AI Agents for Business Operations

Overview

Choose one process for an AI agent, decide what it may do and who checks it, then build, test and run it safely.

Outcomes

What you will be able to do.

• Distinguish rule-based automation, an LLM workflow and an agent, and justify the lowest-autonomy design for a given process.
• Score candidate processes on volume, variability, reversibility, cost of error, data readiness and observability, and exclude those that should not be delegated.
• Produce a process map, an exceptions list, a written definition of done and a test set of at least 30 real cases.
• Write a tool inventory and access plan that applies least privilege, with a single source of truth for each data item.
• Design an approval matrix, a review screen and a review rota that meets the meaningful human involvement test.
• Compare the main agent platforms on capability and usage pricing, and estimate monthly running cost, including review time.
• Evaluate an agent against its test set, run shadow mode, and red-team it for prompt injection and misuse.
• Complete a pre-build gate covering data protection (including a DPIA and processor terms), security, a failure catalogue, a kill switch and an incident plan.
• Define measures against a baseline, report performance honestly, and write a stop rule with a review date.
• (Implementer track) Build a working triage agent in a sandbox, with structured output, scoped tools, an approval step and run logs.

Curriculum

10 modules, 38 lessons.

Every module ends with a practical output on your own business. Open a module to see its lessons.

1. 01 What an agent is, and what it is not 3 lessons

   Much of what is sold as an agent is automation with a language model in one step, and that is often the better design. This module gives the vocabulary to tell the difference and choose the least autonomous design that does the job.

   1. 1.1 Rules, workflows and agents
   2. 1.2 The moving parts
   3. 1.3 What the evidence says about agents in business

   Practical output A shared vocabulary and a first list of candidate processes from your own business.

2. 02 Choosing the right process 4 lessons

   Agent projects fail most often at selection. This module gives a scoring method and the discipline to exclude processes that should not be delegated.

   1. 2.1 Where the time actually goes
   2. 2.2 The suitability test
   3. 2.3 The ones to leave alone
   4. 2.4 Agent, workflow or better process?

   Practical output A scored shortlist, a chosen process and a written design choice.

3. 03 Mapping the process and defining done 4 lessons

   An agent cannot follow a process that exists only in someone's head. This module writes the process down, lists the exceptions and defines what a correct result looks like, which becomes the test set.

   1. 3.1 Map before you build
   2. 3.2 The exceptions list
   3. 3.3 Define done and build the test set
   4. 3.4 The baseline

   Practical output A process map, an exceptions list, a written definition of done, a test set and a baseline.

4. 04 Tools, permissions and data 4 lessons

   An agent's risk is set by what it can touch. This module designs the tool inventory and permissions on the principle of least privilege, and decides where the truth lives.

   1. 4.1 The tool inventory
   2. 4.2 Accounts and credentials
   3. 4.3 One source of truth
   4. 4.4 Connecting systems: APIs, webhooks and MCP

   Practical output A tool inventory with permissions, an account plan and a system map with a source of truth for each data item.

5. 05 Human approval and escalation 4 lessons

   Human review is where agent projects succeed or quietly fail. This module designs the approval matrix, what reviewers see, who covers the queue and how it is recorded.

   1. 5.1 The approval matrix
   2. 5.2 What the reviewer sees
   3. 5.3 Queues, cover and review fatigue
   4. 5.4 The audit trail

   Practical output An approval matrix, a review design, a rota with a sampling plan and a log specification.

6. 06 Choosing a platform and working out the cost 3 lessons

   The platform decision is made last, after the design. This module compares the main routes with current pricing and teaches a cost model that includes running costs and staff time.

   1. 6.1 Build, buy or assemble
   2. 6.2 The platforms in 2026
   3. 6.3 The running-cost model

   Practical output A chosen route and platform with a monthly cost estimate and a spending cap.

7. 07 Building the first version (implementer track) 5 lessons

   A guided build of an inbox triage agent, first on a no-code platform, then optionally the same design in code. The core track can skip it or watch to understand what they are commissioning.

   1. 7.1 Build setup and a sandbox
   2. 7.2 The triage workflow in n8n
   3. 7.3 Adding tools and the agent loop
   4. 7.4 The same design on Zapier, Make or Copilot Studio
   5. 7.5 The same design in code (optional)

   Practical output A working prototype in a sandbox with logs, approvals and limits.

8. 08 Testing before customers see it 3 lessons

   Text demos do not predict live performance. This module turns the test set into a repeatable evaluation and runs the agent in shadow mode before it acts.

   1. 8.1 Running the test set
   2. 8.2 Shadow mode
   3. 8.3 Red-teaming the agent

   Practical output A scored test run, a shadow-mode plan and a red-team log.

9. 09 Security, privacy and failure 4 lessons

   Agents handle ordinary customer content, which can carry hidden instructions and personal data. This module covers the security and data protection controls a small business can actually run, and the plan for when it goes wrong.

   1. 9.1 Prompt injection and excessive agency
   2. 9.2 Personal data and the paperwork
   3. 9.3 The failure catalogue and the kill switch
   4. 9.4 The incident plan

   Practical output A completed Before You Build an AI Agent pre-build gate, including a failure catalogue, kill switch and incident plan.

10. 10 Running it, measuring it and deciding 4 lessons

    An agent is a process change, not a software install. This module covers measurement against the baseline, staff and roles, cost monitoring and the written stop rule.

    1. 10.1 What to measure
    2. 10.2 Reporting honestly
    3. 10.3 Staff and roles
    4. 10.4 Stop rules and the review date

    Practical output A measurement plan, a monthly report template, a staff plan and a written stop rule.

Fit

Who it is for.

Owners and operations managers of UK businesses of roughly 5 to 100 people who must decide what an agent may do, and the technically capable person who will build it.

Questions

Before you ask.