AI is sometimes presented as the answer to every inefficient process. That framing overlooks the technology SMEs already use successfully: forms, databases, rules, scheduled jobs, APIs and workflow systems. It also overlooks the decisions that should remain with people.
The most reliable solutions combine three categories deliberately. Deterministic automation handles predictable actions. AI assists with unstructured information. People make accountable decisions and handle exceptions. The design challenge is to put each step in the right category.
1. Deterministic automation: if X happens, do Y
Deterministic automation follows rules that can be stated and tested. If a valid enquiry is submitted, create a record. If an approval is overdue, notify the owner. At midnight, run an agreed reconciliation. When a status changes to complete, make approved information available to the next system.
This approach is well suited to:
- notifications and reminders;
- routing based on confirmed fields;
- scheduled tasks;
- movement and synchronisation of structured data;
- validation and calculation;
- approval sequences; and
- record creation through APIs.
The same input should produce the same result. Rules are inspectable, which makes them easier to test and explain. They still need exception handling: an unavailable API, duplicate event or missing value must lead to a visible and recoverable state.
Do not use a language model to decide something that a clear business rule can determine. AI introduces variability and cost without adding value to a fixed calculation.
2. AI-assisted tasks: interpreting unstructured material
AI becomes useful when input is less orderly. Emails, notes, documents, images and natural-language requests contain meaning that is difficult to process with fixed rules alone.
Appropriate assisted tasks can include:
- summarising a long enquiry or case history;
- classifying a message into likely themes;
- extracting proposed names, dates or requirements from a document;
- preparing a first draft;
- interpreting a natural-language request into a structured query; and
- highlighting likely anomalies for investigation.
These outputs are probabilistic. They can be incomplete or confidently wrong. The workflow should preserve source material, show the generated result as assistance, define when verification is required and provide a route when the model cannot handle the input reliably.
AI models should receive only the information required for the task through an approved service and permission model. Prompts, evaluation examples and model versions become operational configuration that needs controlled change.
3. Human decisions: judgement and accountability
People should remain central where the work involves commercial judgement, sensitive communications, safety, employment, legal rights, material approvals or unusual exceptions. They understand relationships and consequences beyond the pattern in the input.
Human involvement should not mean mechanically approving whatever the system suggests. A reviewer needs the source, relevant context, a clear decision and the ability to correct or reject the recommendation. Workload must allow genuine attention.
Accountability cannot be delegated to an AI model. Even when an assistant behaves conversationally or an agent takes several actions, the organisation and its people remain responsible for how it is used.
A combined customer-enquiry workflow
Consider this sequence:
Customer enquiry → structured capture → deterministic route → AI summary → workflow task → human review → response and action → recorded outcome → management information
A website form captures contact details, consent and service selection in explicit fields. Validation prevents obviously incomplete submissions. A rule creates a case and assigns an initial route. AI summarises the free-text problem and may suggest additional themes.
A person sees the original enquiry, structured fields and summary. They decide priority, suitability and response. An approved action updates the record, creates follow-up and feeds aggregate reporting.
The AI does not need to send the response or determine a commercial commitment. It reduces reading and preparation effort while the workflow preserves responsibility.
This architecture needs a database for the case, an API or integration for connected systems, a workflow engine for state and ownership, permissions for each role and an audit trail of material actions. A dashboard draws from the operational record rather than another manually maintained tracker.
A combined field-inspection workflow
The pattern also works in the field:
Inspection → structured evidence → validation → AI-assisted summary → exception → work item → human decision → closure evidence → reporting
The mobile application captures defined checks, photographs, location context and notes, including offline where required. Deterministic validation confirms required evidence. AI can prepare a concise summary or suggest an issue classification. A failed or uncertain result creates work for the appropriate reviewer.
The reviewer decides severity and corrective action. Closure requires defined evidence and approval. Dashboards show status from the same records.
Here, AI improves the handling of narrative and media context without replacing inspection requirements or accountable assurance. FieldProof provides a concrete solution framework for this type of joined-up workflow.
The architecture behind a dependable solution
APIs and integrations
APIs move approved information between websites, applications, CRM, finance, field and work-management systems. Define authentication, identifiers, validation, retries and reconciliation. Integration should respect the agreed system of record.
Databases
A structured database holds the state of work: input, ownership, status, decisions, evidence and outcome. It enables reliable search and reporting. AI output should be distinguishable from verified facts.
Workflow engines
The workflow represents stages, responsibilities, deadlines, approvals and exceptions. It should allow authorised correction and show stalled items. Avoid encoding needless bureaucracy simply because the software can.
AI models
Select a model and service according to the task, data, required quality, latency, cost and governance. Test against representative examples. Monitor actual failure modes rather than assuming a single benchmark describes the use case.
Dashboards
Management information should help people act: unowned work, ageing, exceptions, throughput and outcomes. Define metrics from business meaning, not merely the fields that are easiest to count.
Permissions and audit trails
Least-privilege access limits who can view, change or approve. Audit history records source, automated action, AI assistance, correction and human decision where appropriate. Logging should be useful without unnecessarily duplicating sensitive material.
Why these capabilities are now accessible to SMEs
Managed cloud databases, API-based AI models, configurable workflow services and modern web-development frameworks reduce the infrastructure required to build capable systems. An SME can combine services that once demanded a large enterprise programme.
Accessibility does not remove architectural responsibility. Someone still needs to understand the process, protect data, design permissions, test integrations and operate the result. A smaller organisation may gain speed by keeping the design focused and using managed services, but it should avoid creating a chain of poorly understood subscriptions.
The right level may be modest. A business can begin with an approved AI assistant and a repeatable review checklist. It can add workflow or integration only after the benefit and behaviour are understood.
When the answer is not AI
Sometimes the right intervention is a better form that collects complete information. Sometimes it is a database that replaces conflicting spreadsheets. Sometimes it is an API connection, a scheduled job, a clear business rule or a redesigned website journey.
Ask four questions before adding AI:
- Is the input genuinely unstructured or ambiguous?
- Would a fixed rule solve the task more reliably?
- Can the output be checked at a proportionate point?
- Does AI improve the complete outcome after review and exceptions are counted?
If the answer to the first question is no, conventional automation is likely to be the stronger foundation.
Use the simplest capable technology
A well-designed SME workflow is not judged by how much AI it contains. It is judged by whether work moves reliably, information is trustworthy, people can handle exceptions and decisions remain accountable.
Use deterministic automation where the rule is known. Use AI where interpretation of unstructured material creates genuine value. Keep people responsible for judgement and material actions. Connect the parts through secure data, APIs, permissions and observable workflows.
Above all, use the simplest technology capable of solving the problem reliably. That principle underpins Jay Malvern’s work across AI agents and intelligent automation, process automation and web-application development.