“Agentic AI” is two things at once: a real change in what software can do for a small firm, and this year’s marketing sticker. The real part is the move from AI that answers to AI that works. This page is the map of the cluster, and the four guides beneath it go into the detail.
What an AI agent does, in one example
An agent is given a job and carries it out, where a chatbot only talks about it. Ask a chatbot “which of our invoices are overdue?” and, at best, you get advice about credit control. An agent connected to your systems checks the ledger, finds the overdue accounts, drafts each chasing email in your voice, queues them for a person to approve and reports what it did. The model is the same. What changes is that it plans steps, uses tools, and carries the task instead of the conversation.
That is the whole definition worth having: planning, tool use and multi-step work, under supervision. The cluster fills in the rest in this order:
- What is an AI agent? The plain-English anatomy and the vocabulary vendors will use at you.
- Agentic workflows in practice. The five patterns where agents pay their way in a service firm.
- Governance and risk. The six controls that keep an acting system defensible.
- Getting started. The first-agent playbook, from choosing the process to the month-end decision.
Which work in a service firm suits agents
Recurring, rule-shaped work that sits next to a judgement call suits agents best, and professional firms are full of it. Enquiry triage. Onboarding chases. Reconciliation. Report assembly. Watching the sources that matter and surfacing only what needs a person. The agentic workflows guide works through these patterns one by one, including what this site’s own operations automate every day. For the accountancy version of the question, see AI for accountancy firms.
Why agent risk is different from chatbot risk
A chatbot’s mistake is a wrong sentence; an agent’s mistake is a wrong action. That is why governance sits at the centre of this subject rather than in an appendix:
- Permissions bounded: an agent’s tool access sets how much damage it can do
- Actions logged: every step can be inspected after the fact
- Approval gates exactly where the consequences are
- Caps on spend and volume, because an agent can also fail by doing too much of the right thing
Gartner’s much-cited prediction is that over 40% of agentic AI projects will be cancelled by the end of 2027 for cost and unclear value. Most of that failure is avoidable, and the risk and governance guide covers how, written for firms that answer to regulators.
How to adopt agents without the risk
Start with an agent whose every output is a draft a person approves. It does the full job, nothing acts on its own, and a month of real traffic gives you the evidence with no risk from its actions.
Firms that skip the draft-only month are the ones that end up in the cancellation figures. It costs a few weeks and buys you a measured error rate on your own work.
The pattern that works is unheroic: pick one narrow process that hurts, give an agent the job with every output held for approval, keep a person on the gate, measure honestly, then widen. The getting started guide turns that into a step-by-step path, including the build-versus-buy question and what a realistic first month looks like.
Related reading elsewhere in the knowledge base: multi-agent routing for how larger agent systems are structured, and the internal AI assistant playbook for the adjacent build most firms want first. If you’d rather see agents doing real work before believing any of this, the Labs page shows what runs here.