From rules to models
A rule is explicit: a person wrote the condition, and anyone can read it and predict what it will do. A model is learned: its behaviour comes from data, and the people who deployed it can describe its tendencies but not enumerate its outputs. That difference is the whole design problem. Rules fail loudly and consistently, usually at the edges nobody thought about. Models fail quietly and plausibly, often on inputs that look ordinary.
The practical answer is not to replace rules with models but to decide, for each step in a workflow, which kind of logic belongs there. Steps that need judgement over messy input — what is this reply about, which product does this enquiry concern, how should this opening line read for this person — are where models earn their keep. Steps that carry consequences — suppressing a contact, changing a lifecycle stage, issuing a discount, sending to a whole segment — stay as rules, with the model's output as one input among several. When the choice is between a fixed workflow and an agent for a whole task, rule-based automation vs AI agents sets the two side by side with a decision rule. The NIST AI Risk Management Framework describes the same idea in its own vocabulary: map where AI is used, measure how it behaves, manage the risks and govern the whole with named accountability. Marketing teams rarely need the full framework, but the habit it encourages, of treating every model as a component with known failure modes rather than as a clever colleague, is the right one.
A useful way to see a workflow is as a chain of decisions with a trigger at the front and an action at the back. AI can sit at three places in that chain: at the trigger (predicting that something will happen), in the middle (classifying, scoring, routing, composing) and at the action (drafting and, with permission, acting). The rest of this guide walks those positions in turn. The automation governance guide covers naming, ownership and change control for the workflows themselves; everything here assumes that foundation is in place.
Classification and routing
The oldest problem in automation is that most useful input is text. Form messages, email replies, chat transcripts, support tickets, survey comments and social mentions all arrive as prose, and a rule can only match keywords in them. Classification models read the prose and return a label — topic, intent, sentiment, urgency, product line, language — that a rule can then act on. Routing is the natural consequence: a reply classified as a request to unsubscribe goes to the suppression step; one classified as a pricing question goes to sales; one classified as a complaint skips every nurture and lands in a human queue.
Three design choices matter more than the model. The first is the label set. Keep it small, mutually exclusive where possible, and anchored to an action: a label that does not change what happens next is a label nobody will maintain. The second is the confidence threshold. Most classifiers return a score alongside the label, and the workflow should treat low-confidence results as "unknown" and route them to a person rather than guessing. The third is the fallback. When the model is unavailable, times out or returns something outside the label set, the workflow needs a defined path, usually the one it followed before the model existed.
Classification also feeds lead scoring. A scoring model that reads enquiry text alongside firmographic and behavioural fields will usually separate serious buyers from browsers better than points-based rules, but it is also harder to explain to the sales team receiving the leads. Keep the explanation available: the fields and signals that drove the score should be visible on the record, because a lead handed over with a number and no reason is a lead that sales will learn to ignore.
Personalisation at send time
Classic personalisation is merge fieldsA placeholder in an email or message template replaced with a value from each contact's record, such as a first name, with a fallback when it is empty. and segments: the same template, with the name and a few blocks swapped by rule. Send-time personalisation uses a model to compose or select content for each recipient when the message is assembled: the subject line, the product block, the opening sentence, the recommended next article, the send hour. The appeal is obvious. The risk is that the message a given person receives has never been seen by anyone on the marketing team. The rule-based kind is covered on its own terms in the guide to personalising email content, from merge-field fallbacks to testing a personalised version against a generic one.
The way to keep that risk acceptable is to constrain the model's freedom to the parts of the message where variation is cheap and errors are survivable, and to lock everything else. Offers, prices, legal copy, claims about the product and the unsubscribe footer are fixed. The model chooses between approved variants, or writes within a tight brief, for the lines where tone and relevance matter. Every generated element goes through automated checks before send: banned terms, claim patterns, length, links that resolve, placeholders left unfilled. A sample of sent messages is read by a person each week, which is the only way to notice a model that has drifted into a tone the brand would never use.
