AI agents in lifecycle marketing: what we let them do

A client's marketing director asked us a fair question in a planning session this spring. The team had been shown a demonstration where an automated system built an audience, wrote the email, chose the offer and scheduled the send, all from a one-line instruction. Could we do that, and if not, why not?

The honest answer is that most of the pieces in that demonstration work. The part that does not work is the arrangement: one system making four kinds of decision, with no separation between them and no record of which decision caused which outcome.

We do use automated decisioning in the programmes we run. What we hold to is a boundary that has not moved much in two years. The system may decide who, when, and in what order. A person decides what it costs, what it claims, and whether we are allowed to contact this person at all.

Three different decisions get called "the campaign"

Untangling this is most of the work. What looks like a single decision is usually three, with very different consequences when it goes wrong.

The first is decisioning: which message a given person should receive, from a set of candidates, and when. It is reversible, it is measurable, and a mistake costs one slightly less relevant email.

The second is authorship: what the message actually says. A mistake here can be a false claim, a tone that damages the brand, or a statement the business cannot stand behind.

The third is economics: what the offer is worth, who qualifies, and what margin the business is prepared to give up. A mistake here is expensive immediately, and often at scale before anyone notices.

Most disagreements about automation in CRM are really disagreements about which of these three is being discussed. Once they are separated, the answers become much less controversial.

What we let an automated system decide

Inside a bounded set of options, ranking and timing are well suited to automation, and humans are not especially good at them.

• Ranking candidate messages. Given a person and a list of messages they are eligible for, choose the one to send. This is next best action as it has been practised for years, and the scoring underneath it is usually an ordinary propensity model rather than anything new.
• Timing within an allowed window. Choosing the hour of send for an individual, inside a window a person has defined and inside the contact policy.
• Allocating scarce contact slots. When three journeys want the same person this week and the cap allows one, deciding which one wins, according to a priority order a person has set.
• Selecting content blocks from an approved library. Choosing which dynamic content module to show from a set that has already been reviewed and signed off.
• Drafting. Producing variants, subject lines and summaries for a person to review, edit and approve before anything is scheduled.
• Operational work. Writing a first version of a segment query, reconciling two reports, flagging a journey whose volume has dropped, summarising a week of test results.

The common feature is that the system is choosing from options someone has already approved. It is picking from a menu, not writing one.

Drafting deserves its own note, because that is where teams get the fastest return and the most trouble. A model can produce forty subject lines in a minute. Whether any of them sound like your brand depends entirely on whether somebody has written the voice down first. That is the argument we made in keeping a brand voice consistent at scale. Without that, automated drafting produces volume and a slow drift towards a generic register that nobody notices until a customer mentions it.

What stays with a person

Four areas we do not automate, and the reasoning is the same in each case: the cost of being wrong is not proportional to the benefit of being right.

Offer economics. Discount depth, who qualifies, and how often. A system optimising for conversion will find that a larger discount converts better, because it does. Margin sits outside its view unless somebody models it carefully, and by the time a mispriced offer shows up in the management accounts it has been sent to a lot of people.

Brand and factual claims. Anything asserting a fact about the product, a price, a delivery promise, a regulated statement or a comparison with a competitor. These need a human who is accountable for their truth. In regulated categories that accountability is a legal requirement rather than a preference.

Consent and data use. No automated system decides who may be contacted, on what channel, on what lawful basis. Under GDPR and its equivalents, the record of consent and the logic of suppression are the things you will be asked to produce, and they must be explainable by a person who understands them. Suppression logic in particular should be deliberately dull and deliberately human.

Anything irreversible at scale. Sending to the whole file, changing a global contact policy, altering data in the source system, deleting records. A person presses that button.

Guardrails that make the rest safe

The division of labour only holds if it is built into the system rather than written in a policy document nobody reads.

1. A bounded action space. The system chooses from an enumerated set of approved actions. It cannot invent a new message type, a new audience or a new offer. Everything it can do is something a person has already said yes to.
2. Hard limits it cannot override. Frequency caps, suppression lists, send volume ceilings and market restrictions are enforced outside the decisioning layer, so no optimisation can trade them away.
3. A logged decision for every send. Which options were considered, which was chosen, on what score, under which model version. Without this you cannot debug a strange outcome, and you cannot answer a customer asking why they received something.
4. Staged exposure. New automated logic starts on a small share of the audience, with a control group held back, and grows only when the numbers justify it.
5. A stop control that works in seconds. One person, one switch, reverting to the previous behaviour without an engineering deployment.
6. A named owner. Every automated decision has a human whose job it is to defend it. If nobody can be named, the automation does not ship.

None of this is novel. It is the same discipline that applies to any automated system touching customers, applied to a newer set of tools.

Read a sample of what it did, every week

Automated decisioning fails quietly. Performance does not collapse; it drifts, and the metrics that would reveal the drift are averages that hide it.

So we sit down weekly and read actual decisions. Twenty or thirty at random, plus the ones at the extremes: the person who received the most messages, the person who received none, the offer that went to the smallest segment. You are looking for decisions that are defensible in aggregate but obviously wrong for that individual.

The accountability question is the one still waiting for a good answer. A system can hold a decision, but it cannot own one. When an automated send goes wrong in a way that reaches a customer or a regulator, the organisation needs a named person who understood the logic, agreed to it, and can explain it. Designing that person into the process from the start is much easier than appointing them afterwards.