The future of marketing, 80% automated and 20% conversational.

I want marketers to spend more time deciding where growth will come from and less time building the campaigns needed to get there.

Too much of the working week still goes into moving a brief between teams, creating assets, setting up audiences and checking that everything is ready. That work matters. It also leaves less time to understand why customers buy, where they lose interest and what we should do differently.

Briefly has started by helping us organise the brief and delivery process. The next ambition is bigger, a system that can understand an objective, decide what work is needed and coordinate its delivery within rules we set.

That is where I think marketing is heading.

An 80% automated operating model

My proposed target is for 80% of campaign activity to come from established automated journeys, with 20% initiated by a marketer in response to gaps and opportunities. Both would use automated delivery. The split describes what starts the activity. It is a working target to test, and the right balance will vary by business.

The automated activity would cover the full customer lifecycle. Acquisition leads into consideration and conversion. After the sale, onboarding, product adoption, retention and expansion need the same attention. A signed contract should move the customer into a journey that helps them get value from what they bought.

These journeys should respond to what a person or buying group needs. Someone researching a problem needs different help from a customer who has bought a product and has yet to use a key feature.

The decision should consider their stage, previous activity, recent contact, consent and what Sales or customer service is already doing. A service issue may be a reason to pause a promotional message. Strong buying intent may justify passing an opportunity to Sales.

Sometimes the right action will be to wait.

The other 20% starts with a conversation

I would reserve the remaining 20% for situations that need a fresh decision. A segment falls behind plan. An event creates demand. A new customer problem appears. An existing journey stops producing useful outcomes.

The response should start with understanding the gap. Sending another email may do very little if the problem is the offer, the audience or a slow Sales follow up.

In an illustrative future interaction, a marketer could ask Briefly:

“We are behind our pipeline target in this segment. Show me where the journey is breaking down and propose an action using the customers and prospects we can contact.”

Briefly would examine the available evidence, explain its recommendation and ask for information that materially changes the decision. The marketer could challenge the proposal, change the offer or narrow the audience through the same conversation.

Once the intent and permissions are clear, the system would handle the campaign build. The ambition is to remove the marketer's manual production work. The marketer would still own the objective, judgement and any approval the organisation requires.

A conversation should produce a plan we can inspect, with an audience, message, timing, expected outcome and reason for acting.

Agents with a defined job

I see this working through agents with specific responsibilities. A planning agent interprets the objective. An audience agent checks eligibility against approved data. A content agent works from approved product information and brand guidance. A build agent creates the assets and configures the journey.

A review agent checks the output against the brief, the rules and the expected behaviour. A measurement agent follows what happened and records the result for future decisions.

Each agent needs clear boundaries and evidence of its work. Adding more agents does not automatically make the system better. Predictable tasks may be better handled by a fixed workflow, with agents used where interpretation is needed. Anthropic describes these workflow and agent patterns.

Consent, suppression and contact pressure rules must be enforced by the systems that control delivery. They cannot depend on a model remembering an instruction. The same applies to budget limits, Sales handovers and stop conditions.

An agent reviewing another agent is useful, but it cannot guarantee correctness. We would need tests, logs and a way to stop a journey and roll back changes where possible. The system should earn wider authority through measured performance on a limited scope.

The intelligence should belong to the organisation

I believe the knowledge behind these decisions will become one of an organisation's most valuable assets.

That includes which messages work for different customers, which signals deserve attention, where a journey fails and which decisions improve an outcome. It also includes the judgement that experienced colleagues carry in their heads today.

We should capture that knowledge in a form the organisation controls and can continue using as models and suppliers change. Otherwise, we risk renting the judgement that makes our marketing distinctive.

I would challenge the idea that every organisation needs to train its own large language model from scratch. That is a substantial technical and financial commitment. Owning the intelligence does not always require owning the underlying model.

Existing models can run within an organisation's own cloud environment or on its premises. Cohere documents both private deployment options.

My starting point would be a model connected to approved company knowledge. Retrieval brings relevant information into a request. Further model training may help where there is a proven need for specialised behaviour. These are different choices, and we should test which one solves the problem. AWS compares retrieval and further model training.

Data control also depends on access, retention, contracts and where connected services process information. A private model still needs those controls across its data sources, logs and tools. Calling a model “our own” does not settle those questions.

For me, the valuable intellectual property is the combination of company knowledge, decision rules, tested workflows and learning from outcomes. We should own and protect that knowledge so it can outlast a particular model or supplier.

What the marketer owns

This changes what we ask of a marketing team. Understanding customers, setting priorities and judging results become more central to the job. Manual production should take up less of it.

We should judge progress through customer and business outcomes, alongside quality, cost and the number of exceptions that need human attention. Producing more campaigns would be a poor measure on its own.

My ambition for Briefly is that a marketer can identify a growth problem, have a useful conversation about it and see an agreed response delivered without rebuilding the process each time.

The team would spend the time it gains testing better ideas and fixing the parts of the customer journey that still fail people. That is the work I want automation to make room for.