Where send-time personalisation uses customer data in the prompt, the data-handling rules in the disclosure and data privacy guide apply: only the fields the message needs, nothing sensitive, and a vendor whose terms say the inputs are not used for training. The email automation hub covers the flows themselves — welcome, abandoned cart, re-engagement — where this kind of personalisation is usually tried first, and the nurture sequences guide covers the longer lead-facing flows where it is tried next.
Agents that draft and act
An agent, in the sense marketing teams now meet it, is a model given a goal, a set of tools and permission to take several steps without a person between them: read the record, look up the order, draft a reply, update a field, schedule a follow-up. The difference from a classifier or a composer is that the agent decides what to do next, not just what to say. That is powerful for the long tail of one-off tasks that never justified a workflow, and it is the use with the widest range of ways to go wrong.
The governing idea is to separate drafting from acting. Drafting is low-risk: an agent that prepares a reply, a segment definition, a report summary or a campaign brief for a person to approve saves time without touching anything. Acting is where permissions matter, and permissions should be granted per action, not per agent. An agent might be allowed to tag a contact, add a note and draft a reply, but not to send it, change a stage, apply a discount or delete anything. Each action it is allowed should be logged with the input it saw and the reason it gave, so that when something goes wrong, and something will, the trail exists.
Scope creep is the characteristic failure. An agent set up to draft replies to product questions is soon asked to handle complaints, then refund requests, then anything in the inbox, and the permissions that were sensible for the first task are now applied to the last. Treat each new task as a new deployment with its own review. The accuracy and review workflows guide covers how to set review tiers so that low-risk drafting flows and high-risk actions are checked differently.
Prediction-driven triggers
Traditional triggers are events: a purchase, a form fill, a date, a score crossing a line. A predictive trigger fires on a probability: this customer is likely to churn, this lead is likely to convert within the quarter, this subscriber is likely to disengage, this account is ready for an upsell. Models for these predictions are now a common feature of automation and customer data platforms, and they let a workflow act before the event rather than after it.
The useful question is what the workflow should do with a prediction, because a prediction is not a fact. A churn-risk flag that triggers a discount will, over time, teach the model's inputs that looking at the cancellation page earns a discount. A conversion-likelihood flag that pulls a lead forward will frustrate sales if the model is confident and wrong. Good designs use predictions to prioritise and to time, not to decide: the churn-risk cohort gets a service check-in and a content series rather than money; the high-likelihood lead gets a faster human touch rather than an automatic quote.
Two measurement habits keep predictive triggers honest. The first is a holdout: a slice of the flagged population that receives nothing, so the team can see whether the intervention changed the outcome or merely accompanied it. The second is calibration review: compare predicted probabilities with what happened, by cohort, each month, and retire or retrain a model whose predictions have drifted from reality. Lifetime value (LTVTotal value a customer brings over time.) predictions in particular drift as the product, pricing and acquisition mix change, and a model trained years ago will confidently describe customers who no longer exist. The LTV and CAC guide explains what those predictions are built on.
Guardrails and approvals
Guardrails are the controls around a model's output; approvals are the points where a person signs off before an action happens. The right combination depends on what the step can do, not on how clever the model is. The table sets out the usual pattern.
