The rest of this cluster covers what agents are, where they pay and how they stay safe. This page is the route map: the shortest defensible path from no agents to one that clearly earns its keep, written for a firm without a technology team.
Why the first agent matters
The first agent sets the pattern for every one after it. If it goes well, you end up with staff who know how to supervise an agent, a governance setup the next few agents can copy, and evidence that makes the second decision easy. Those are worth more than the hours it saves.
If it goes badly, it becomes the story people in the firm tell for two years about why this stuff doesn’t work. The five steps above are designed to avoid that.
Choosing the process
The right first process recurs often, has simple rules, and keeps its consequences inside the firm. Score each candidate on four questions:
- Does it recur at least weekly?
- Do its rules fit on a page?
- Do its consequences stay internal, or arrive as drafts a person approves?
- Does someone currently hate doing it?
Four yeses is your first agent. The usual first choices are enquiry triage, a chasing routine or a reconciliation, for reasons the workflows guide sets out. Don’t start with the most valuable process. Start with the best-shaped one. If several candidates tie, ranking them by effort against reward is what an Automation Audit is for.
Buy versus build
Buy when the process lives inside one system you already pay for. Build when it crosses systems or follows your firm’s own way of working.
| Buy (configure what you own) | Build (commission your shape) | |
|---|---|---|
| The process lives in | One system you already pay for | Several systems, or your firm’s own shape |
| Speed to start | Days | Weeks |
| Fit | The vendor’s idea of the process | Your actual process |
| Governance | Only as good as the product’s logs and gates | Specified: logs, gates and caps by design |
| Cost shape | Subscription | Fixed build + measured running costs |
More of the software you already pay for now ships agent features. If your chosen process sits entirely inside one of those systems, configuring what you own is the sensible first move, and the AI-washing guide will help you judge whether the feature is real.
Build when the process crosses systems, when it’s specific to your firm, or when the bought version fails the governance questions: can you inspect its logs, are its permissions bounded, and does a person approve the parts that matter? Whichever route you take, the six controls stay the same. A bought agent without logs saves money only until something goes wrong and nobody can say what it did.
The first month: drafts only
For the first month, the agent does its whole job but every output lands as a draft for a person to approve. Nothing goes out unchecked, so an error costs nothing beyond the correction. In return you get a month of evidence from real work, staff who are supervising rather than speculating, and edge cases showing up while they’re still free to fix.
At the end of the month the numbers decide: volume, error rate against reviewers, hours returned. Either the routine work earns some autonomy, sign-off stays where the consequences are, and the next process is queued, or you stop cheaply and write down why. Both are better than the pilot that never proves itself and never gets switched off.