Custom AI agents: when a template falls short

Custom AI agents are built around a particular business's tasks, tools and working rules. They make sense when a ready-made agent gets most of the job right but repeatedly misses a step that matters. That might be a required check before a report goes out, access to an internal tool the template cannot reach, or a company rule that does not apply anywhere else. The sensible starting point is usually a template. It gives you a working role to test before you spend time defining your own.

The decision to build a custom AI agent should come from a specific problem, not a general desire for something more sophisticated. If an existing research analyst can gather sources and prepare a useful report, there may be little reason to replace it. But if your research process requires checking every figure against an internal pricing sheet before the report reaches a manager, the standard version might need a change. The question is whether the difference is important enough to justify a dedicated agent.

Start with a template and observe the gaps

A template gives you a reference point. It has an established job, a starting brief and a defined purpose. You can give it a real piece of work, inspect the deliverable and see where its approach differs from yours.

Take a support triager. Its job might be to read incoming tickets, identify the subject, assess urgency and prepare a suggested response. That is a useful starting point for a team handling customer requests. After reviewing several runs, however, you might find that your company has a particular way of classifying requests from existing customers. Perhaps a billing question needs to be treated differently when an account has an unresolved dispute. A general support classification may not capture that distinction.

Write down the problem each time it occurs. Be precise. 'The agent needs to understand our business better' is difficult to act on. 'The agent classifies an invoice dispute as an ordinary billing question when the account already has an open complaint' describes something you can test.

Keep the examples. A single mistake might be an unclear instruction or an unusual input. The same mistake across different pieces of work suggests that the agent's standard approach does not fit your process. That is a much stronger reason to consider a custom version.

Three signs that you need a custom agent

Most reasons to customise an agent fall into three categories. Each points to a different kind of change, and each should be described before anyone starts building.

• A recurring step keeps going wrong. The template completes the main task but misses a condition that your team expects it to check. For example, an outreach writer prepares a suitable introduction but does not distinguish between a new prospect and someone who has already received a proposal. That distinction changes what the next message should say.
• The agent cannot reach a required tool. Your process depends on information held in a system that the existing agent cannot access. A research task might require checking an internal knowledge base alongside external material. If the necessary integration is unavailable or its access is too limited, the agent cannot complete the intended job as described.
• Your business has a rule that generic instructions cannot reliably express. Perhaps a recruiting team uses a specific screening rubric, or an agency requires a particular review before any client-facing material is prepared for publication. The rule belongs to your organisation, so the agent needs to work with that rule rather than assume a general industry practice.

These problems should not be treated as interchangeable. A missing step calls for a clearer process instruction. A missing tool calls for an integration and an appropriate access scope. A business-specific rule calls for an explicit policy that the agent can follow and a person can review. Giving an agent a longer brief will not solve a missing integration.

Define the difference before building

Before creating a custom AI agent, write down what it should do differently from the template. This sounds simple, but it prevents a common mistake: trying to customise an agent around an abstract description of the business rather than a particular job.

A useful specification covers the work itself. What starts the task? Which information should the agent read? What tools does it need? What conditions change the way it handles a case? What should it produce when the work is complete? Which actions must wait for a person?

Consider a bookkeeper working with an organisation that has its own expense categories. The standard bookkeeping role might prepare transaction matches and flag missing information. A custom version might need to apply the organisation's expense classification rules before preparing the results. The specification should identify those rules, explain how exceptions are handled and say what evidence must accompany an uncertain match.

Do not add every preference the team has ever expressed. Separate rules that affect the result from stylistic preferences that can remain in the brief. If a rule changes how a transaction is classified, it belongs in the agent's instructions. If someone simply prefers a shorter explanation in the final report, that may not justify a separate custom agent.

It also helps to write down what the agent must not do. An agent that prepares a financial reconciliation should not decide that an uncertain discrepancy can be ignored. One that drafts an external message should not send it just because the wording looks finished. The boundary needs to be part of the design, not an assumption made by the person assigning the work.

Custom does not mean unrestricted

A custom agent may have a narrower job than its template counterpart. That is often a better design. Giving an agent access to every system and permission available to the team creates unnecessary room for mistakes. Its tools and access should match the task it has been assigned.

The same applies to decisions. An agent can prepare a recommendation, explain its reasoning and assemble the information needed for a person to decide. That does not mean it should carry out every action associated with the recommendation.

At Delagents, we built the workspace around this distinction. A person writes a brief, an agent carries out the assigned work and checkpoints pause the delegation for sign-off. The approvals inbox holds actions that affect something outside the workspace. Each approval card shows the exact proposed action, the cost so far and a deadline. A person can approve it, change it or reject it.

Some actions always require a person, regardless of how an agent has been customised. These include first contact with someone outside the workspace, moving money, issuing refunds, deleting or overwriting records the agent did not create, changing roles or its own policy, connecting integrations or widening access scopes, accepting terms or signing, publishing publicly and acting in another workspace.

Every delegation also has a hard budget cap. An append-only activity log records tool calls, approvals and policy changes. These controls matter just as much for a custom agent as they do for a standard one. A business-specific instruction should never become permission for the agent to bypass the approval process.

Test the custom version against real exceptions

Once the custom requirements are written down, test them against actual examples of the work. Include an ordinary case, an incomplete case and one that triggers the special business rule. If the agent is supposed to recognise an existing customer with an unresolved complaint, give it an example that contains both details. Check whether it applies the correct classification and explains the reason.

Look at the deliverable, not just the agent's explanation of what it intended to do. Did it use the right source? Did it follow the required classification? Did it identify missing information instead of filling the gap with an assumption? Did it stop at the correct checkpoint?

Pay particular attention to cases where the instructions conflict. A company rule might require a second review, while the incoming request asks for an immediate response. The agent should recognise the conflict and stop at the appropriate point. It should not quietly choose whichever instruction is easiest to follow.

Keep a record of the remaining failures. If the custom version still misses the same condition, revise the instruction or reconsider the process. If the problem comes from unavailable information, investigate access rather than adding more wording. Customisation is a way to address a defined gap, not a guarantee that every unusual case will be handled correctly.

Know when to keep the template

There is a point at which customisation creates more work than it removes. If the template handles the important parts of a task and the remaining differences can be covered in an individual brief, keep using it. A new agent needs its own instructions, access decisions, tests and ongoing review. That effort should have a clear purpose.

It is also worth checking whether the supposed exception is actually a process problem. If two people give different instructions for the same situation, building an agent around one person's interpretation will not settle the disagreement. Agree on the business rule first. Then put it into the agent's instructions.

A useful threshold is this: customise when the same important requirement keeps appearing, when the existing agent lacks a necessary tool, or when the work depends on a rule that your business must apply consistently. Otherwise, the template is probably doing its job.

The Delagents agent roster and builder gives teams a place to start with roles such as research analyst, support triager, bookkeeper, outreach writer and meeting-prep agent, alongside inbox triage, content repurposing and recruiting screening. Start by choosing the closest role, then compare its intended job with one real task from your own operation. Write down the first requirement it cannot meet, decide whether that is an instruction, tool or policy gap, and use that specific example as the basis for your custom agent.