| AI use in the workflow | What can go wrong | Guardrail | Approval |
|---|---|---|---|
| Classifying inbound text | Wrong label; low-confidence guesses | Fixed label set; confidence threshold; an "unknown" route | None for routing; a person for anything that suppresses or escalates |
| Lead scoring with a model | Opaque scores; drift as the market changes | Visible contributing signals; monthly calibration check | Sales and marketing review score bands each quarter |
| Send-time copy | Off-brand tone; invented claims; broken links | Locked offer, legal and footer blocks; banned-term and claim checks; link validation | Weekly human sample; full review for new templates |
| Product or content recommendation | Irrelevant or unavailable items; sensitive inferences | Stock and eligibility filters; exclusion lists; category fences | None per send; review of the exclusion rules |
| Agent drafting | Confident errors in facts or policy | Source-grounded prompts; draft only, never send | A person approves every draft until the error rate is known |
| Agent acting | Wrong record, wrong action, runaway loops | Per-action permissions; rate limits; dry-run mode; full logging | Approval per action class; owner reviews logs weekly |
| Predictive trigger | Acting on a probability as if it were a fact | Prioritise and time, do not decide; holdout group | Owner reviews holdout results before scaling |
| Model or vendor change | Behaviour shifts silently | Versioned prompts and models; regression test set | Owner signs off on test results before the switch |
Two rows deserve emphasis. Rate limits and dry-run modes for acting agents are what turn a bad hour into a bad minute: a workflow that can only take a bounded number of actions before pausing for review cannot empty a list or spam a segment. And the regression 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. — a fixed collection of inputs with known good outputs that is re-run whenever the prompt, model or vendor changes — is the cheapest control on this list and the one most teams skip. When a workflow with an AI step has already sent the wrong messages, the diagnosis for AI automation that sent the wrong messages traces one message back to the guardrail that was missing. Before such a step is switched on, the AI automation go-live checklist turns this table into a gate to tick, from locked blocks and per-action permissions to a stored test set and a sampling plan. The platforms' own automation, bidding and audience expansion included, has its own documented controls; the AI in ad platforms guide covers those as the help centres describe them. The tutorial on writing a sampling plan for an automated AI use writes a sampling plan for one use, from the reader to the pause rule.
A subscription box adds AI to its retention flow
A meal-kit subscription business with about 40,000 active subscribers rebuilds its retention workflow. A churn model scores every subscriber weekly; the top 8% by risk enter a flow, with a 10% holdout that receives nothing. Flagged subscribers get a service check-in email whose opening line and recipe suggestions are composed at send time within a locked template; the offer block, where there is one, is a fixed rule based on tenure, not a model choice. Replies are classified into five labels; "pause request" and "complaint" go to a person within an hour, and the rest are answered by an agent that drafts only, with a team member approving each reply for the first six weeks. After a quarter, the flagged-and-contacted group retains 4 percentage points better than the holdout, the agent's drafts are approved unedited 87% of the time, and the team moves routine "delivery day" replies to auto-send while keeping every other label on approval. The weekly sample reading catches one drift: the model began opening emails with the same jokey phrase, and the prompt is tightened.
Measuring automation
An automated workflow is a programme with a cost and a result, and AI inside it adds both. Measure at three levels. At the workflow level, the question is whether the flow moves its objective: conversion rate (CVRShare of visits or clicks that convert.) for a lead flow, retained or reactivated customers for a retention flow, measured against a holdout rather than against the previous period. At the component level, the question is whether the AI step does its job: label accuracy on a reviewed sample, the share of low-confidence routes, the share of drafts approved unedited, calibration of predictions against outcomes. At the operational level, the question is cost and risk: time saved per task, review hours consumed, incidents and their severity, and the volume of actions taken without a person.
Pick one KPIA metric tied directly to an objective. per level and report them together, because they move against each other. A flow whose approved-unedited rate climbs while its conversion rate falls is a model that has learned to write acceptable, ineffective copy. A flow whose incident count is nil and whose review hours are also near nil is probably one nobody is looking at. The AI for analysis and reporting guide covers using AI in the reporting itself, with the same caution about verification, and the CRM connection guide covers getting the outcome data back into the systems where these measures are built. The applications hub sets this guide beside its siblings on content, creative, analysis and ad platforms; the governance hub holds the policy that all of them work under.
Common mistakes
- Replacing rules with models wholesale. Consequential steps stay deterministic; models supply inputs.
- No "unknown" route. A classifier forced to pick a label will pick a wrong one with confidence.
- Letting agents act before their drafts have been measured. Drafting first, acting later, permission by action.
- Treating predictions as events. A probability prioritises and times; it should not, on its own, issue money.
- No holdout. Without one, the retention flow's result is indistinguishable from what would have happened anyway.
- Leaving offers and legal copy to the model. Offers, claims, legal copy and footers are rules, never generated.
- Changing the model or prompt without a regression set. Behaviour shifts quietly; the test set is how you notice.
- Scope creep without re-review. Each new task an agent takes on is a new deployment.
