AI instructions are the text where you tell an agent how your organization works. The agent reads them before starting and uses them to decide what to look at, how to read what it finds and what to prioritize.
They turn your policy into practice. If your manual says a cash transaction above a certain amount needs supporting documentation, the instruction is where that becomes concrete agent behavior.
Where to set them
Agent instructions
In Settings > AI & Agents > Agents configuration, inside the agent's "Instructions" section.
They apply to that agent across everything it works on, whatever the template. This is where you set its general judgment: how it writes, how much detail it uses, when it leaves something pending review.
Template instructions
In Settings > Cases, inside each case template, and in Settings > Onboarding, inside each onboarding template. The field is called "AI Instructions" and sits at the bottom of the form.
They apply to every case or onboarding created from that template, whichever agent works on it. This is where you put what belongs to that case type: what a PEP due diligence reviews, what documentation a legal entity onboarding requires.
Each template has its own instructions, independent from the rest.
How they combine
When an agent works a case, it reads both and combines them: the agent's instruction brings the judgment, the template's brings the case context.
A practical rule for deciding where to write each thing: if it holds for everything the agent does, it goes on the agent. If it holds for this case type even when another agent or a person works it, it goes on the template.
Both fields accept up to 50,000 characters and support basic formatting such as bold, lists and tables.
Recommendations
These are the four shown on the configuration screen, and what they mean in practice.
Be specific about the behavior and tasks you expect from the agent
A vague instruction produces a vague report. Name the data point, the threshold and the action.
Weak: "Review the customer's documentation carefully."
Better: "Check that the address on the proof of address matches the declared one. If they differ, state it in the report and request a proof issued within the last 90 days."
The difference is that the second one says what to compare, against which criterion, and what to do with the result.
Set the tone and format of the responses
The agent writes reports your team reads, and in some cases a regulator too. Tell it how you want them.
"Write the report in the third person and in past tense. Always open with the conclusion, followed by the evidence supporting it."
"Cite the source of every data point in parentheses: the list and version date for a screening hit, the document number for a file."
"Do not use bullet points in the conclusion. The report is exported to PDF and long lists get cut off."
Include examples or edge cases when they help clarify
Ambiguous cases benefit the most from an instruction. One concrete example beats an abstract rule.
"A namesake with the same date of birth but a different ID number is not a confirmed match. Leave it pending review and explain which data points you compared."
"If the customer declares salary income and receives transfers from a company other than their employer, that is an inconsistency. Document it even when the amount is small."
"A gap of up to two days between the document date and the upload date is not a finding. Do not report it."
That last example shows something people often forget: it also pays to tell the agent what not to report. A report full of irrelevant findings costs as much to review as an empty one.
Keep them concise but comprehensive
The agent reads the whole instruction, so every line competes for its attention. There is no need to restate general compliance rules, it already knows them. Use the space for what is specific to your organization.
Before adding a paragraph, ask yourself whether it changes any decision the agent makes. If it does not, cut it.
Examples by scenario
Legal entity onboarding. "Pay special attention to the chain of ultimate beneficial owners. Flag any shareholder above 25% ownership. If the structure includes companies in another jurisdiction, state the country and check that the documentation is apostilled."
PEP cases. "This case involves a politically exposed person. Apply enhanced due diligence. Verify the linked companies and look for indirect ownership structures. Document every source you consult with its access date."
Cash transaction monitoring. "Compare the amount against the declared transactional profile. If it goes over, state by how much and over what period. Check for amounts split below the reporting threshold and describe the full sequence of transactions."
Fast-track onboarding. "This is a flow for pre-approved customers. Focus only on blockers: identity that cannot be verified, expired documentation or restricted jurisdiction. Minor formatting issues must not delay approval."
Limits worth knowing
Instructions do not grant permissions. The agent can only do what its roles and its tools allow. If an instruction asks for an action it has no access to, it does not run it. To widen what it can do, adjust its roles or its tools, not the text.
Results can vary. The models were tested with Complif's default instructions. When you write your own, review the first cases the agent works and adjust the text based on what you see.
Where to start
If this is your first setup, load an instruction from the library in the agent's "Instructions" section and use it as a base. They are texts already tested for that agent type, in your organization's language and country.
Then adjust what is specific to your operation, let the agent work a few real cases, and correct the text based on what you find in the reports.