Two ways to automate
Rule-based automation is what most marketing teams already run. A marketing automation workflow waits for a trigger, such as a form fill, a purchase, a date or a score crossing a threshold, then follows branches and actions that a person designed: send this email, wait, check whether the contact clicked, update this field, notify sales. Lead ScoringAssigning points to leads for fit and behaviour to decide which are ready for sales. rules, Lead NurturingKeeping in useful contact with leads who are not yet ready to buy until they are. sequences, abandoned-basket emails and data hygiene jobs all work this way. Their defining property is that the behaviour is fully described in advance. Given the same inputs, they do the same thing every time.
An AI agent, as the AI for marketing automation guide describes it, is a model given a goal, a set of tools and permission to take several steps without a person in between: read the record, look something up, draft a reply, update a field, schedule a follow-up. The behaviour is not written down; it is decided at run time by the model, based on the goal, the instructions and what it finds. Two similar inputs can produce different sequences of steps. That is what lets an agent handle the long tail of tasks that never justified a workflow, and it is also what makes it harder to govern.
The choice between them is a choice about where judgement sits: with the person who designed the rules, ahead of time, or with the model, at the moment of acting.
Side by side
Rule-based automation vs AI agents
| Row label | Rule-based automation | AI agents |
|---|---|---|
| Predictability | The same input always produces the same steps and output | Steps and output can vary between runs on similar inputs |
| Auditability | The logic can be read before it runs; logs show which branch fired | Needs logs of what the agent saw, decided and did, and why, to reconstruct a run |
| Cost of an error | Errors are systematic and repeat until fixed, but are easy to trace | Errors are varied and harder to anticipate; one run can do something new |
| Handling unexpected inputs | Breaks or falls through to a default branch | Can adapt, sometimes well and sometimes wrongly |
| Setup | Every step and branch has to be designed and built | Goal, instructions, tools and permissions have to be defined; fewer explicit steps |
| Upkeep | Changes when the process, data or offer changes | Changes when the process changes and when the model or tool is updated |
| Testing | Test each branch with representative records | Test a varied set of inputs, several runs each, and judge output quality |
| Data access | Touches only the fields each step uses | Often needs wider read access to decide what to do; access should be limited per tool |
| Oversight | Monitoring of volumes, errors and outcomes | Sampling review of runs, approval gates for actions, an owner who can pause it |
| Running cost | Low per run once built | Higher per run; each step may involve a model call |
The rows group into two themes. Rules are strong wherever the business needs to know in advance what will happen: predictability, auditability, testing and data access. Agents are strong wherever the task varies too much to script: unexpected inputs and setup for messy, one-off work. Upkeep and oversight are heavier for agents, because their behaviour depends on a model the team does not control and can change when the vendor updates it.
Where rules win
Rules win whenever an action has consequences that are expensive to reverse or that must be explained afterwards.
- Sending to customers. Which email goes to whom and when should be decided by logic someone can read, with suppression lists, frequency caps and consent checks that cannot be reasoned around.
- Spending money. Discounts, refunds, budget changes and bids above set limits need fixed thresholds.
- Changing records. Lifecycle stages, lead status, opt-in fields and ownership drive reporting and other workflows. A wrong change spreads.
- Regulated or contractual steps. Consent handling, unsubscribe processing and required disclosures must behave identically every time.
- High-volume, uniform tasks. Where inputs are similar and the right action is known, a rule does it faster, cheaper and more consistently than a model.
Rules also win on accountability. When a workflow misfires, someone can read the logic, find the branch at fault and fix it, and the automation governance guide covers the naming, ownership and review that keep a workflow estate in order. Reconstructing why an agent took a particular path is a harder investigation.
Where agents earn their place
Agents are worth their extra oversight where the steps cannot be written down in advance and the result can be checked before it matters.
- Research and preparation. Gathering information about an account from several internal sources and drafting a briefing for a salesperson; preparing a summary of last month's campaign results with the queries attached.
- Varied, low-volume requests. Internal requests that differ every time, such as pulling a list, checking a setting, or drafting a brief from scattered notes.
- Drafting responses that need context. Reading a customer's history and a question, then drafting a reply for a person to approve.
