Pick the dullest process you can find. High volume, well documented, clear right answers, mistakes that are cheap to fix. That kind of work produces evidence quickly, exposes few surprises and teaches the organisation how to run an agent before anything important depends on it. The exciting use case can come second, on foundations the boring one built.
Why do organisations start with the exciting use case?
Because it wins the budget. A proposal to automate credit decisions, complex claims or investment advice gets attention from the board. A proposal to classify incoming post does not. Sponsors want a first agent that shows what the technology can do, and the most visible processes are usually the most complex.
They are also, almost always, the hardest place to start. Complex decisions rely on judgement that is poorly documented, involve rare cases that matter a great deal, carry real harm when wrong and often fall under the strictest rules. Creditworthiness assessment of individuals, for example, is listed as high-risk under the EU AI Act. A first agent in that territory has to clear the highest bar with the least organisational experience.
What makes work good for a first agent?
Five properties, and the more of them a process has, the better.
| Property | Why it helps |
|---|---|
| High volume | Evidence builds in weeks, not months; rare cases appear sooner |
| Well documented | Procedures already describe the work, so the agent has clear instructions |
| A clear right answer | Each outcome can be scored, which makes certification and edit tracking meaningful |
| Low harm when wrong | Mistakes are cheap to find and fix; nobody is seriously affected |
| Stable inputs | Documents and requests arrive in predictable forms |
Typical candidates include classifying and routing inbound documents, drafting replies to routine enquiries, matching invoices to purchase orders, preparing standard letters for a person to send, and updating simple records such as contact preferences. Our guide on which process to give an AI agent first offers a scoring method.
Why does boring work produce better evidence?
Because evidence is a function of volume and clarity. An agent handling thousands of routine enquiries a month generates enough scored cases to judge each capability properly within a single quarter. Every case has a right answer that a reviewer can confirm. Edit rates mean something because the standard is clear. Rare cases still appear, because volume brings them out, and when they do they are easy to recognise.
A complex process with low volume might take a year to generate the same amount of evidence, and many cases will be contested: two experienced people might reasonably decide them differently. Certification on contested cases is slow and hard to defend. Our guide on measuring an AI agent's work in its first 90 days shows what a quarter of evidence looks like when the work suits it.
What does the organisation learn from a boring first agent?
Almost everything it will need for the exciting one. Running any agent in production means naming approvers, sizing the held queue, reading refusals, keeping sources current, tracking certificates, deciding when to widen and handling incidents. None of that depends on how complex the work is. All of it is easier to learn when mistakes are cheap.
A first agent on boring work lets the team make those mistakes safely. The first held queue that backs up, the first stale source, the first argument about whether to widen, the first incident: each is a lesson, and each is far less costly on routine enquiries than on credit decisions. By the time the second agent arrives, the operating model has been tested.
The first ceiling should be low, even for boring work. Start every capability at read or draft, so a person sends everything, and widen only as evidence arrives. Our guide on setting the first ceiling for a new agent explains how to choose it. Boring work tends to earn depth quickly, which is part of the point: the team experiences the full cycle of building evidence and widening within months, not years.
What if the boring process has no modern interface?
Often it does not. Routine back-office work tends to live in older systems, spreadsheets and shared mailboxes. That is not necessarily a barrier. Agents can work through files, mailboxes and document stores, and for some systems through the same screens people use, though each approach has its own controls to consider. Our piece on whether an AI agent can work without an API looks at the options. Starting there also forces the identity and access questions to be answered early.
How do you measure success for a boring agent?
By what matters for that work, not by headline numbers. For routine enquiries, it might be the share resolved correctly on first contact; our explainer on first-contact resolution covers that measure. For document handling, it might be correct classification and routing time. Add the operational measures that tell you the agent is governed well: reviewer edit rate, held queue age, refusal patterns, certification on time. A boring agent that is well run is a better result than an exciting one that is not.
What is the trade-off?
A boring first agent is a harder sell. It delivers modest, real savings rather than a dramatic story, and some sponsors will feel the organisation is aiming low. There is a risk that early success on simple work creates overconfidence about harder work.
The opposite risk is larger. A first agent on complex work that stalls, misfires or cannot be certified can set the whole programme back by a year and harden scepticism among staff and risk teams. Look at your shortlist for the first agent and rank it by volume, documentation, clarity of answers, harm and stability of inputs, not by visibility. Start with the top of that list. The deployment sequence it will follow is set out in our guide to deploying an AI agent: train, shadow, certify, widen.
Where Pryme fits
Pryme's Pre-Trained Agents are built for common jobs of exactly this kind, such as document intake and routine enquiries. They still start low for each customer: each agent is certified to a depth by Attest evidence on the customer's own cases, up to its role's ceiling, and certificates lapse after 90 days unless renewed.
Questions people ask
Is boring work worth automating if the savings are small?
Often, yes, because the main value of a first agent is what the organisation learns. Savings on high-volume routine work also add up, and they come with less risk than savings on complex work.
How long should the first agent run before starting a second?
Long enough to go through at least one full cycle: build, shadow, certify, widen a capability, and renew a certificate. That usually means a few months. Starting the second agent sooner risks repeating the first one's early mistakes.
What if there is no boring process available?
There almost always is one, though it may be a part of a larger process rather than a whole one. Split the complex process and give the agent its most routine step first, such as gathering documents or drafting a summary for a person.