Introduction
Small businesses are no longer only subscribing to AI tools. Many are assembling workflows, agents and API integrations of their own. A new Pax8 survey says 67% of the AI-engaged businesses in its sample had built at least one AI solution in regular use, while 35% had found unapproved AI tools, workflows or agents operating during the previous six months.[3]
That combination is the real story. Building has become accessible. Operating safely, predictably and economically has not become automatic.
What happened
Pax8 released its Q3 2026 SMB AI Pulse on 12 October. The online survey covered 1,150 owners, executives, functional leaders and managers involved in technology decisions at companies with 5 to 499 employees.[3] Respondents came from the United States, United Kingdom, Germany and Sweden, with country samples of 400, 350, 200 and 200 respectively.[3]
Among AI-engaged businesses, Pax8 reports that 52% were developing automated workflows connecting AI to other business systems, 46% were building or testing agents for multi-step tasks and 39% had built custom API integrations.[3] Among businesses that paid for AI in the previous year, 49% had received at least one bill or usage charge higher than expected.[3]
The survey was released by Pax8, a marketplace that sells cloud and AI solutions to SMBs. Its framing supports the commercial role of technology partners. The sample is also limited to four countries and technology decision-makers, while several questions apply only to businesses already using, building or paying for AI. The findings should not be generalised to all European SMEs or treated as Belgian market statistics. Pax8 states a nominal margin of error of about 2.9 percentage points for the combined sample, with larger margins for each country.[3]
Why it matters to business
The barrier to building an AI workflow can now be lower than the barrier to buying traditional enterprise software. A team can connect a model to email, CRM, documents or accounting data in days. That speed is useful, but it bypasses the routines that normally accompany production systems: requirements, access reviews, testing, budget ownership, support and change control.
Shadow AI is therefore not only an employee using a public chatbot. It can be a departmental workflow that sends customer data to an external model, an agent using a shared administrator account or an integration generating variable usage charges nobody monitors.
The hidden cost is not limited to tokens. Decision-makers in the survey reported spending time reviewing and correcting AI output. Pax8 says 55% spent at least three hours per week on review, editing or fact-checking, while 7% spent ten hours or more.[3] If that work is invisible, an automation can appear efficient while merely moving labour into quality control.
Practical actions for the next 30 days
- Run a lightweight discovery. Ask teams which AI tools, agents, browser extensions and automations they use. Compare answers with identity, browser and expense records.
- Create one register. Record owner, purpose, data, integrations, permissions, model, expected cost and current status for every use case.
- Set three autonomy levels. Read or draft, reversible action, and sensitive action. Attach approval requirements to the level.
- Make review explicit. Define which outputs require checking, who checks them and how much review time is acceptable.
- Set cost controls. Add monthly budgets, usage alerts, rate limits and an owner for investigating anomalies.
- Test with a fixed evaluation set. Keep representative examples and known edge cases. Re-run them when prompts, models or data sources change.
- Provide an approved path. Give teams a fast way to request a new tool or workflow. Governance that only says no will be routed around.
A practical measurement framework
For each automation, track four groups of measures. Outcome: cycle time, conversion, error reduction or cash collected. Quality: acceptance rate, correction rate and serious failures. Control: approval frequency, denied actions and unresolved alerts. Economics: model cost, integration cost and human review time.
Review the measures after two and six weeks. A pilot should expand only when the business outcome improves without unacceptable quality, control or cost signals. If review effort consumes the time supposedly saved, redesign the workflow rather than polishing the presentation.
How VDS can help
VDS Consultoria helps SMBs move from scattered AI experiments to controlled operational systems. We can inventory existing automations, identify data and permission risks, consolidate integrations and build a reusable control layer for approvals, logging, evaluation and cost monitoring.
The objective is not to slow teams down. It is to make successful experiments deployable and weak experiments easy to stop. We build AI systems that run your operations, with ownership and evidence attached.
Sources
[3] https://pax8.com/en-us/news-post/pax8-research-finds-small-businesses-are-building-ai-faster-than-they-can-govern-it — Pax8 Research Finds Small Businesses Are Building AI Faster Than They Can Govern It