Untrained AI users cost more and deliver less. Education fixes both

Barry Hatfield is CloudSmith’s Chief Information Officer. (Image: AI-enhanced)

---

Right now, the loudest message in business technology is simple: get your people using AI. Boards want it, vendors are selling it, and every second LinkedIn post celebrates another productivity win. Adoption is racing ahead, leaving skills to rapidly fall behind.

But there is a quieter conversation happening in CIO offices. AI spend is arriving at a rate nobody planned for. It doesn't show up as one large line item. It shows up as subscriptions that grow, API bills that double, and consumption charges that no one can quite explain at month-end. For South African organisations, the pressure is sharper, as most AI services are billed in US dollars, so every unplanned token is also exposed to the exchange rate. 

The root cause is only partly the technology. AI illiteracy plays a major role: people using powerful tools without knowing how to use them well. What I'm seeing, both inside CloudSmiths and across our clients, is that an untrained user costs more and delivers less, and the fix is quite simply – education. Users who know which tool to use for which output will allow organisations to drive costs down and quality up at the same time.

AI is priced by consumption. Every prompt, every document pasted in, every long back-and-forth is measured in tokens, and tokens cost money. No single request is expensive. The problem is the aggregate – thousands of small, unoptimised interactions every day, across every team.

Model choice is the clearest example. In a tool like Claude, choosing between a lighter model such as Haiku and a more powerful one such as Opus or Fable is a simple dropdown. The price difference between them is significant. Most users have no idea it exists, and even fewer know which model suits the task in front of them.

So they default to the biggest model for everything. Summarising a meeting, drafting an e-mail or reformatting a table rarely needs the most capable model available. Matching the model to the use case is a skill which can be taught. Once users learn it, the same work costs less and often comes back faster.

In CloudSmiths' analysis, one finding really surprised the company. When it looked at usage internally and within some of its clients’ organisations, some of the heaviest AI users were not producing the best results. High token consumption did not equal high output. In many cases, it pointed to a skills gap. The wrong model, poorly structured prompts, or AI being used for work it isn't suited to. Organisations are falling into the trap of focusing on improving the technology, without bringing their users along for the ride. The people are being left behind, and with them, the value opportunity.

AI literacy is not only about prompting. It is knowing where AI adds value, and where it just adds cost.

For judgment-style work, AI excels: interpreting, drafting, summarising, reasoning through ambiguity. But we often forget that it is probabilistic, not deterministic. Ask the same question twice and you may get two different answers. That makes it a poor substitute for a well-defined process.

The risk is subtle. Instead of AI driving a process, people start using it to create process on the fly, every time. A process that should be automated once, reliably and cheaply, gets re-designed by a large model on every run. That is expensive, inconsistent and hard to audit.

Literate users know where this line sits. They use AI where judgment is needed, conventional automation where the steps are known, and a lighter model where a heavier one adds nothing. That is how output goes up while spend comes down.

Illiteracy has costs beyond the bill. Untrained users also create new risks the CIO must manage:

• Cyber security. What data are users pasting into AI tools and where does it go?
• IT governance. Who is allowed to use which tools and models, for what, and with what oversight?
• Optimisation. Are we getting measurable value or just measurable consumption?

At the same time, no CIO wants to be the one who slowed the organisation down. Locking AI away is not an option. Missing the train is a bigger risk than the bill.

Education is the main lever, and governance makes it stick. In CloudSmiths' view, that means moving away from generic, open-ended front ends towards governed solutions built for specific use cases. When a tool is designed for a job, the right model, data access and guardrails are set once, centrally. Educated users then know when to use it, and get a better result for less.

CloudSmiths feels this problem more sharply than most, because it is deeply entrenched in AI. It has highly skilled technical teams who optimise AI solutions for clients every day. Then there is the company's internal team, which looks a lot like its clients' users. Usage and cost vary hugely between them, depending on skill level and use case.

That is why CloudSmiths is driving its teams to certify with Anthropic. And it is not just technical or implementation staff. The leadership is certifying too.

Dr W Edwards Deming put it well: "It is not enough to do your best; you must know what to do, and then do your best." For me, that applies directly to AI. A leader can't govern what they don't understand. If you want to set sensible policy on models, usage and spend, you need hands-on knowledge of how these tools actually behave.

The opportunity is to close the skills gap before the cost of it compounds. A few practical starting points:

1. Make usage visible. Track consumption by team, tool and model, not just total spend.
2. Look past the heavy users. High usage is a signal to investigate, not a badge of honour. Check whether output matches cost.
3. Teach model selection. Show people when a lighter model is good enough, and set sensible defaults so they don't have to choose every time.
4. Separate judgment from process. Use AI where it adds judgment. Automate the repeatable steps conventionally.
5. Govern by use case. Replace open-ended access with purpose-built, governed solutions for your highest-volume tasks.
6. Lead from the front. Make sure your leadership team is as AI literate as you're asking everyone else to be.

AI illiteracy may be the most expensive line item nobody tracks. The organisations that win won't be the ones that use the most AI. They'll be the ones whose people know how to use it well.