Cutting Through the Category Confusion
"AI agent" has become a label attached to everything from a scripted FAQ widget to a genuinely autonomous system operating across multiple business tools. The confusion is expensive, because it leads businesses to buy the wrong thing for the wrong problem.
A useful definition: an AI agent is a language model that has been given tools — defined operations it is permitted to perform — and the latitude to decide which tools to call, in what order, to satisfy a request. A chatbot talks. An agent talks and acts.
That distinction sounds academic until you see the failure modes. A chatbot that gets something wrong produces a bad answer. An agent that gets something wrong produces a bad answer and a wrong entry in your CRM. Which is precisely why deployment discipline matters more than model choice.
Agents Versus Rule-Based Automation
The instinct after using a capable model for a week is to route everything through it. This is a mistake, and an expensive one.
| Dimension | Rule-based automation | AI agent |
|---|---|---|
| Behaviour | Identical every run | Varies between runs |
| Cost per run | Effectively zero | Model tokens, meaningful at volume |
| Unstructured input | Handles poorly | Handles very well |
| Auditability | Complete by default | Requires deliberate logging |
| Best used for | Fixed rules, known inputs | Interpretation, classification, drafting |
The strongest production systems we build are approximately 80% deterministic pipeline with AI inserted at the two or three points where interpretation is genuinely required. The model reads the inbound email and extracts intent; deterministic logic does everything after that. This keeps the system cheap, debuggable, and predictable.
Four Deployments That Reliably Work
1. Support triage and first-line resolution
An agent connected to your documentation, order database, and ticketing system can resolve a substantial share of routine enquiries — order status, policy questions, account changes — and escalate the rest with a clean summary attached. The measurable win is usually not headcount but response time: instant, at 2am, in the customer's language.
The critical design decision is the escalation threshold. An agent that tries too hard to answer produces confident nonsense. An agent instructed to escalate the moment confidence drops produces trust.
2. Inbound lead qualification
Enquiries arrive as free text: an email describing a project, a form with a rambling requirements field, a WhatsApp message. An agent can read this, extract budget signals, timeline, and scope, score the lead against your criteria, and route it — while drafting a context-aware first reply for a human to approve and send.
3. Internal knowledge retrieval
Most organisations have institutional knowledge scattered across documents, wikis, and old email. A retrieval-augmented agent indexes that corpus and answers staff questions with citations pointing back to the source. Lower risk than customer-facing deployment, because the audience can immediately verify the answer, and often a strong first project for exactly that reason.
4. Content and document drafting
Producing first drafts from structured inputs — proposals from scoping notes, product descriptions from specification sheets, reports from data. The output is a starting point rather than a finished artefact, and the honest framing matters: this compresses the blank-page problem, not the review process.
Where Agents Fail
These are the failure modes we see most often, and each has a design answer.
- Confident fabrication. When the knowledge base lacks an answer, an under-constrained model will invent one. The fix is retrieval grounding plus an explicit, tested instruction to say "I don't know" and escalate.
- Silent irreversible actions. An agent with permission to issue refunds, delete records, or send external messages will eventually do so incorrectly. Any irreversible operation needs a confirmation gate — human approval or a hard rules-based constraint.
- Knowledge base decay. The agent was accurate at launch. Six months later, pricing changed and the documentation did not. The agent is now confidently wrong at scale. Knowledge base ownership must be assigned to a named person before launch, not after the first incident.
- Prompt injection through untrusted input. If your agent reads customer emails or web content, that content can contain instructions aimed at the model. Never let an agent treat data it retrieves as instructions, and never give it permissions broader than its narrowest legitimate task requires.
- Cost drift. Token costs look trivial in testing and become material at production volume, particularly with long context windows and multi-step reasoning. Instrument cost per interaction from day one.
A Deployment Sequence That De-Risks the Project
- Pick one narrow task. Not "handle support" — "answer order status questions for existing customers."
- Prepare the knowledge source. This is usually the majority of the total effort, and skipping it is the single best predictor of a failed deployment.
- Define tools and permissions explicitly. Read access is cheap to grant. Write access needs justification per operation.
- Run in shadow mode. The agent drafts; a human sends. Two to four weeks of this surfaces the edge cases no test plan would have found.
- Release with a visible escape hatch. Users must always be able to reach a human in one step.
- Review transcripts weekly. Not as a formality — as the mechanism by which the system actually improves.
The Honest Summary
AI agents are genuinely transformative for a specific class of problem: high-volume tasks involving unstructured input where a good-enough answer delivered instantly beats a perfect answer delivered tomorrow. They are a poor and expensive substitute for deterministic automation, and a worse substitute for fixing a broken process.
Our related piece on what to automate first is worth reading before scoping an agent project, and our view on the wider shift is set out in AI won't replace designers. If you want an assessment of whether an agent genuinely fits your workflow — including an honest answer if it does not — get in touch with our team.