There’s a version of the documentation conversation that gets skipped in most AI readiness discussions. Everyone agrees you need to document your process before you automate it. What gets less attention is whether the process itself is actually in good shape before you write it down.

Documentation captures what you’re doing. It doesn’t fix how you’re doing it. And if the underlying sales process has gaps, inconsistencies, or stages that were defined years ago and never revisited, documenting it accurately just means you’ve written down a process that isn’t working as well as it should. Then you connect an AI agent to it and amplify the problem.

This post is about the process design work that needs to happen before documentation — and before any AI build. It’s for leaders whose sales process exists but has evolved informally over time, whose stages in the CRM don’t quite match how deals actually move, or who have a sense that something isn’t quite right but haven’t had the time to identify exactly where.

The Difference Between a Process and a Habit

Most sales teams have habits, not processes. A habit is something your team does consistently because it has worked before. A process is something your team does consistently because it has been deliberately designed to produce a specific outcome, and you’ve verified that it does.

The distinction matters for AI because habits are implicit and processes are explicit. When you configure an AI agent, you’re writing instructions. Those instructions have to describe the process in terms specific enough that the agent can follow them without judgment. Habits can’t be transferred that way — they live in the intuition of the people who developed them and can’t be communicated in writing with enough precision to be actionable.

The practical test is this: if you asked your three best salespeople to each independently describe how they approach a new lead from first contact through handoff to a more senior rep, would you get three consistent answers? If the answer is no, you don’t have a process yet. You have three experienced people doing something that works for them individually, which is valuable but isn’t scalable and isn’t something you can hand to an AI.

What a Process Needs to Define

A sales process that’s ready for AI needs to answer a specific set of questions clearly and without ambiguity.

The first is entry criteria. What conditions need to be true for a lead to enter the outreach process? Are you reaching out to every lead that comes in, or only ones that meet certain criteria — company size, industry, persona, lead source? If there are criteria, they need to be specific enough that the agent can evaluate them against the data in your CRM without needing a judgment call.

The second is the goal of each stage. What is the outreach process trying to accomplish at each point of contact? The first message isn’t trying to close the deal. It’s probably trying to establish whether there’s enough interest to continue the conversation. The second message has a different goal. The handoff to a human has a different goal again. Each stage needs a defined purpose and a definition of what success looks like before moving to the next one.

The third is escalation criteria. When does the agent stop and a human takes over? For the client we built this for on a recent episode of AI in Action, this was codified as the three-reply rule — after three exchanges, the conversation goes to a person. But there are also qualitative criteria that need to be defined: what kinds of responses signal enough intent that a rep should step in even before the turn limit, and what kinds of responses signal that the lead isn’t a fit and the conversation should be closed.

The fourth is failure handling. What happens when a lead doesn’t respond? After how many messages and how many days does the agent stop trying? What, if anything, happens to that lead after it ages out of active outreach? These questions don’t have universal answers, but they need to have your answers written down before the build starts.

The CRM as a Process Mirror

One of the most useful exercises for assessing whether your process is actually defined is to look at your CRM stages and ask whether they reflect how leads actually move through your system, or whether they were set up a few years ago and haven’t been revisited since.

In most organizations that have been running Salesforce or HubSpot for a while, there’s meaningful drift between the stages that exist in the CRM and the stages that actually describe the sales motion today. Deals get stuck in stages that don’t have clear exit criteria. Leads pile up in a status that used to mean something but now just means “we haven’t touched this yet.” The pipeline looks full but isn’t actionable because nobody can tell from the data alone what actually needs to happen next.

Before connecting an AI agent to a CRM in this state, the stage definitions need to be revisited. Not necessarily rebuilt from scratch — that’s often more work than is needed — but reviewed against how leads actually move today and updated where the drift has become significant. An agent that’s configured to operate within a CRM structure that doesn’t match reality will produce confused, inconsistent behavior, not because the agent is poorly built but because the map it’s following doesn’t match the territory.

ICP Specificity and Why It Matters More Than You Think

The ideal customer profile is another area where most sales teams have something, but what they have often isn’t specific enough to inform AI outreach meaningfully.

A useful ICP for AI purposes isn’t just a description of the kinds of companies that tend to buy from you. It’s a description specific enough to drive messaging decisions. It should answer: what problem is this persona experiencing right now that makes your product relevant to them? What language do they use when they talk about that problem? What objection do they most commonly raise, and what response moves them past it? What does a successful first conversation with this persona look like, and what does it not look like?

If you can answer those questions for each of your primary personas, your ICP is in good shape for AI use. If you can describe the persona at a demographic level but not at a behavioral or motivational level, the agent’s outreach will be personalized by label but not by substance. An email that says “as a CTO at a mid-market SaaS company” and then delivers generic content isn’t actually personalized. It just looks like it is.

The work of getting ICP specificity to the right level often surfaces useful things about your sales motion that you didn’t have explicit language for before. The exercise of writing down what your best customers actually care about and how your best reps talk to them tends to produce insights that improve human-led sales too, not just AI-assisted sales.

Handoff Design as a Process Decision

One of the decisions that often gets treated as a technical configuration choice is actually a process design choice: how the transition from AI-led to human-led conversation is structured.

The design question isn’t just when the handoff happens. It’s what state the prospect should be in when it happens, what information the rep needs to have at the point of handoff to continue the conversation without a reset, and what the rep’s first action should be when they take over.

For the client in the AI in Action episode, this was designed intentionally. The three-reply rule was chosen because three exchanges was enough to surface genuine interest without running the risk of an extended automated conversation that might drift off-brand. When a handoff triggered, the rep received the full conversation thread in Salesforce and had a clear picture of what had been said, what the prospect had asked about, and where the conversation had left off. The rep’s first action was to continue the thread, not to introduce themselves as if the prior conversation hadn’t happened.

That design — the threshold, the information transfer, the rep’s entry point — is a process decision that needs to be made before the build, not during it. Leaving it to be figured out in configuration adds time and often produces a handoff experience that isn’t as clean as it should be.

How to Know When Your Process Is Ready

The readiness signal for process design is similar to the readiness signal for documentation: can someone unfamiliar with your business follow your process and produce consistent, quality output without needing to make judgment calls you haven’t given them guidance for?

If the answer is yes, you’re ready to document and then build. If the answer is no, the specific places where the answer breaks down are the places where more design work is needed. Those gaps won’t be filled by the AI implementation. They’ll be exposed by it.

The practical path from here to ready isn’t complicated. It’s a few focused conversations with your most experienced salespeople, a review of your CRM stages against how deals actually move, and a session where you write down the answers to the questions in this post. For most businesses, that work takes a day or two of focused attention, not weeks of project planning.

Once it’s done, the documentation work described in How to Document Your Sales Process So AI Can Actually Use It becomes straightforward because you’re writing down a process that’s actually well defined rather than trying to document something that’s still being figured out.

And once the documentation exists, How to Build an AI Lead Nurturing Agent in Salesforce covers what the build phase looks like and what it requires from your team.

Listen to the full podcast episode here.

Related Resources

• Why Your AI Implementation Will Fail Without These Business Fundamentals
• How to Document Your Sales Process So AI Can Actually Use It
• How to Build an AI Lead Nurturing Agent in Salesforce
• What an AI Sales Agent Actually Does All Day
• How to Use AI to Nurture More Leads, Close More Deals, and Scale Your Sales Team Without Hiring