The AI Audit: 10 Signs Your AI Investment Is Burning Money

Your company spent six or seven figures on AI tools this year. Licences, infrastructure, training, maybe a consultant to help roll it out. Ask your CFO what that spend produced and you will get one of two answers: a vague reference to “productivity gains” or silence.

We walk into these situations regularly. Companies across financial services, healthcare, tech, media, and government that made real investments and got polite adoption but no measurable return. The tools work. The spend is not working. And nobody inside the company is saying it out loud because everyone approved the budget.

Here are ten patterns we see repeatedly. If you recognize five or more, your AI investment is not paying for itself. That is not a judgement. It is a diagnostic. And unlike most problems at this scale, the fixes are specific.

Your most used AI feature is “rewrite this email.”

This is the single most common pattern we see. The company bought enterprise licences. Adoption metrics look fine. But when you look at what people are actually doing, 70% of usage is email polishing, meeting summaries, and paragraph rewrites. These are real features. They are also worth about $8 a month per person, not $80.

If the highest value use case your team has found is making their Tuesday standup notes sound more professional, you have an adoption problem dressed up as a success metric.

You measure usage, not obvious business outcomes.

“83% of our employees have logged in to the AI platform this quarter.” Great. What did they do there? CircleCI found this year that the median engineering team’s output barely moved despite massive tool adoption. The metric that matters is never how many people are using it. It is how many days does the month end close take now? or how fast do claims get resolved? or how many releases ship clean?

Your AI pilot has been “almost ready for production” for six months.

The pilot works. The demo is impressive. It has been impressive since Q3. Nobody can explain what is left before it goes live. The truth is usually one of two things: the pilot was built on clean data that does not exist in production, or there is no internal owner willing to sign off on the operational risk. Either way, a pilot that does not ship is a sunk cost with a slide deck.

You still think that it’s an IT project.

IT rolled out the tools. A few eager people on each team started experimenting. Nobody sat down and asked: which steps of this process should a human do, which should a machine do, and where are the handoffs?

Without that person, what you get is scattered usage. One person on the finance team figures out something useful. Nobody else on the team knows. The insight stays with the individual. Nothing changes structurally. Six months later, that person leaves and the knowledge walks out with them.

Your best people are not using it.

This is the most revealing signal. Your top performers already have systems that work. They are fast, they know the shortcuts, they have muscle memory. Asking them to switch is asking them to get slower before they get faster. Most will not bother unless someone shows them exactly where it fits their existing rhythm.

Anthropic’s data from this week confirms the pattern: experienced users who have figured out how to brief the tools properly are pulling ahead. But most teams never get to that point because the onboarding was a generic training deck, not a workflow specific integration.

AI writes your first draft faster. The five approvals after it still take a week.

The most expensive version of this mistake. AI made step one faster. Steps two through fourteen stayed the same. The total cycle time barely changed because the bottleneck was never the step you automated. It was the review chain, the approval loop, the handoff between teams.

If you have not mapped your process end to end and asked “which of these steps should not exist at all,” you are optimizing inside a structure that was built for a world where these tools did not exist.

You are paying for AI tools on top of the software licences AI should be replacing.

This is the one nobody audits. Your company added AI tool licences to the budget. But it also kept every existing SaaS subscription, every reporting tool, every workflow platform that the AI tools can now handle. Nobody looked at the full stack and asked: which of these do we still need?

We regularly find clients spending six figures a year on software that does less than what a well built AI workflow can do in house. Document processing tools, scheduling platforms, basic analytics dashboards, template generators. These were worth every dollar before. They are not anymore. A custom AI workflow built around your actual process will outperform the generic tool, cost less over time, and eliminate a vendor dependency. The waste is not in the AI spend. It is in the legacy software stack sitting underneath it that nobody has questioned.

Your vendor cannot show you results on your data.

We covered this in our Tuesday piece on agent washing, but it applies broadly. If the only evidence you have that your AI investment works is the vendor’s demo environment and their case studies from other companies, you do not have evidence. You have a pitch. The gap between demo data and production data is where ROI goes to die.

Nobody can point to the ROI. Not a feeling. An actual number.

Ask your CFO, your VP of Ops, your CTO. One question: what is the ROI on our AI spend this year? Not “we think people are more productive.” Not “feedback has been positive.” A number. Cycle time cut from eleven days to four. Error rate down 30%. Claims turnaround improved by two days. Revenue per head up 15%.

If nobody in the room can name a single metric that moved, you do not have an ROI problem. You have a measurement problem. The money went somewhere. Either it produced a result and nobody tracked it, or it did not produce a result and nobody wants to say so. Both are fixable. Neither fixes itself.

You are already planning to buy more tools before the first ones worked.

The vendor roadshow never stops. Every quarter there is a new product, a new capability, a new reason to expand the licence. If your organisation is evaluating the next purchase before it has extracted full value from the last one, you are in a buying cycle, not an implementation cycle. More tools will not fix a usage problem. They will add to it.

The Score

Count how many of the ten you recognized.

1 to 3: Normal growing pains. You probably need a workflow audit, not a strategy overhaul.

4 to 6: Your tools are working. Your processes are not. The ROI is sitting on the table. Someone needs to pick it up.

7 or more: You are funding a line item, not a capability. The fix is not more training or more tools. It is an honest look at how the work is designed and who owns the outcome.

Every one of these patterns is fixable. None of them fix themselves.

If you scored higher than you wanted to, OP can help you figure out where the value is stuck. We do this for companies in financial services, healthcare, tech, media, and government. Start with a conversation.