AI governance
Two-Thirds of Small Businesses Are Now Building Their Own AI. Only 23% of UK Firms Formally Review It.

The business has become the builder
A small business adding an AI tool to its work is no longer always just buying software. It may be connecting a model to a finance system, building an automated workflow or testing an agent that carries out several steps. That changes who is responsible for keeping the capability reliable.
The Q3 2026 Pax8 SMB AI Pulse Report, published on 12 October, surveyed 1,150 business leaders at firms with 5–499 employees in the United States, United Kingdom, Germany and Sweden. Among AI-engaged small and medium-sized businesses (SMBs), 67% had built one or more AI solutions already in regular use, and another 19% were prototyping. Only 13% relied exclusively on off-the-shelf tools.
These figures apply to AI-engaged businesses, not all SMBs. But they point to a material shift: organisations are building technology as well as buying it.
The new work is varied. Among builders, 52% were developing automated workflows connecting AI to business systems, 46% were building or testing agents that complete multi-step tasks, and 39% had built custom integrations using an AI provider's API.
A supplier may provide the model, platform or API. But when your organisation assembles these components into a working process, it also takes on responsibilities normally managed by a software vendor: deciding who can change it, maintaining it, securing it and paying for its use.
That does not mean every internal AI solution needs enterprise-scale controls. It does mean that "we built it quickly" is not a complete operating plan.
Why the UK review figure matters
Only 23% of UK AI-engaged businesses reported formal, documented review of internally built AI. The figure was 45% in Sweden, 36% in the United States and 32% in Germany.
A useful review does not have to be a large committee or a long approval queue. It does need to leave evidence of what was assessed, who accepted the risks, what conditions apply and who is responsible for the system once it is in use.
The report also found that 35% of AI-engaged businesses had discovered shadow AI (tools, workflows or agents operating without approval) during the previous six months. A policy explaining what employees may do with public AI tools is useful, but it does not, on its own, assess an internal workflow connected to customer, finance or operational systems.
This gap is manageable when responsibility is clear. It becomes harder when a process depends on an agent nobody has inventoried, a personal API key or a workflow that only its original creator understands.
What can go wrong when nobody owns the build
The risks are operational, not abstract.
- Data exposure: Security or privacy exposure was the most-cited risk of employee-built AI in every market, at 30% to 35%.
- Unplanned cost: Among businesses that paid for AI in the previous year, 49% received at least one bill or usage charge higher than expected. Usage-based costs were particularly difficult to forecast.
- Unclear accountability: If AI makes a mistake, leaders need to know which person owns the process, who reviews its output and how the work is corrected.
- Fragile operations: A workflow can stop working when its builder leaves, a supplier changes an API or a system's permissions are updated.
- Hidden review effort: Although 69% of AI-engaged businesses said AI had freed up at least some capacity, 55% of decision-makers personally spent three or more hours each week reviewing, editing or fact-checking AI-generated work. Seven per cent spent ten hours or more.
These findings do not mean organisations should stop building. They mean the expected benefit needs to include the effort to check and maintain the result, not just the time spent producing a first output.

