Outlook
Part of The marketing automation outlook for England, with four labels that keep forecasts honest
Four kinds of AI in marketing automation, and which ones have evidence behind them
Assess marketing automation AI applications in England by separating live evidence, supplier announcements, legal permission and controlled tests.
What to take away
- Four kinds of AI in marketing automation cover most of what suppliers sell: large language models for drafting, predictive scoring models for audience and journey logic, natural-language query assistants for analytics, and retrieval-augmented generation agents for customer contact.
- DSIT's AI Adoption Research found natural-language processing and text generation were the most common technologies among surveyed AI adopters.
- Supplier announcements describe product capability but do not prove accuracy, attribution or any business outcome.
- Businesses remain responsible for unlawful conduct by customer-facing AI agents under CMA guidance.
- A conservative option lets AI propose changes while a named approver controls publishing and consent.
- Document models, data and prompts, and stop the application if rollback cannot be demonstrated.
Start with the evidence already available
These four kinds of AI in marketing automation can assist with drafting, audience logic, journey construction, analysis and customer interactions. An England-based organisation still needs to test each narrow task with representative data. Product capability does not create legal permission, and a generated recommendation is not causal evidence.
This article records desk research completed on 5 September 2026. The reviewer examined UK official evidence, regulator guidance and current supplier pages. No platform account, prompt or customer record was tested. No model, message or outcome was either.
Product descriptions are supplier announcements, not verified results. The article makes no adoption, saving or performance forecast.
Observed evidence, UK, fieldwork in 2025: DSIT's AI Adoption Research found that natural-language processing and text generation were the most common technologies among surveyed AI adopters.
Its 3,500 telephone interviews covered businesses generally, not marketing teams, and it could not observe shadow use. That supports investigating text tasks. It does not show that automated copy is accurate or effective.
Observed evidence, UK, 2025 to 2026: the UK Business Data Survey found that overall AI estimates varied with definitions and population. Use that warning when someone presents a broad adoption percentage as evidence for a particular marketing application.
Four kinds of AI in marketing automation, and the evidence behind each
Large language models for drafting: Jasper, Copy.ai and HubSpot Breeze
Draft assistance runs on large language models, the text generators built by OpenAI, Anthropic and Google and resold inside marketing tools. Jasper markets brand voice and campaign copy, with published plans typically from about $39 a month.
Four AI kinds and evidence
Draft assistance
- Example tools
- Jasper, Copy.ai
- Evidence type
- DSIT text generation
- Controlled test
- Editorial protocol
- Main risk
- Accuracy, brand
Audience and journey
- Example tools
- Microsoft, Salesforce
- Evidence type
- Supplier announcement
- Controlled test
- Not stated
- Main risk
- Journey logic
Analysis
- Example tools
- Not named
- Evidence type
- Not stated
- Controlled test
- Not stated
- Main risk
- Unsupported answers
Customer agents
- Example tools
- Not named
- Evidence type
- Not stated
- Controlled test
- Not stated
- Main risk
- Rights questions
Copy.ai positions itself for go-to-market teams, with paid plans from around $49 a month. HubSpot Breeze includes a content assistant inside Marketing Hub, priced by tier and contact limits.
Draft assistance can produce a first version of a subject line or content block. A controlled test should check factual claims, prohibited language, accessibility and brand requirements before a human approves any send. This is an editorial test protocol, not a claim that AI will improve response.
Evidence for draft assistance: DSIT observed text generation as common, which supports investigating the task. Supplier pages describe capability, not accuracy. The controlled test above is the strongest evidence available before you rely on it.
Predictive scoring models for audience and journey logic: Microsoft, Salesforce and Adobe
Audience and journey tools lean on predictive scoring, usually gradient-boosted decision trees such as XGBoost or logistic regression, plus lookalike models that rank contacts by similarity to past buyers.
Microsoft's 2026 Customer Insights release plan is a supplier announcement that lists journey creation and outreach optimisation agents. Some entries are previews and release dates may move.
Salesforce Marketing Cloud Journey Builder with Einstein and Adobe Journey Optimizer translate a brief into proposed rules. Braze and Klaviyo offer journey orchestration. Klaviyo publishes a free tier up to 250 contacts and paid plans from about $20 a month.
Buyers should verify region, edition, current status and the exact records an agent can read or change. Evidence for this kind rests on supplier announcements. None of those announcements proves a business outcome.
Natural-language query assistants for analytics: Mailchimp, Google and Salesforce
Analytical assistants answer plain questions, often by turning them into SQL or a chart query first. Mailchimp announced Analytics AI in May 2026 and described availability and planned functions. Those statements come from Mailchimp. They do not prove attribution, accuracy or a business outcome.
Google Analytics 4 includes Gemini in Analytics for English-language questions, included at no extra cost for many users. Salesforce Einstein Analytics and Adobe Customer Journey Analytics offer similar query features.
Test known questions, reconcile answers to exported records and log unsupported explanations. Evidence for analytical assistance comes from supplier announcements and controlled tests. A generated recommendation is not causal evidence.
Retrieval-augmented generation agents for customer contact: Salesforce, Intercom and Zendesk
Customer-facing agents usually use retrieval-augmented generation, where a language model writes the reply and a search layer pulls the policy or help article behind it.
The CMA's March 2026 guidance says businesses remain responsible for unlawful agent conduct and calls for monitoring and human oversight. A qualified UK consumer-law reviewer must apply that guidance to the actual interaction.
Salesforce Agentforce lists consumption pricing from about $2 per conversation. Intercom Fin publishes from about $0.99 per resolution. Zendesk AI agents publish from about $1 per automated resolution. HubSpot Breeze Customer Agent is also available.
Evidence for customer-facing agents combines legal permission and supplier claims. The CMA guidance sets the responsibility. No causal outcome evidence is supplied with the pricing pages.
Build a bounded 2027 scenario
Editorial scenario, England, horizon ending December 2027: the organisation permits AI to propose content and workflow changes, but not to publish, alter consent status, change prices or answer rights questions without a named approver. This is a conservative operating option, not a prediction about the market.
The ICO's AI guidance says AI can require organisations to reassess governance and risk appetite. The NCSC guidance recommends documenting models, data and prompts and managing the supply chain.
Recheck these records, supplier terms and the product version before expanding autonomy. Stop the application if required evidence, human review or rollback cannot be demonstrated.
Before you act
- Test each narrow task with representative data
- Check claims, language, accessibility and brand
- Verify region, edition, status and records
- Reconcile analytical answers against exported records
- Apply CMA guidance with a UK consumer-law reviewer
- Recheck records, supplier terms and product version
Before you act
- Test each narrow task with representative data before relying on it.
- Check factual claims, prohibited language, accessibility and brand requirements before any send.
- Verify region, edition, current status and the records an agent can read or change.
- Reconcile analytical answers against exported records and log unsupported explanations.
- Apply CMA guidance to the actual interaction with a qualified UK consumer-law reviewer.
- Log the model name and version behind each task so a change can be traced.
- Recheck records, supplier terms and product version before expanding autonomy.
Common questions
What evidence supports using AI for text tasks in marketing?
DSIT's AI Adoption Research, based on 3,500 telephone interviews in 2025, found natural-language processing and text generation were the most common technologies among surveyed AI adopters. That research covered businesses generally, not marketing teams, and could not observe shadow use. It supports investigating text tasks but does not show automated copy is accurate or effective.
Why should a broad AI adoption percentage be treated with caution?
The UK Business Data Survey, covering 2025 to 2026, found that overall AI estimates varied with definitions and population. Use that warning when someone presents a broad adoption percentage as evidence for a particular marketing application. A general figure does not establish anything about a specific use.
What must a business do before giving an AI agent more autonomy?
Recheck records, supplier terms and the product version before expanding autonomy. The NCSC guidance recommends documenting models, data and prompts and managing the supply chain. The ICO guidance says AI can require organisations to reassess governance and risk appetite. Stop the application if required evidence, human review or rollback cannot be demonstrated.
