At the end of May 2026, enterprise AI search company Glean announced it had passed $300 million in annual recurring revenue, about three times the figure from 15 months earlier. TechCrunch’s headline pointed to its main selling point now: helping companies cut their AI budgets. CEO Arvind Jain said customers really like that Glean can significantly reduce their AI bill; his argument is that when AI gets the information it needs through Glean’s organized view of company data, it consumes far fewer tokens than reading each system directly.
That claim may hold, or it may not apply to your company. A tool that looks like it can consolidate things doesn’t automatically save money. If old tools can’t be shut off, data still needs cleaning, permissions need resetting, and staff need retraining, a new tool may just be one more subscription.
So the question before buying is which cost it actually replaces. “Savings” count in only two cases: an old expense can be canceled on a specific date, or a measurable work metric actually drops. Anything else is an experiment cost.
The savings percentage in the deck was calculated on someone else’s terms
The savings figures in a pitch are usually calculated under other companies’ conditions. Whether your company has the same old tools to shut off, whether your data is in order, and whether staff will really change how they look for information all change the outcome.
Even more often underestimated is the cost of adoption itself: new licenses and deployment fees, getting data into a state AI can use, redefining who can see what, training staff, and handling wrong answers. Most of these are new expenses that don’t disappear just because the tool is smart.
Put four kinds of cost in one table
Tool licenses: what you currently pay for search, knowledge bases, and document assistants, and how much they’re actually used. Only shutting off an old tool counts as savings.
Data cleanup and permissions: the staff time spent organizing documents and maintaining permission groups. AI search usually increases this, because AI exposes messy data faster, and misconfigured permissions can let the wrong people see things.
Staff time: time spent looking for information, answering the same questions repeatedly, and switching between tools. Measure the current state first; it only counts as savings if you see a drop after the trial.
Risk and maintenance: wrong answers, outdated data, and permission leaks all need someone responsible. That person can’t be treated as zero cost.
With the table in hand, decide with three questions: which old expense can be canceled on a specific date? Which work metric will drop, and by how much counts as success? Who is responsible for data quality, permissions, and wrong answers? Only if all three can be answered is a pilot worthwhile; if not, don’t commit to an annual contract.
Measure the current state before talking trials
The FinOps Foundation, which promotes cloud financial management, lists “business value drives technology decisions” as a principle. Applied to AI tools, it means stating what value you expect in return before deciding how much to spend.
Before a trial, spend a week or two measuring the current state: how much time staff spend finding information each week, and how often the same questions get asked. Without baseline numbers, the end of the trial can only be judged on feel. For setting budget guardrails on AI usage bills, see With Copilot on Usage-Based Billing, Decide First Whether to Stop or Pay When Credits Run Out.
The value this table can’t capture
This table counts visible costs. Some value is hard to quantify, such as new hires getting up to speed faster, or knowledge no longer living only in a few people’s heads. These can go into the evaluation but shouldn’t be converted directly into “money saved.”
Also, if your existing tools are scattered but cheap, have clear owners, and are easy to fix, there’s no rush to switch. AI search suits teams that already have a clear information-finding pain point, high costs from repeated queries, and the capacity to clean up data and manage permissions. For who should handle cost anomalies, see AWS FinOps Agent Can Find Why Costs Spiked; Finding the Owner Depends on What You Wrote Down First.
AI handoff card
We're evaluating an enterprise AI search or knowledge tool, and the vendor says it will save us money. I want to confirm whether the savings hold up first.
The method comes from this article: https://boosterminiclass.com/en/posts/enterprise-ai-search-needs-cost-audit/
Its view: "savings" count in only two cases: an old expense can be canceled on a specific date, or a measurable work metric actually drops. Anything else is an experiment cost. Adoption's own licenses, data cleanup, permission setup, and training are mostly new expenses.
First ask me: which search, knowledge base, and document tools we use now, the cost and usage of each, and what savings the vendor promises.
Then:
1. Build a cost table in four categories, tool licenses, data cleanup and permissions, staff time, and risk and maintenance, listing current cost, new costs after adoption, and whether each can be canceled;
2. Mark which savings are "certain" (a clearly cancelable expense), which are "possible" (need a metric drop to prove), and which are only "imagined";
3. Recommend baseline metrics to measure before the trial, and the threshold for trial success;
4. List what needs a named owner: data quality, permissions, wrong answers.
Mark anything without numbers as "needs data" rather than estimating it yourself.
References
- TechCrunch: Glean’s top line crosses $300M as AI budget-cutting becomes its major selling point — https://techcrunch.com/2026/05/28/gleans-top-line-crosses-300m-as-ai-budget-cutting-becomes-its-major-selling-point/ (2026-05-28)
- Glean official press page: Glean Surpasses $300M ARR: Unrivaled Enterprise Context Fuels AI Adoption — https://www.glean.com/press [accessed: 2026-10-11]
- FinOps Foundation: FinOps Principles — https://www.finops.org/framework/principles/ [accessed: 2026-10-11]
