


We either save you money or make you money



We either save you money or make you money

How to commission AI agents for marketing, decide which tasks to automate, retain human control and assess the finished work.
AI agents for marketing can be designed to investigate a defined issue, use connected information and prepare or carry out permitted actions. For a business commissioning that work, the useful question is specific: which task should the system complete, to what standard and with whose authority?
A campaign review, content research process or enquiry assessment can involve several steps. Each needs an agreed source of information, a useful output and a clear owner. The label attached to the software matters less than the reliability of the process your team receives.
WLW’s marketing engineering approach connects those processes with senior marketing judgement, customer understanding and commercial priorities.
A fixed workflow follows an established sequence. An agent can choose its next step or tool according to the information it receives. This distinction is set out in Anthropic’s explanation of workflows and agents. The same guidance recommends starting with the simplest approach that meets the requirement.
For example, notifying a named account manager when a form arrives may need an ordinary automation. Investigating why a campaign attracts unsuitable enquiries could require a more flexible sequence of evidence gathering and analysis.
Ask your supplier to explain that choice. Adding more independent decisions increases the number of behaviours you need to test, support and control. The specification should justify the complexity through a useful business outcome.
Define the expected output in operational language. “Review marketing performance” leaves too much open. “Prepare a weekly review of these campaigns, using these date ranges and the sales team’s accepted-lead definition” gives the work a boundary.
The brief should identify authorised accounts, relevant markets, reporting periods and the person who receives the output. It should explain what counts as sufficient evidence, what the system must leave unchanged and when it should ask for human review.
Define completion carefully. A report is complete when its evidence is current enough for the decision, its calculations are correct and its recommendations are usable. A successful connection to an account only proves that the connection worked.
Consider an illustrative campaign review for a specialist equipment supplier. The workflow could bring together search terms, landing-page performance and the sales team’s assessment of recent enquiries. It would identify a possible mismatch, show the supporting evidence and prepare a proposed change.
The campaign owner would review the recommendation against margins, stock, territory and sales capacity. Only an approved action would proceed within the agreed spending limits. The approval checkpoints are part of the proposed operating specification.
Separate read access from permission to edit. Specify which fields can change, the approval required and how the business can inspect the previous version. A recommendation process should not acquire publishing or spending authority simply because its account connection permits those actions.
A dependable test includes awkward cases. Supply an empty report, a stale dataset, a missing account and a duplicated enquiry. Check whether the system recognises each condition and gives the right explanation.
Missing evidence must remain missing evidence. A failed data request should not become a claim that traffic, enquiries or sales have fallen to zero. An incomplete period should not be compared with a complete one without qualification.
Test uncertain writes as well. When a system times out after an action, it should inspect the destination before trying again. Your acceptance criteria should cover duplicate prevention, escalation and the ability to stop the process without losing the work already completed.
Specify what the next person needs to make a decision. A campaign recommendation should identify the affected activity, supporting evidence, proposed action and any approval required. A research summary should retain the sources and dates behind its conclusions.
Build the handover into the workflow. Decide where the output is stored, who reviews it and how an exception reaches the right person. Include a clear response when the evidence is too weak to support a recommendation.
Test the process from the original request to the finished handover. Your team should be able to inspect the work, understand its limits and decide what happens next. These requirements belong in the acceptance criteria before the system is put into routine use.
Agree who updates instructions, checks changes in connected systems and investigates exceptions. Keep a record of approved versions and provide a route for your team to report a recurring mistake.
Corrections should improve a controlled process. A system that changes its own operating instructions needs stricter testing than one that prepares an amendment for review. Decide which approach is appropriate before it is deployed.
Your business should know which accounts it owns, which components depend on a supplier and what support remains necessary. Documentation should explain how to operate, pause and recover the workflow, as well as how it was initially built.
Choose a process whose output someone can assess confidently. Compare the completed work with the previous method, including review time, error correction and ongoing costs. Expand only when the evidence justifies a wider remit.
Get in touch with WLW about AI agents for marketing. Bring a task your team needs to complete more reliably. We will discuss the workflow, level of automation and controls it requires.
Featured photograph: Circuit board detail by quapan, Wikimedia Commons. CC BY 2.0. Original photograph reproduced for the website.
Phone:
Email:
Address: