A safer way to use AI for CRM follow-up

What shipped

A recent report on a powerful personal AI assistant raised a useful question for SME owners: how much access should an AI tool get before it becomes a business risk? The assistant in question connects to email, messages, calendar, device activity and other sources so it can act on a user’s behalf. Testers praised the convenience, but also raised concerns about privacy, data retention and the ability to make commitments for the user.

That is not a reason to ignore AI in sales. It is a reason to design the workflow properly.

For an Irish SME sales team, the practical opportunity is much narrower and more useful: a CRM follow-up and pipeline hygiene assistant. Instead of giving a broad agent open-ended permission to roam across every system, the workflow reviews recent emails, meeting notes and CRM activity, then produces a list of suggested next steps.

It might flag that a quote was promised but not sent, a renewal date is missing, or a prospect asked for a call next week and no task was created. It can draft a short follow-up email, suggest a CRM note, or propose an updated deal stage. Crucially, it does not send emails or change CRM fields on its own.

The named human review step is the Sales Review Gate. A salesperson or sales manager checks the suggested updates, edits them where needed, then approves any message or CRM change before it is used operationally. The AI is decision support, not an autonomous sales rep.

This kind of controlled workflow is a good fit for Artellis workflow design, because the value is in the hand-off, the review point and the rule set, not in buying a large AI platform.

Who should care

This matters most to owner-managers, sales managers and operations leads in SMEs where the CRM is always nearly right, but rarely trusted.

The pattern is familiar. A salesperson has a good call, promises to send details, then moves on to the next meeting. A manager asks for a forecast and gets a mix of stale deal stages, missing close dates and notes that live in someone’s inbox. Follow-ups depend on memory, and the CRM becomes a reporting chore rather than a working tool.

The cost is not just admin time. Forgotten follow-ups make the business look disorganised. Deals stall because the next step is unclear. Managers spend pipeline meetings asking for updates that should already be visible. New sales hires struggle because too much context sits in private inboxes and informal notes.

A CRM follow-up assistant is most useful when the sales process is already partly defined. It does not need a perfect CRM. It does need agreement on basic rules, such as what counts as a live opportunity, when a proposal should trigger a follow-up, and which fields matter for forecasting.

It is not a good first project if the business has no consistent sales stages, no shared CRM ownership, or no appetite for human review. In that case, the starting point is workflow clean-up, not automation.

The privacy lesson from broad personal assistants is also important. Sales inboxes often contain customer pricing, contract terms, credit notes, personal data and commercially sensitive information. An SME should avoid tools that require excessive access, unclear retention terms, or permission to act without approval. A smaller, bounded assistant that only reads defined sources, produces reviewable outputs and logs decisions is usually the safer and more useful choice.

A worked example

Take a 35-person Irish services firm with four salespeople and one working sales manager. They use a CRM, but activity is patchy. Some meetings are logged, some are not. Email follow-ups are written manually. Friday pipeline reviews often begin with the same question: what actually happened with this deal?

The business starts with one salesperson’s last 30 days of data. That includes CRM activity, recent sent and received emails linked to active opportunities, meeting notes and open tasks. The aim is not to analyse every customer interaction forever. The aim is to prove whether the assistant can find useful gaps and reduce manual admin.

The first version reviews each active opportunity and produces a short queue for the salesperson:

1. Suggested CRM note based on the latest email or meeting note.
2. Suggested next action, such as send proposal, book demo, confirm budget, or follow up after no reply.
3. Suggested due date, based on what was promised or on agreed follow-up rules.
4. Draft follow-up message, grounded in the actual conversation and company-approved wording.
5. Confidence flag, showing whether the suggestion is clear, uncertain, or needs manager review.

For example, an email from a prospect says: “Thanks for the call. Please send the revised support option and we can discuss internally before Friday.” The CRM has no next task and the opportunity is still marked as early stage.

The assistant suggests a CRM note: “Prospect requested revised support option following call. Internal discussion expected before Friday.” It suggests a task: “Send revised support option by tomorrow morning.” It drafts a short email with the promised material referenced, but leaves placeholders where pricing or attachments must be checked.

At the Sales Review Gate, the salesperson reviews the item, corrects the note if needed, attaches the right document, and approves the CRM update and email. If the assistant is wrong, the salesperson rejects the item and selects a reason, such as wrong deal, missing context, or no follow-up needed.

After two weeks, the manager can measure practical signals: time saved on CRM admin, number of missing follow-ups found, rejected suggestions, edited drafts and whether pipeline review takes less time. A realistic value signal is saving 2 to 4 hours per week per salesperson while reducing forgotten follow-ups.

This is also where AI literacy matters. Salespeople need to know that the assistant can summarise and draft, but it may misread context, overstate certainty or miss a side agreement. A short enablement session and written rules help the team use it safely. For more general guidance on choosing contained AI use cases, see the Artellis AI insights library.

Where I'd start

Start small and make the review process explicit.

Pick one salesperson and one CRM pipeline. Pull the last 30 days of CRM activity, emails and meeting notes for open opportunities only. Exclude anything that is clearly outside the sales workflow. Define five to ten follow-up rules in plain English. For example: if a proposal was sent and there is no reply after five working days, suggest a follow-up. If a customer asks for pricing, create a task unless one already exists. If an email mentions a renewal date, suggest the CRM field to update.

Then build a reviewable list, not an autonomous agent. The first output should be a table or queue showing the deal, evidence, suggested action, draft wording and approval status. Every item should point back to the source message or note so the salesperson can check it quickly.

Decide what the assistant is not allowed to do. In most SMEs, it should not send email, change deal value, change close date, mark a deal as won or lost, or create a contractual commitment without human approval. Those actions should sit behind the Sales Review Gate.

Finally, test with real users before rolling out. Track accepted suggestions, edits, rejections and time saved. If the assistant creates noise, tighten the rules. If it finds useful missed actions, extend it to the next salesperson.

The best version of this workflow is boring in the right way. It helps the team remember what was promised, keeps the CRM cleaner, and leaves the commercial judgement with the people who own the customer relationship.