Start with a bounded process
Choose one workflow with a clear trigger, input, output and owner. Avoid starting with “an autonomous agent that runs the company”. Narrow systems are easier to evaluate, secure and improve. A bounded process is the difference between a system you can prove and a project you can only hope works.
Fife accountancy practice: instead of building a company-wide AI assistant, the team scoped one bounded process — extracting invoice fields from supplier PDFs. Clear trigger (PDF arrives), clear input (invoice), clear output (structured fields for review), named owner (the partner). This is a system that can actually be built and measured.
One process, one outcome, one ownerDesign the system around the model, not the other way round
Authentication, permissions, business rules, retrieval, validation, queues, human approval, monitoring and rollback are often more important than the model itself. The model is one component inside a controlled process.
Define acceptance criteria before you build
Decide what a good output looks like, which errors are tolerable, which are unacceptable, what evidence must be shown and when the system should refuse or escalate. Write these down before anyone writes code — otherwise “it looks good to me” becomes the only standard.
For invoice extraction: 95% of fields extracted correctly on a fixed 40-invoice test set; 100% of total-value fields must match the line-item sum; anything uncertain is flagged for human review rather than guessed.
Pilot, observe, expand
Start read-only or draft-only. Monitor failures and user corrections. Add reversible actions only when the evidence supports more autonomy. Expansion should be gated by measured success, not enthusiasm.
A system that "looks fine" on a demo can fail quietly on the tenth real case. Keep a small test set, run the system against it, and measure pass/fail — not just whether the answer sounds confident.
Pilot first
- One bounded process with a baseline, as in the Fife invoice example above
- Measure before and after, then expand only on evidence
- Failures stay cheap to find and fix
The default for a first implementation.
Big-bang rollout
- Whole-company or whole-department launch in one go
- Baselines and acceptance criteria are harder to pin down
- One quiet failure can ripple across the business
Rarely justified before a pilot has produced evidence.
The difference is not ambition but measurement. A pilot is how you earn the right to scale: it produces the baseline, the failure data and the acceptance evidence a wider rollout needs.
A production readiness checklist
- Named owner accountable for the outcome.
- Authentication and least-privilege permissions.
- Data mapped, minimised and routed through an approved path.
- Acceptance criteria written and tested against a fixed set.
- Human approval for consequential actions.
- Monitoring, logging and a way to stop or roll back.
- Cost limits and an incident plan.
Common failure modes to design around
For Scottish SMEs the biggest implementation risk is rarely the model — it is skipping the boring engineering because the team is small and time is short. A bounded pilot with a baseline is more valuable than a grand rollout with no measurement.
Frequently asked questions
Do we need an AI expert to implement AI?
For a narrow, bounded pilot, a capable person who can configure integrations and a no-code or low-code tool may be enough. Larger or higher-risk systems warrant specialist help. Start small either way.
What is the hardest part of AI implementation?
Usually the surrounding system: data access, validation, human approval, monitoring and rollback. The model is the easy part; the engineering around it is where value and risk live.
How do I know it is ready for production?
When you can answer: what data, why AI, what can fail, who approves, how we detect failure, and how we recover. If you can't answer all six, it is not ready.
- ICO — AI and data protection guidance
- NCSC — Guidelines for secure AI system development
- Alan Turing Institute — AI ethics and society briefing papers
The tooling behind this guide is being built in public. It is open source on github.com/BryanFiFife, so you can inspect rather than take our word for it:
Open source does not mean certified. Read the code, run the tests, and treat it like any tool: check it against your own requirements before you depend on it.
A reliable implementation is not “an LLM with a prompt”. It is a controlled business system with a model only where a model adds value.