A proportionate review process for internally built AI
A review process should match the system's impact. A tool that drafts internal meeting notes does not need the same controls as an agent that can update customer records or initiate payments. In both cases, however, the business should be able to answer basic questions about purpose, data, ownership and oversight.
1. Make an inventory before scaling
Record each internally built AI capability, including prototypes that access live data or systems. For each one, note:
- The business task it supports and the people who use it.
- The named business owner and technical contact.
- The AI provider, connected systems and data involved.
- Whether it can only recommend actions or can make changes itself.
- Its current status: prototype, approved pilot or regular use.
Start with a simple register if that is proportionate. The aim is visibility, not paperwork for its own sake.
2. Assess access and impact
Check what the system can see and what it can do. Confirm that permissions are limited to what the process needs, and that sensitive information is handled appropriately.
Then consider the consequences of an incorrect output or action. Could it affect a customer, employee, payment, safety decision or contractual commitment? Where consequences are material, require a human review before the action is taken. Define a route to stop or escalate the process if it behaves unexpectedly.
3. Test, approve and record changes
Before a system moves into regular use, test it against normal tasks, edge cases and foreseeable errors. Keep a record of who tested it, what was found and who approved the move into use.
Reassess it when the model, instructions, connected data, permissions or business process changes. A successful pilot is not permanent proof that a changed system remains suitable.
4. Assign maintenance and cost responsibility
Name a person responsible for routine checks, supplier changes, access reviews and incident escalation. Document enough for someone else to support the system if its original builder is unavailable.
Set a budget owner and monitoring thresholds for usage-based charges. Where suppliers support them, use spending alerts, usage limits or rate controls. Check costs against the work the system actually completes; a low API charge is not good value if staff spend hours correcting the output.
5. Measure benefit and review regularly
Baseline the task before deployment: time taken, volume, quality checks and common errors. After launch, compare like with like. Count the time saved after review effort, not before it. Track incidents, exceptions and cost alongside productivity.
Most organisations can begin with a lightweight review and revisit active systems at a sensible interval, such as quarterly or when a significant change occurs. Fresh Mango AI's AI Governance & Security Project and Quarterly AI Success Review are two ways to structure that work. Which approach is suitable depends on the systems in use and the risks they create.

Treat control as part of the business case
The report found that 70% of SMBs would pay at least 6% more for greater control over how their data is stored and used. That signals that data control is not simply an IT preference. It is part of the value organisations expect from AI.
Similarly, 66% of SMBs said their business would not remain competitive without AI in the next three years. That is a reason to assess AI carefully, not a reason to approve every build. The commercial question is whether a specific capability improves a real process enough to justify its operating cost, review effort and risk.
A useful starting point is a short register of what has been built, followed by a review of the systems that access sensitive data or take actions in business systems. Prioritise those first. Then decide whether to approve, fix, monitor or retire each one.
For a wider view of use cases, data readiness and priorities, consider an AI Productivity & Readiness Assessment. The purpose is to make a clearer decision about what to build next, and what should not proceed yet.

What happens next
Small businesses are becoming AI builders. The corresponding responsibility is to manage each build as an operational capability, not as a one-off experiment that can be forgotten once it works.
You do not need to review every idea with the same level of detail. You do need to know what is running, what it can access, who owns it, what it costs and how its performance is checked. That is the foundation for scaling useful AI without losing control of the work it supports.
A practical next step
If you want to discuss what your organisation has built and what needs reviewing first, a free 30-minute discovery session is a straightforward place to start.
There is no obligation, and if a different next step is more suitable, we will say so.
Common questions
Frequently asked questions
Does building AI in-house mean we need enterprise-scale controls?
No. Controls should match the system's impact. A tool that drafts internal meeting notes needs less oversight than an agent that can update customer records or initiate payments. In both cases the business should be able to answer basic questions about purpose, data, ownership and oversight.
What counts as a formal review of internally built AI?
A useful review does not require a large committee or a long approval queue. It does need to leave evidence of what was assessed, who accepted the risks, what conditions apply and who is responsible for the system once it is in use.
Is a policy on public AI tools enough to manage shadow AI?
A policy explaining what employees may do with public AI tools is useful, but on its own it does not assess an internal workflow connected to customer, finance or operational systems. The Pax8 research found 35% of AI-engaged businesses had discovered shadow AI in the previous six months.
How often should internally built AI be reviewed?
Most organisations can begin with a lightweight review and revisit active systems at a sensible interval, such as quarterly or whenever a significant change occurs to the model, instructions, connected data, permissions or business process.
Sources
- Pax8 Q3 2026 SMB AI Pulse Report (opens in a new tab)
Published 12 October 2026. Survey of 1,150 business owners, executives, functional leaders and managers involved in technology decisions at companies with 5–499 employees across the United States, United Kingdom, Germany and Sweden. Results for smaller subgroups within individual markets should be interpreted as directional.
- Pax8 report announcement and methodology (opens in a new tab)

