THE SHORT ANSWER
AI can interpret documents, classify requests, extract information, summarize, draft, reason over context and select tools. This extends automation beyond fixed fields and rules, but outputs remain probabilistic and can introduce hallucination, cost, latency, privacy and monitoring risks.
From fixed steps to bounded knowledge work
| Capability | Possible use |
|---|---|
| Natural language | Let people request work in ordinary language |
| Document understanding | Classify and extract from varied documents |
| Summarization | Prepare cases, meetings or reports |
| Generation | Draft messages and content |
| Reasoning/tool use | Choose bounded steps from context |
| Orchestration | Coordinate models, rules, tools and approvals |
Probabilistic work needs evaluation
A fluent result may contain invented facts, miss context or vary after a prompt or model change. Measure task-level accuracy, acceptance, failure severity, latency and cost with representative cases.
- Minimize sensitive data.
- Ground outputs in appropriate sources.
- Require review for consequential actions.
- Keep deterministic checks around AI steps.
- Monitor drift, cost and rejected outputs.
- Provide a fallback when the model is unavailable.
Evidence & context: NIST · UK Information Commissioner's Office
Add AI only where variation creates real friction
Use rules for stable logic and AI for a bounded variable task that has evaluation criteria. Do not turn an entire workflow into an agent because one step needs classification.
Continue with AI Agents & Automation or AI cost and performance.
Apply the idea to one real workflow
Choose one current workflow. Record the present outcome, the proposed change, the accountable owner, the most important exception or failure, and one before-and-after measure. Test the smallest safe version before expanding it.
Sources & further reading
- Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.
- Demystifying evals for AI agents
Anthropic. A provider's engineering guidance on multi-turn agent evaluation, checked 13 September 2026. Examples inform evaluation design but do not establish universal pass thresholds.
- Using tools
OpenAI Developers. Official documentation showing how models can be given built-in, function and remote tools. Checked 13 September 2026; product-specific tool names and availability can change.
- How should we assess security and data minimisation in AI?
UK Information Commissioner's Office. UK regulatory guidance, checked 11 September 2026. Jurisdiction-specific context, not individual legal advice or permission for a particular use.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