AI sales automation: which tasks to give rules, AI or a person

This guide is for sales operations leads, sales managers and founders deciding which parts of their sales process to automate with AI. It sorts sales tasks into those that suit simple rules, those that suit AI and those that should stay with a person, then covers a spec for each workflow, the ways AI automation fails, published costs, a worked time study and a two-week test. For software that pursues a goal across many steps on its own, see the guide to AI sales agents.

Where the time goes

Salesforce's State of Sales report (7th edition), from a survey of 4,050 sales professionals in 22 countries in August and September 2025, found that reps spend 40% of an average week selling and 60% on other work, which the report describes as nonselling work "like data entry and prospecting." That 60% is the target for automation. The aim isn't to remove reps from the sale; it's to give them back hours now spent on typing, copying and chasing.

Rules, AI or a person?

Rules-based automation follows instructions you wrote: when this happens, do that. It's predictable, cheap and easy to audit. AI automation reads and writes language, so it can handle a reply that says "not now, try me after our budget review in March", but it can also be wrong in ways a rule never is.

A good test for any task: if the AI gets it wrong, who notices, and when? If the answer is "the customer, next week", keep a person in the loop.

For the research row, Bebop does this for each account in a list: it finds matching companies and contacts, explains why each fits and writes a playbook your reps can check before outreach.

Write a spec for each workflow

Vendors demo automations working. Your spec decides what happens when they don't. One page per workflow is enough.

Workflow: [name, e.g. post-meeting follow-up]
Owner: [person accountable]      Review date: [date]

TRIGGER: [what starts it, e.g. a meeting with an external attendee ends]
INPUT: [what it reads, e.g. transcript, CRM opportunity, attendee list]
STEP TYPE: [rule / AI / person] for each step:
  1. [step] ([type])
  2. [step] ([type])
OUTPUT: [what it writes or sends, and where]
APPROVAL: [who approves before anything reaches a customer or changes a deal]
NEVER: [fields or actions it may not touch, e.g. close date, amount, pricing]
FALLBACK: [what happens if a step fails or confidence is low, e.g. task to the rep]
LOG: [where every run is recorded]
METRIC: [minutes saved per run, error rate on a weekly sample of 20]
COST CAP: [maximum actions or spend per month]

How AI sales automation fails

• Errors compound. A workflow with five AI steps that are each right 95% of the time finishes correctly only 0.95 to the fifth power, about 77% of the time. Check the end result, not each step in isolation.
• Silent CRM damage. A wrong close date or deal amount written by AI doesn't announce itself. It shows up later as a bad forecast. Keep forecast fields out of reach of automation, and see the guide to AI sales forecasting for why.
• Hidden instructions in inbound text. An AI that reads email or web pages can be steered by text written for it. The OWASP Top 10 for LLM applications (2025) lists prompt injection as the first risk and recommends least-privilege access and human approval for high-risk actions.
• Sending damage. Automating outreach multiplies volume. Google's sender guidelines tell senders to keep the spam rate in Postmaster Tools below 0.1% and never reach 0.3%, and require SPF, DKIM, DMARC and one-click unsubscribe for senders of more than 5,000 messages a day to personal Gmail accounts.
• Opt-outs lost between tools. The FTC's CAN-SPAM guidance says the law "makes no exception for business-to-business email," that opt-outs must be honored within 10 business days, and that "you can't contract away your legal responsibility." If three tools send email, all three need one suppression list.
• Loops and runaway costs. Two automations replying to each other's auto-replies can burn through a month's credits overnight. Set a cost cap per workflow.

Worked example: a cybersecurity vendor automates meeting follow-up

Illustrative example. The company and numbers are invented to show the method.

The team: Saltmarsh Cyber (saltmarsh.example) sells security awareness training and phishing simulation to companies with 100 to 1,000 employees. Its 6 account executives hold about 12 customer meetings a week each.

The baseline: for one week, each AE logs time spent after every meeting on notes, CRM fields and the follow-up email. The median is 20 minutes. That's 72 meetings × 20 minutes = 24 hours a week across the team.

The automation: after each meeting, an AI drafts the summary, proposes CRM field updates and writes the follow-up email. The AE reviews and sends. Close date and amount are on the spec's "never" list.

Two weeks of results:

At the BLS average employer cost of $54 an hour for a full-time private-industry worker (June 2026), 16.8 hours is about $907 a week, or about $43,500 over 48 working weeks. The usage cost, at an illustrative 4 actions per meeting and Salesforce's published rate of 20 Flex Credits (about $0.10) per agent action, is 72 × 4 × $0.10 = $28.80 a week, about $1,380 a year.

The six errors in 40 (15%) were the real finding. Saltmarsh changed the spec so the AI marks a next step "proposed" unless the transcript contains the buyer agreeing to it, and kept the AE's review step permanently.

A two-week test you can run

1. Pick one workflow with high volume and an output you can check, such as meeting follow-up or reply sorting.
2. Time it for a week before automating. Ask reps to log minutes per run. Without a baseline, every saving is a guess.
3. Write the spec, including the "never" fields and the fallback.
4. Run it for two weeks with approval on. Every output gets a human look.
5. Sample 20 outputs a week and count errors by type.
6. Calculate net time saved: old minutes minus review minutes, times runs.
7. Decide: keep it if net savings are real and errors are caught before they reach a customer or a forecast. Remove approval only for outputs that never reach customers and had no errors in the sample.

What it costs

Published prices checked on 26 October 2026:

Per-action pricing is cheap per run and unpredictable per month. Estimate runs per week times actions per run before you sign, then set a cap.

What to ask a vendor

• Which steps are rules and which are AI? Can I see each run's inputs, outputs and log?
• Can I lock fields and actions the automation must never touch?
• How does it handle low-confidence results: guess, skip or hand to a person?
• How are opt-outs synced across every tool that sends email?
• What protects it from instructions hidden in inbound email or web pages?
• What exactly is billed per run, and can I cap monthly spend?

Self-check before you switch an automation on

• It has a written spec with an owner, a fallback and a cost cap.
• Nothing it produces reaches a customer without approval, unless it passed a clean sample.
• It can't change forecast, pricing or contract fields.
• Opt-outs from every channel land in one list.
• You have a baseline, so you'll know whether it saved time.

Next step: if the outbound sequence is the workflow you want to automate, read the guide to AI SDRs and its six-week pilot plan.