Before you buy an AI agent, fix the journey it will inherit.

If marketing and sales cannot agree when a prospect should stop receiving nurture emails, giving an AI agent more freedom will not resolve the disagreement. Someone still has to own that decision.

Marketing teams can reduce manual work with AI. Leaders still need to decide how work is organised, who approves decisions and how results are measured. Access to a model is only the starting point.

Five stages, used as a working model

I use “World of CRM” as a five stage framework for B2B SaaS marketing. This is my proposed model, not an industry standard or a claim that every business follows the same path. Assess a particular journey, not the whole company with one score. A business might run strong onboarding while still sending generic renewal messages.

Stages three and four are separate here for a reason. A journey can branch within email without being coordinated across product, sales and other channels.

1. Broadcast: the calendar decides

The starting point is a newsletter, product update or promotion sent to a selected list. Segmentation may be broad and the send date usually follows a campaign plan.

This can be appropriate for an update that the audience expects. It becomes limiting when the same treatment is used for a new trial, an active customer and someone already talking to sales.

Keep delivery, complaints and unsubscribes visible. Define the intended action as well. Open rate alone cannot tell you whether a campaign helped the business, particularly given the limits introduced by Mail Privacy Protection. Apple support

2. Sequences: the lifecycle decides

A signup, trial start or meeting creates a sequence with a defined goal. Messages can use the person's role, product plan or completed actions. An onboarding sequence should stop or change when its purpose has been achieved.

For an illustrative trial, measure the proportion of eligible accounts that reach a meaningful activation event and then become paying customers. Define that event with the product team. A login is evidence of access, but may tell you little about whether the customer experienced value.

The next step is to connect reliable product events and agree the entry and exit rules. Writing seven emails before agreeing those rules creates work that may not be needed.

3. Branching: behaviour changes the path

The journey now responds to what has happened. Someone who completed setup receives a useful next step. Someone who has not completed it receives help. A person who books a meeting exits the demo request sequence.

This is still achievable with explicit rules. It does not require an autonomous agent. Keep the branches understandable and test missing events, late events and repeat events, as well as the expected path.

The measure is whether the branch improves the outcome against an appropriate comparison. A busier journey diagram is not evidence of maturity.

4. Coordination: the account has one conversation

Email, in product messages, sales activity and paid media should work from compatible customer and account information. They do not all need to act on every signal.

An illustrative demo request could create a sales task and pause the relevant prospecting sequence. An active support issue might suppress an expansion offer. The team needs to agree which communications take priority and which service messages should continue.

Look at accepted opportunities, conversion, sales cycle length, retention and expansion, using measures that fit the journey. Record pipeline influence if it helps, but do not present attribution as proof of incremental revenue.

For many organisations, dependable coordination is the most useful near term target. It creates the conditions for more flexible decisions later.

5. Bounded agency: the system chooses an allowed action

The goal is a system that chooses the next useful action from current customer information, within agreed permissions. “Wait” and “ask a person” must be valid outcomes. A generated email is only one possible action.

Parts of this exist today. Adobe documents AI offer ranking, eligibility controls and frequency caps. On the same page, it marks some journey path, channel and arbitration features as coming soon. That mix is a reason to validate a specific use case in the selected product, rather than buy an umbrella promise of autonomous marketing. Adobe product capabilities

An account visiting pricing twice is a signal to interpret, not proof of a buying decision. Check fit, existing conversations and product behaviour before acting. Nor does every event need to reach the CRM instantly. Agree the freshness required for the decision and measure the actual delay across the systems involved.

The leadership work starts before the agent

I would put one business owner in charge of the journey and its outcome. Marketing operations, product, sales, data and security need defined responsibilities around that owner. People should know who can change a rule, stop a journey and approve a wider scope of automation.

Technology companies may have useful starting conditions, such as product event data and software engineering skills. I would not assume that makes organisational change easy. Separate budgets, conflicting incentives and unclear ownership still have to be addressed.

Start by letting the system recommend actions while the team reviews them. Record disagreements and failure cases. Allow automatic execution only for a bounded use case once its quality is demonstrated. Keep an audit trail, an escalation route and a way to stop it. NIST's voluntary AI Risk Management Framework provides a reference for managing risk across the design, use and evaluation of AI systems. NIST AI RMF

Make the return testable

Some benefits will take time to prove. Set a baseline, fund the work in stages and agree what evidence will justify the next investment.

Build the case from identifiable costs: software, data connections, migration, training, ongoing review and model usage. Separate released staff time from a cash saving. Time becomes a financial saving only when a corresponding cost is removed or avoided. Any revenue benefit needs its own evidence and must not be counted twice.

Here is an illustrative three year scenario, not a benchmark or a forecast. Suppose setup costs £120,000 and ongoing costs are £80,000 a year. Suppose each year produces £60,000 of verified cost reduction and £90,000 of incremental gross profit. Total costs are £360,000 and total benefits are £450,000. The undiscounted net benefit is £90,000, or 25% of cost. The result changes materially if adoption is slow or those benefits fail to appear. A full investment appraisal should also consider timing, discounting and downside scenarios.

I would begin with one journey, such as trial activation. Agree the baseline, a comparison group and the conditions for stopping or expanding the pilot. Check quality and customer impact alongside effort. Then fund the next step on the evidence.

Give the system more responsibility as the evidence shows it can handle it.