AI automation is worth buying when it removes a defined operational bottleneck, has a named owner and can be checked against a clear standard. Start with one repetitive, high-volume workflow where errors are recoverable. Do not start with a company-wide chatbot or an instruction to “add AI”. A useful first project might classify incoming enquiries, draft replies for human approval, extract fields from routine documents or assemble internal reports.
The buying decision should combine commercial value, data protection, security and day-to-day usability. A convincing demonstration is not enough: ask how the system behaves with incomplete inputs, unusual cases and service outages.
Which process should you automate first?
Choose a process that staff can explain from beginning to end. Map the trigger, inputs, decisions, systems touched, output and exceptions. If nobody owns the current process, automation will usually make its ambiguity faster rather than make it better.
Score candidate workflows against five questions:
- How much repeated manual effort does the process create?
- Are inputs sufficiently consistent to test?
- Can a person verify the result before harm occurs?
- Is there a measurable baseline, such as handling time or rework?
- Can the workflow be paused and completed manually?
Avoid using a first pilot for decisions with serious legal, financial, employment or safety consequences. Automated decision-making and profiling can create additional UK GDPR obligations. The ICO guidance on AI and data protection is a sensible starting point when personal data is involved.
What should a good discovery phase produce?
A paid discovery should end with decisions, not a slide deck of possibilities. Expect a process map, prioritised use case, data inventory, integration map, risk register, success measures, prototype plan and delivery estimate. It should identify which actions remain human-approved and what evidence will be logged.
Define acceptance criteria in ordinary language. For example: the system extracts the required fields, cites the source document, flags uncertainty, never sends externally without approval and records who approved the action. Test criteria should include invalid files, conflicting instructions, missing data and duplicate submissions.
How should an SME handle data and privacy?
First establish what information enters the system, where it is sent, how long it is retained and whether a supplier uses it to improve shared models. Record the lawful basis for processing personal data and provide appropriate privacy information. Minimise the data supplied: an automation rarely needs every field available in a CRM.
A data protection impact assessment may be required where processing is likely to create high risk. The ICO provides guidance and a DPIA process. Seek specialist advice for your circumstances rather than treating a supplier’s standard terms as compliance.
Your contract and architecture should answer:
- Which organisation is controller, processor or sub-processor?
- In which countries is data processed and stored?
- What retention and deletion controls are available?
- Can authorised staff inspect and correct outputs?
- How are incidents reported and investigated?
- What happens to data and configurations when the contract ends?
What security questions should you ask?
AI does not replace ordinary cyber security. Require separate user accounts, least-privilege access, multi-factor authentication where supported, encrypted transport, protected secrets, audit logs, dependency updates and tested backups. Integrations should receive only the permissions needed for their task.
Ask the supplier to explain prompt-injection risks, malicious file handling and how untrusted content is separated from system instructions. The NCSC guidance on secure AI system development covers secure design, development, deployment and operation. The NCSC’s Cyber Essentials guidance also provides a practical baseline for common technical controls.
How much should AI automation cost?
Treat any range as a planning estimate that requires a scoped quote. A narrow prototype may be planned in the low thousands of pounds, while a production workflow with several integrations, security controls, a management interface and ongoing support can move into tens of thousands. Complex platforms can cost more.
The important variables are discovery effort, number and quality of integrations, data preparation, user interface needs, model usage, testing, assurance, hosting, monitoring and support. Compare total ownership cost, not only build cost. Include usage charges, maintenance, supplier subscriptions, staff review time and the cost of switching.
Request a staged proposal with assumptions and exit points. A fixed price is meaningful only when scope and acceptance tests are clear. Where uncertainty is material, fund discovery first and price delivery from its evidence.
How do you evaluate an AI supplier?
Ask for a live walkthrough of a comparable workflow and its failure handling, but do not expect access to another client’s confidential system. The supplier should be able to discuss trade-offs rather than promise perfect accuracy.
Supplier checklist
- Defines the business baseline and success measure before building
- Explains model, hosting and integration choices in plain English
- Documents data flows, sub-processors and retention
- Includes human review and a manual fallback
- Tests ordinary, edge and hostile inputs
- Provides logs, monitoring and incident ownership
- States what your business owns and can export
- Separates one-off fees from recurring costs
- Includes support response expectations and change control
What should happen before launch?
Run a controlled pilot with representative users and real-world examples that have been lawfully prepared. Compare results with the agreed baseline. Review false positives and false negatives separately because their business consequences differ.
Before release, assign a process owner, technical owner and incident contact. Set thresholds for escalation, usage limits and a kill switch. Train staff on what the system can and cannot do. Schedule an early review after launch, then regular checks for changed inputs, model behaviour, costs and user workarounds.
Frequently asked questions
Does an SME need its own AI model?
Usually not. Many useful automations combine a managed model with your rules, approved knowledge and existing systems. Dedicated hosting may be justified by security, performance or contractual requirements, but it adds operational work.
Can AI send customer messages automatically?
It can, but start with human approval. Move to limited automatic sending only after testing, monitoring and clear rules for sensitive or uncertain cases.
Who owns an AI automation?
Contract terms differ. Specify ownership and licences for source code, prompts, workflows, documentation, data and generated outputs. Also require practical export and transition arrangements.
How quickly can a pilot launch?
A contained pilot may take weeks, but integration access, data quality, approvals and testing often determine the schedule. A supplier should confirm timing only after scoping.
