microsoft, Midjourney, and whichever code assistants your engineers favour. Include personal cards if your reimbursement policy allows AI purchases, because that is where the quiet spend lives.
Step two — separate seats from usage. Flat monthly seats and metered API bills behave completely differently. A $20 seat costs $20 whether it is used once or a thousand times. A metered bill scales with real work. Mixing them in one total hides the waste.
Step three — get a per-head figure and compare it to the alternative. Divide your monthly AI total by headcount, then check that number against what a consolidated plan would cost. An AI subscription cost calculator does this comparison directly: you select the plans your team is actually on, and it returns the monthly total, the annual total, and the delta against a single unified subscription. It takes about two minutes and turns an argument about instinct into an argument about a number.
Step four — ask who logged in last month. Most vendors expose last-active dates in team billing. Seats that have not been touched in thirty days are not fallbacks; they are forgotten.
The three consolidation strategies, honestly compared
Once you have the number, you have three real options. Each has a genuine cost.
Strategy A: Standardise on one vendor
Pick a single assistant and cancel the rest. Cheapest to administer, simplest to explain, and the easiest to get approved.
The cost is capability. You inherit that vendor’s weaknesses permanently — its refusal behaviour, its context limit, its blind spots on recent events. And you inherit its roadmap. If a competitor ships something transformative in March, you find out in a Slack thread and cannot act on it until your renewal.
Fits: small teams with a narrow, consistent workload.
Strategy B: Route to a multi-model platform
Replace several seats with one subscription that provides access to models from multiple labs, with a shared allowance across all of them. This is the structural fix rather than the disciplinary one, because it removes the thing that caused the sprawl: the requirement to buy a whole seat from every vendor whose model you occasionally need.
Platforms in this category — Perspective AI is one of the established examples — bundle the major model families into a single account, let you switch between them inside one conversation, and meter usage from a shared credit pool rather than charging you a full seat per lab. The practical effect is that keeping a fallback stops being expensive, so people stop hoarding subscriptions.
The cost is dependency and, occasionally, latency on brand-new releases: you are trusting one platform to keep its roster current. Check how quickly new model versions appear before committing.
Fits: teams whose work spans several task types — writing, code, research, analysis.
Strategy C: Move to metered API access with an internal interface
Buy raw API access and put your own thin client in front of it. You pay only for tokens consumed, which for light users is dramatically cheaper than a seat.
The cost is engineering. Somebody has to build and maintain chat history, file upload, permissions, and a usable interface, and that person’s time is not free. This is a real answer for teams with spare engineering capacity and a wrong answer for everyone else.
Fits: technical organisations with low per-person usage and someone who wants to own it.
What to do in the first month
A sensible sequence, in order of effort:
- Name an owner. One person is accountable for AI spend. Without this, nothing below survives the quarter.
- Cancel the dormant seats. Anything untouched for thirty days goes. This is free money and requires no debate.
- Run the calculator. Establish the consolidated baseline so the comparison is concrete.
- Pilot with the heaviest users, not the lightest. The people running six tools are the ones who will find the gaps in a consolidated plan fastest. If it holds for them, it holds for everyone.
- Write down the exceptions. Some specialist tools genuinely should stay — a video generator, a domain-specific coding assistant. Document why, and revisit each quarter rather than never.
- Set a review date. AI pricing changed several times in the last eighteen months. A decision made today should be re-examined in six months, not treated as permanent.
Frequently asked questions
How much does the average professional spend on AI subscriptions? Estimates in 2026 cluster around $110–$150 per month for a heavy individual user, covering three to five assistant subscriptions plus one or two specialist tools. Light users land closer to $40. The variance within a single company is usually larger than the variance between companies.
Is consolidating to one AI vendor risky? It concentrates risk in a way that is easy to underestimate. Any single vendor’s outage, price rise, policy change, or capability gap becomes yours in full. Multi-model access reduces that exposure, which is a legitimate reason to prefer it beyond pure cost.
What is shadow AI? AI tools bought and used inside a company without IT or security approval — typically on personal cards or free tiers. It is the main reason AI spend is routinely underestimated, and the main reason sensitive data ends up in accounts nobody is monitoring.
Do consolidated AI platforms give you the same models as buying direct? Usually the same model families, sometimes with different rate limits or a short delay on brand-new versions. Verify roster and update cadence against your actual requirements before switching, rather than assuming parity.
The takeaway
AI tool sprawl is not a failure of discipline. It is the predictable result of buying per-vendor seats for a category where the useful behaviour is switching between vendors. You can fight that with policy and lose slowly, or change the purchasing structure and stop generating the problem.
Either way, start with the number. Pull three months of charges, divide by headcount, and compare it to the consolidated alternative. Most teams discover the gap is larger than they guessed and easier to close than they feared.