Small businesses are not waiting for a perfect AI strategy. They are already using AI for practical work, but they draw a clear line around tasks involving compliance, money, employees, and legal exposure. That is not resistance. It is sensible risk pricing.
American Banker reported findings from a Stax Payments survey of 500 U.S.-based small and midsize business technology decision-makers conducted in June 2026. Nearly all respondents had tried AI in some form. Stax reported positive perceptions around decision speed, confidence, and customer service, while adoption was concentrated in lower-risk activities.
Because Stax is a payments company and the survey reports respondent perceptions, the results should not be treated as independent proof of ROI. They still provide a useful map of where businesses feel comfortable and where trust breaks down.
What happened
Marketing and content creation was the most common reported use case at 61%, followed by graphic design at 49% and business planning at 45%. Adoption was much lower for work with stronger legal or compliance consequences: human resources and accounting or tax preparation each stood at 25%, legal services at 22%, and insurance at 16%.
The contrast matters. Businesses are willing to use AI where a person can review an output cheaply and correct a mistake before it causes harm. They are more cautious where an error can affect payroll, taxes, employment decisions, contracts, customer funds, or regulatory obligations.
The survey also found that 80% expected their organization's AI use to increase during the following six months. That means the sensitive-work gap is unlikely to remain empty. Vendors, banks, payment companies, ERP providers, and consultants will try to fill it with integrated products. The business question is how to move from experimentation to controlled use without turning a productivity tool into an invisible decision-maker.
Why it matters for growing businesses
The risk is not simply whether an AI answer is accurate. Sensitive workflows combine data quality, authority, deadlines, and evidence. An accounting assistant may classify a transaction correctly but use the wrong tax treatment. A hiring tool may summarize candidates quickly but introduce inconsistent criteria. A legal assistant may draft a useful clause while missing a jurisdiction-specific obligation.
Growing businesses need more than a generic “human in the loop.” They need to name the human, define the decision point, preserve the evidence reviewed, and specify what happens when confidence is low or source data conflicts.
ERP integration raises the stakes further. Connecting AI to accounting, inventory, payments, or customer records can eliminate duplicate entry and accelerate analysis. It can also spread a bad assumption across several processes. The ERP must remain the controlled system of record, with validation rules, role-based permissions, and reconciliations that do not disappear just because the interface became conversational.
A risk-based adoption ladder
- Start with reversible support. Use AI to summarize meetings, classify inquiries, draft internal notes, or prepare first versions of marketing content.
- Ground outputs in approved data. Connect only to defined folders, reports, or ERP views. Do not give an assistant unrestricted access to every document because search is convenient.
- Separate advice from execution. Let AI propose a journal classification, supplier response, or cash-flow explanation before it can post, send, or pay anything.
- Introduce structured review. Reviewers should see source records, confidence indicators, exceptions, and changes, not only a polished final answer.
- Measure errors by consequence. Track not just accuracy, but false approvals, missed exceptions, correction time, customer impact, and financial exposure.
- Use thresholds. Low-value, routine cases may proceed with light review. High-value, unusual, or regulated cases should require explicit approval.
- Reconcile independently. Bank, tax, payroll, and ledger controls must remain separate from the AI that proposes entries or explanations.
- Retain accountability. A named process owner must approve policy, monitor results, and suspend automation when assumptions or source systems change.
A 30-day pilot for finance-adjacent work
Choose one narrow task, such as preparing a weekly accounts-receivable summary. Define the required fields, approved sources, expected calculations, and escalation rules. Have the AI produce a draft using read-only ERP data, then compare it with the existing manual report for four weekly cycles.
Record missing invoices, incorrect customer groupings, unsupported explanations, and time spent reviewing. A successful pilot should reduce preparation effort without increasing reconciliation errors or hiding uncertainty. If it passes, the next step might be drafting collection messages for approval, not autonomously contacting every overdue customer. Ambition is useful; sequencing keeps it solvent.
How VDS Consultoria can help
VDS Consultoria helps businesses select AI use cases based on value, reversibility, data sensitivity, and control requirements. We can connect approved data sources, preserve ERP authority, design review queues, define approval thresholds, and build dashboards that show whether an automation is actually saving time.
The survey's message is straightforward: SMBs want AI, but they do not want blind trust. The winning approach is not to force sensitive processes into autonomy. It is to earn expansion through bounded pilots, visible evidence, and controls that remain understandable to the people accountable for the result.
Source
American Banker, “What small businesses want from their banks' AI,” 19 August 2026: https://americanbanker.com/payments/news/stax-research-shows-how-small-businesses-use-ai