- An AI agent pairs a model with instructions and permitted tools, and works best in one bounded role, such as turning an unstructured enquiry into a proposed brief.
- Adoption is rising quickly: ONS survey data for June 2026 puts AI use at around 35% of UK businesses with 10 or more employees, up from around 12% in late 2023.
- Many agent projects will not survive. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, blaming escalating costs, unclear business value or inadequate risk controls.
AI agents for business use a model, instructions and tools to carry out a defined task or sequence of tasks. They are useful when work involves interpreting information and choosing a next step, but need clear permissions, evaluation and escalation when the answer is uncertain.
Distinguish an agent from a chatbot and a fixed workflow
A chatbot provides a conversational interface. A fixed workflow follows a predetermined sequence. An agent can use a model to decide which permitted tool or step to use next. The same business system may combine all three.
That distinction matters when choosing a solution. Sending a reminder after a known deadline usually needs a rule. Reading an unstructured enquiry and preparing a proposed brief may benefit from AI. Anthropic’s engineering guidance similarly distinguishes predefined workflows from systems where models direct more of the process. More autonomy adds complexity, so give it a specific purpose.
Choose a bounded role with visible output
| Role | Useful output | Review point |
|---|---|---|
| Research assistant | A brief with source links and open questions. | Check relevance and evidence before using the conclusions. |
| Enquiry assistant | Proposed requirements and a draft response. | Confirm the customer record and approve the reply. |
| Document assistant | Extracted fields with source references. | Review missing or contradictory values. |
| Content assistant | A draft based on an agreed brief and brand context. | Editorial and factual checks before publication. |
A role should have defined inputs, permitted actions, completion criteria and an owner. “Help with everything” makes both testing and accountability difficult.
Give the agent the minimum access it needs
A worker preparing a proposal does not automatically need permission to send it. Separate read, draft, approve and execute capabilities. Apply the relevant user and customer permissions when retrieving records, rather than relying on the model to remember confidentiality instructions.
Set limits for time, spending and repeated tool calls. Treat external documents and messages as data that could contain misleading instructions. If the system cannot determine whether an action is allowed, it should escalate or stop, with the context needed for a person to decide.
Evaluate the work, including the awkward examples
Create a test set from representative tasks, with appropriate permission to use the data. Include missing details, conflicting messages, outdated documents and requests outside the role. Define what a correct result looks like and which errors would be unacceptable.
Review accuracy, completeness, time per task, cost per accepted output and the amount of human correction. A fluent answer can still be wrong. Re-run relevant tests when prompts, models, tools or source data change, and keep records of the configuration used.
Introduce autonomy gradually
Begin with recommendations or drafts. Run alongside the existing team before enabling external actions. Expand only when observed performance supports the change and the business has agreed the approval rules. A rollback or disable switch needs to work in practice.
Socialroom combines company context, AI writing and approval workflows. Klientflo applies AI assistance to communication and client requirements. These show specific product roles rather than a claim that AI can independently run every part of a business.
Continue with the AI readiness checklist, our platform approach or the AI agent FAQs.
Sources
Every figure in this article links back to the source below it was checked against.
Written for Orwell Lab’s practical guide collection. Statistics sourced and checked on 30 September 2026.
Discuss your own requirements ↗