- Triage of messy inputs. Reading inbound messages that do not fit categories cleanly and proposing what should happen next.
The common thread is that the agent prepares and proposes, and a person or a rule decides. Before an agent goes into use, the method in testing prompts and models applies: a test setA collection of real inputs for an AI task, each noting what a good output must contain, run to compare prompts or models and rerun when they change. of real inputs, several runs each, a rubric and a pass bar, rerun whenever the model or instructions change.
Mixing the two: agents inside guardrails
The most useful designs in marketing are rarely pure. A rule-based flow handles triggers, routing, permissions and sending; a model, sometimes acting as an agent, does one bounded job inside it; an approval step sits between the model's output and anything a customer sees.
A typical pattern: a workflow triggers when a high-value lead submits a demo request. A fixed rule checks consent and ownership. An agent reads the lead's form, website visits and company record, and drafts a personalised follow-up and a short briefing for the salesperson. The workflow then holds both for the salesperson to edit and approve, and only the approved email is sent, through the normal sending rules with their suppression and frequency checks. The agent contributes judgement where rules struggle; the rules keep control of everything consequential.
Guardrails for agents inside flows follow from the comparison table. Grant permissions per action rather than per agent, so it can draft but not send, tag but not delete. Log each run with the inputs it saw and the steps it took. Cap volumes and set limits that stop a run that behaves unusually. Name an owner who reviews a sample of runs on a schedule and can pause the agent. The NIST AI Risk Management Framework's functions of govern, map, measure and manage are a useful checklist here: who is accountable, where the agent acts and what could go wrong, how its behaviour is measured, and what happens when it misbehaves. When an agent or a send-time model has already sent the wrong messages, the diagnosis for AI automation that sent the wrong messages asks which of these guardrails, permission, limit, sample, source or test, was the one missing.
A payroll software firm splits one process between rules and an agent
Northvane Payroll, an invented payroll software company, receives around 400 inbound demo requests a month. Its first plan gives an agent the whole job: qualify each lead, write and send a follow-up, and update the CRM stage. In a 2-week trial on 60 requests with sending switched off, reviewers find 7 drafts that promise features on the roadmap rather than in the product, and 4 stage changes that would have moved existing customers back to "prospect". The team redesigns. Rules now handle consent, routing by company size and all stage changes. The agent only drafts the follow-up and a 5-line briefing, which the assigned salesperson approves before sending. Over the next month, salespeople approve 312 of 380 drafts with light edits, rewrite 61 and discard 7, and no stage change is made by the agent.
How to decide
Work through four questions for each task.
- Does the task send, spend or change records? If yes, those actions belong to rules. An agent may prepare them, but a rule or a person executes them.
- Can the steps be written down in advance? If yes, write them down and build a workflow. It will be cheaper, more predictable and easier to audit.
- If not, can a person check the result before it matters? If yes, an agent can draft or prepare, with an approval step. If not, redesign the task so a check exists, or keep it manual.
- Can the team test, monitor and own it? An agent needs a test set, logs, a sampling reviewA person reading a random sample of AI answers or generated copy on a fixed rhythm, plus every case that ended badly, and routing findings to the owners. and a named owner. If those do not exist, the task is not ready for an agent.
Measure the result the way any automation is measured: the outcome KPIA metric tied directly to an objective. the task serves, plus the share of agent outputs approved without changes, the share rejected and the incidents logged. A falling approval rate after a model update is the signal to re-test.
Common mistakes
- Using an agent where a workflow would do. If the steps can be written down, a workflow is cheaper, safer and easier to fix.
- Giving an agent send, spend or record-change permissions by default. Grant actions one at a time, starting with drafting.
- Skipping the approval step to save time. The time saved is lost many times over when an unapproved message reaches customers.
- Testing an agent on a handful of runs. Agents vary between runs; test on a varied set, several times each.
- Letting scope creep. Each new task for an agent is a new deployment with its own test and permissions.
- No logs. Without a record of what the agent saw and did, a bad run cannot be explained or prevented.
- Treating rule-based flows as old-fashioned. Predictability is a feature, not a limitation, for anything consequential.
