When AI seems too expensive, what else concerns businesses?

Among the 547 eligible businesses citing high AI costs, 197 also reported both a lack of staff with the required skills and difficulties with data availability or quality. Using the survey weights, that is an estimated 37.5%. The unweighted share is 36.0%.^[1][8]

Both percentages use the same businesses and answers. The unweighted figure counts every business equally. The weighted figure gives each record the contribution assigned by the survey design, so it is our main estimate. Weighting changes the estimate, not which answers count as reporting both concerns.

The 95% sampling interval runs from 29.8% to 45.2%. That range is a reason to describe the result as roughly four in ten, rather than suggest that 37.5% is an exact measure of every similar business.

Because all three answers come from the same record, the analysis measures the overlap directly. We can see whether the businesses raising a cost concern also explicitly reported both staff and data concerns.

It also leaves the other responses open. A business outside the confirmed-both group might report one of these concerns, neither, or an unknown answer. The complement cannot be described as a group whose only problem is price. The questionnaire contains other reasons too, and our headline does not rank them.

The percentage describes this historical Brazilian survey group. A business elsewhere can use the distinction to frame its own research question.

The cost question asks whether costs seem too high. It does not record a supplier quotation, the amount a business could afford, or which expense the respondent had in mind. We therefore refer to a perceived cost concern. The answers cannot tell us how far prices would need to fall, or whether a discount would change a decision.^[2]

The staff question concerns people inside the business who have the skills to use the technologies. The questionnaire first screens for an IT area, department or specialist. Having that support does not itself establish whether the people involved have the skills needed for a proposed AI task.

The data question combines availability and quality. A yes answer could concern whether suitable information exists, whether it is usable, or both. Because the survey does not separate those possibilities, it cannot establish which kind of preparation a particular business needs.

The seven AI-use questions cover more than generative AI. They include technologies such as text analysis, speech recognition, machine learning and workflow automation. The result therefore concerns that wider list of technologies.

Any unknown staff or data answers remain in the full cost-concern group. Our main percentage counts explicit confirmation of both concerns; it does not fill unanswered questions with a guess. Eligibility also allows unknown or unanswered AI-use responses, so “did not report using” should not be read as confirmed nonuse.

For a business owner or marketing lead, the practical implication is to keep three questions separate: what the proposed task costs, who can judge the work, and which information it needs. The observed overlap gives a reason to ask those questions together. To investigate AI implementation challenges in your own business, start with a defined task and evidence from the people who would do the work.

Start with a description of the work that someone could check. “Use AI in marketing” is too broad to reveal what the team needs. “Group incoming product questions into topics for the sales team to review” names an output and the people who will use it. The narrower task makes it possible to discuss the price of something specific.

Then ask what the proposed price includes. A tool subscription and the staff time needed to prepare information or review an output are different entries in a plan. This is a suggested planning distinction, not a breakdown measured by the survey. A team can record those entries without assuming that an unfamiliar task will be expensive or that it will save time.

A question about review can be more useful than a general question about confidence with AI. Who would recognize a wrongly grouped inquiry? Which source would they consult when an answer is uncertain? Who decides whether the output is usable? These questions identify responsibilities for a proposed task; they do not diagnose a person’s skills from a survey statistic.

Keep the next step proportionate to the decision. A business might need to clarify a quotation, locate an authoritative product document or ask a colleague to review a small example. Choose the check that answers the specific unresolved question before committing to a larger project.

Consider a fictional regional wholesaler whose sales team receives product inquiries through its website, email and social messages. A manager proposes using AI to group those inquiries into topics. The team wants to see which questions deserve clearer product information and which should reach a salesperson. This is an illustrative planning example, not a business observed in the survey.

The first discussion stalls at the price. Instead of deciding that cost explains everything, the manager writes down the proposed output: a list of inquiry topics, linked to the original messages for a person to review. The proposal does not include sending automatic replies or making product recommendations to customers.

The sales lead notices that similar product names could be confused. She agrees to review the grouping against the current catalog and to flag an inquiry when its meaning is unclear. The relevant question becomes whether she has the information and familiarity needed for that review, rather than whether someone in the company is generally “good with AI.”

A colleague checks the information the task would use. Product descriptions live in more than one document, and some inquiry messages contain customer details. The team decides to identify the current product source and check which messages it may appropriately use before preparing an example.

The manager can now ask for a more specific proposal. What does the tool do, what must a person check, and what information must be prepared? The team records those tasks alongside the quoted price.

The next decision is whether a limited review can answer the remaining questions. The team might proceed, revise the task or defer it. The manager records the reason for that decision and what new information would prompt another review.

Copy this template into an existing project brief or meeting note. No Influence account is needed. Complete it for one proposed task.

1. Intended task: [specific work] → [output someone can inspect] → [person or team using it]. Write what is outside the proposal as well as what is included.

2. Cost question: [quotation or estimate being discussed] · [what it includes] · [what is still unknown]. Record staff preparation or review work separately if it is relevant; do not invent a saving to justify the proposal.

3. Review responsibility: [named reviewer] · [what they will check] · [source they can consult] · [what happens when they are unsure]. A job title alone may not answer these questions.

4. Required information: [source documents or records] · [who maintains them] · [whether they are current and appropriate to use] · [unresolved question about availability or quality].

5. Small next check: [question to resolve] · [owner] · [evidence that would inform the decision]. Describe the check before interpreting its result.

6. Decision record: [proceed, revise or defer] · [reason supported by the check] · [remaining uncertainty]. Keep a concern distinct from evidence that the concern has been resolved.

In the fictional wholesaler example, the output is an inquiry-topic list, the sales lead reviews it, and the current catalog supplies the product reference. The next check concerns whether that arrangement produces something the sales team can assess. Those entries make the proposal more concrete without borrowing a success rate from an unrelated survey.

We used the public, confidentiality-treated TIC Empresas 2025 microdata from NIC.br/Cetic.br. Interviews took place in Brazil from February 2025 to January 2026. The survey covers active formal private enterprises with at least ten people working in them, within its defined sectors.^[1][3][4]

Our analysis starts with the questionnaire’s IT-area, department or specialist screen. We then keep businesses with no affirmative answer across the seven listed AI technologies and a yes answer to the perceived-cost question. The main calculation divides the survey weight of businesses explicitly reporting both staff and data concerns by the weight of this entire eligible cost-concern group. Any unknown staff or data answers remain in that denominator.

We calculated sampling uncertainty using the supplied 200 combined bootstrap replicate weights. The intervals describe uncertainty for individual estimates, not every source of possible error. We display only results meeting our predefined support and disclosure rules; omitted breakdowns should not be treated as zero.^[8]

We fixed the research question and calculations before examining the joint results. During source checks, we documented an inconsistency in released response codes outside our analysis group and corrected an overly precise comparison with published reference tables. The download records what changed and when. The changes did not alter the analysis population or the target calculations. An independent calculation reproduced the published result.^[8]

The public file has undergone confidentiality treatment. Cetic.br checks its utility against the published tables, but that does not certify that every new joint analysis is identical to one using the original interviews.^[7]

The general idea that staff, data and financial concerns can coexist is established. Earlier research grouped firms by combinations of barriers, and the 2025 producer report discusses people and data among implementation requirements. Our contribution is a direct estimate of selected concerns reported together within this specific cost-concern group, not the discovery of a new cause of non-adoption.^[5][6]