AI agents, minus the hype

Every tool you already pay for is about to grow an "agent." Your CRM, your helpdesk, your accounting software, all of it. Analysts expect around 40% of enterprise applications to embed task-specific AI agents before long, and yet fewer than one in four organisations have actually got one running in production. We find that gap genuinely exciting, because it's where the real work lives, and it's the honest place to start any conversation about agents.

What an agent actually is

Underneath the marketing, an agent is simply a system that can take actions, not just produce text. A chatbot answers a question. An agent is handed a goal, works out the steps, and uses tools to get there: querying a database, calling an API, filling in a form, then checking whether it worked. The difference lives in the verbs. A chatbot says; an agent does.

That's also why we treat them with respect. The moment a system can act on your behalf, the cost of getting it wrong stops being an awkward sentence and becomes a deleted record, a mis-sent email, or a refund that should never have gone out.

Where they genuinely work today

The work agents do well right now has a clear shape: high-volume, rule-heavy and measurable. Triaging and routing inbound support tickets. Pulling figures from three systems to draft a weekly report a person signs off. Reconciling records between two databases and flagging the mismatches. Drafting a reply, a summary or a proposal that a human reviews before it goes anywhere.

The common thread is that a person stays in the loop on anything that matters, and you can measure whether the agent is genuinely saving time. If you can't tell whether it's working, it isn't ready to run on its own.

Where they don't, yet

Agents still struggle with long, unsupervised chains of steps, and the maths is unforgiving: a step that's 95% reliable sounds excellent until you string ten of them together, at which point a clean run is closer to a coin-toss at around 60%. They're a poor fit for high-stakes, irreversible actions without a human checkpoint, and for anything where being confidently wrong is expensive. The model doesn't know when it's guessing, so we design the system to know for it.

Why most stall before production

When we look at why agents don't make it past the demo, the answer is rarely the model. Roughly three-quarters of organisations point to data integration and quality as the biggest barrier: the agent simply can't reach or trust the systems it needs to act on. The other recurring trap is bolting an agent onto an existing process instead of rethinking the process around what the agent makes possible.

The teams that succeed don't ask "where can we add an agent?" They ask "which workflow is worth redesigning?"

How to start without betting the company

Our advice is the same one we follow ourselves. Pick a single workflow that's high-volume, measurable and forgiving, somewhere a mistake costs minutes rather than customers. Keep a person approving anything with real consequences. Instrument everything, so you can always see what the agent did and why. Prove the return on that one workflow before you let it anywhere near the next. Agents are real, and the good ones are genuinely useful. They reward the businesses that start narrow and honest far more than the ones that buy the whole roadmap on day one.