Eight marketing automation strategy mistakes, beginning with calling activity an outcome

These eight marketing automation strategy mistakes each stop a team deciding whether an automated change worked or should continue. Each one also surfaces when a marketing automation platform is selected, reviewed or run, because a vendor demo, trial or review has to answer for it. The non-ranked list draws on current government, regulator and standards material.

Each mistake turns into a question to put to the platform you are reviewing: what outcome does a journey claim, which denominator backs each rate, which permission fields exist per channel, how are errors and withdrawal handled, and what limits apply to any AI component. Ask for evidence rather than a demo.

Several of these mistakes only become visible when a platform is put under review: a journey canvas that hides exclusions, a metric card without a population, a consent field copied across channels, or an AI score with no challenge route. The checks below say what to ask for and what to keep as evidence.

What to take away

1. Calling activity an outcome

Messages sent and workflows launched are outputs. They do not establish a service or commercial effect. HM Treasury's 2026 Magenta Book separates inputs, activities, outputs and outcomes, and asks evaluators to expose alternative explanations. Write the causal steps and the evidence needed at each one.

Worked example: a UK retailer's welcome journey sends an email, then an SMS after 48 hours if the email is unopened. The dashboard reports open and click rates, which are outputs. The corrected metric is repeat purchase rate within 90 days, with the denominator set as new customers who received at least one welcome message.

What good looks like: a metric definition sheet naming the outcome, denominator, exclusions, owner and review date.

2. Starting with a platform feature

A visual journey builder makes it easy to automate the wrong task. The Magenta Book evaluation sequence begins by understanding the intervention, evidence and question before selecting an approach. Define the customer problem, eligible population, exclusions and accountable owner first. Tool capability belongs after the strategic choice and operating specification.

Journey canvases and template galleries, of the kind sold by Mailchimp, Klaviyo and HubSpot, sit behind this mistake. When you read a platform review, check whether it explains how the canvas handles exclusions, per-channel permission fields and version history, not only how quickly a journey can be drawn.

What good looks like: a problem statement covering the customer problem, eligible population, exclusions and accountable owner, agreed before tool selection.

3. Using a metric without a denominator

A percentage is not reproducible unless the event, population, period and exclusions are known. GOV.UK guidance on performance data distinguishes raw measurements from rates and recommends deciding what to collect early. Give every metric an owner and data-quality check.

What good looks like: a metric dictionary entry listing the event, population, period and exclusions, with a named data-quality check.

4. Copying one permission rule across channels

Email, SMS, live calls and tracking technologies have different PECR questions. The ICO's direct-marketing hub routes organisations to separate guidance for electronic mail, calls and storage technologies. Record lawful basis, channel analysis, consent or exception and preference evidence in distinct fields.

The ICO's PECR penalties are capped at 500,000 pounds.

What good looks like: a permission field schema with separate fields for lawful basis, channel, consent or exception and preference evidence, each with a review date.

5. Treating a claim as harmless copy

An automated message can repeat an unsupported comparison at scale. The CAP Code misleading-advertising section covers substantiation, material omissions and comparisons in non-broadcast communications. Keep the approved claim, evidence, limitations and expiry together; pause the message when the support changes.

What good looks like: a claim file holding the approved wording, substantiation, limitations, expiry date and owner.

6. Testing only the happy path

Duplicate events, blank fields, inaccessible forms and failed suppressions are part of the design. WCAG 2.2 sets criteria for matters including keyboard access, focus order, labels and status messages in the W3C standard. Test errors, refusal, withdrawal, manual override and rollback before testing volume.

What good looks like: a test plan naming an owner and a pass condition for errors, refusal, withdrawal, manual override and rollback.

7. Giving an AI component open authority

An AI-generated score or message needs a defined purpose, permitted inputs, human review, logging and a route for challenge. The ICO's AI and data-protection hub links risk and explanation guidance for organisations. A vendor label does not settle accuracy, fairness or legal responsibility.

What good looks like: an AI register entry stating the purpose, permitted inputs, human review step, logging and the challenge route.

8. Letting a pilot continue by default

The Magenta Book treats evaluation as work that informs decisions before, during and after implementation. Set the end date, review owner and stop conditions before launch. A failed suppression, unrecoverable denominator, unexplained guardrail movement or unmanageable exception queue should trigger a pause. At review, choose continue, revise, stop or retire and write down why.

What good looks like: a pilot charter with the end date, review owner, stop conditions and the four decision options.

Keep a repair ledger

For each mistake found, record the affected rule, evidence, risk owner, immediate containment and decision date. Do not rewrite history by replacing an old dashboard definition or permission field without retaining its effective period.

Review the ledger before approving expansion. Repeated defects may indicate that the operating model, rather than one message, needs redesign. Expansion can wait for a decision supported by actual results.

Before you act

Common questions

Why is calling activity an outcome a mistake?

Messages sent and workflows launched are outputs, not outcomes. They do not establish a service or commercial effect. Write the causal steps and the evidence needed at each one.

What should be defined before selecting a platform feature?

Define the customer problem, eligible population, exclusions and accountable owner first. Tool capability belongs after the strategic choice and operating specification. The Magenta Book evaluation sequence begins by understanding the intervention, evidence and question before selecting an approach. Starting with a platform feature makes it easy to automate the wrong task.

How should an AI component be governed in marketing automation?

An AI-generated score or message needs a defined purpose, permitted inputs, human review, logging and a route for challenge. The ICO's AI and data-protection hub links risk and explanation guidance. A vendor label does not settle accuracy, fairness or legal responsibility. Open authority for an AI component is a mistake.