AI in Sales: Proposals, Research and Follow-up

Sales is where AI produces the fastest visible return in most businesses, because the work it replaces is the work sellers least want to do. Research, proposal drafting, CRM updates and follow-up sequences consume the majority of a sales week, and none of it is selling.

This guide covers the four sales workflows where AI pays off, the checks each one requires, and the commercial guardrails that keep a faster pipeline from becoming a litigated one. It is part of the Use Cases & Workflows pillar.

AI can make the pipeline faster. Only a human can make it honest — so commercial terms and claims about the client stay with a person.

Why sales qualifies

The time allocation is the argument. Salesforce’s State of Sales Report 2026 (n=4,050) found that around 60% of a sales rep’s week is non-selling activity, with only about 40% spent selling. Proposal drafting, account research and CRM hygiene sit squarely in that 60%.

Applying the four tests:

• Frequent. Proposals, research and follow-up happen weekly.
• Expensive. Seller time is among the most expensive time a business has.
• Verifiable. Research claims can be checked; proposals can be reviewed against capability and terms.
• Contained. One document type, one team, one set of sources.

The containment caveat matters: proposals touch pricing and delivery commitments, so the workflow must place a human gate precisely there.

Workflow 1: Account and prospect research

What AI does: produces a structured briefing on a named company from supplied sources — public information, the CRM record, prior correspondence, meeting notes.

What the human does: opens every source, confirms the claims that will be used in outreach, and confirms the account’s relationship history.

The check: no claim about a prospect is used in outbound until a human has confirmed it. A wrong detail about a client’s business is the fastest way to lose a meeting.

The prompt pattern: a fixed briefing structure — company, recent developments, likely priorities, relevant prior contact, suggested angle. Structure is what makes the output comparable between accounts.

Workflow 2: Proposal and bid drafting

What AI does: drafts standard sections against a template and an approved content library — company background, methodology, team, approach.

What the human does: writes or approves the commercial sections, confirms every claim of capability, and signs off the final document.

The check: pricing, terms, delivery commitments and any statement of what the business can do are human-set and human-approved, without exception.

Why it matters: proposal production is a recurring, deadline-driven document workload with a measurable outcome — win rate — which makes it one of the few AI use cases with a directly attributable commercial metric. For the full methodology, see the RFP Win Desk.

Workflow 3: Follow-up sequences

What AI does: drafts follow-up emails from the meeting record, the proposal and the agreed next step.

What the human does: confirms the commitments made, the dates, and the tone appropriate to the relationship.

The check: any date, price or commitment in a follow-up is verified against what was actually agreed. A follow-up that promises something the business cannot deliver is a self-inflicted problem.

What to avoid: automated sequences that run without a human reading them. The failure is not the automation; it is the unread commitment.

Workflow 4: CRM hygiene

What AI does: turns meeting notes and email threads into structured CRM entries — next steps, dates, contacts, stage changes.

What the human does: confirms the stage and the forecast, because the forecast is a judgement rather than a transcription.

The check: anything that affects forecasting is human-set. The data quality problem in CRM is well documented — only 35% of teams fully trust their CRM data, and 47% say accuracy has worsened (Salesforce, State of Sales 2026) — and automation can improve it or accelerate the drift, depending on whether a human confirms the consequential fields.

The commercial guardrails

Four rules keep the workflows safe.

• Terms are human. Price, discount, delivery date and scope are set by a person, never drafted as a default.
• Capability claims are verified. Nothing about what the business can do goes into a proposal without confirmation from the person who will deliver it.
• Client references are checked. Any figure, metric or outcome attributed to a client is confirmed, and used only if permitted.
• Outbound is read. Every outbound message is read by a human before it sends, without exception.

These four rules cost minutes per proposal and prevent the category of error that damages relationships rather than just documents.

A worked sequence for a proposal

A services firm responds to twelve proposals a quarter. The current cycle is six hours per proposal, of which two are research and two are drafting standard sections.

• Day 1 — research (45 minutes). The seller supplies the account brief and the CRM record; AI produces a structured briefing; the seller opens every source and confirms the claims that will be used.
• Day 2 — drafting (60 minutes). AI drafts the standard sections against the approved content library; the seller edits, and writes the commercial sections without AI.
• Day 3 — review (45 minutes). A second person checks the capability claims against delivery, the terms against the pricing sheet, and the client references for permission.
• Day 4 — submission. The seller reads the final document end to end and submits.

Net change: roughly three hours saved per proposal, with a second-person check added at the point where errors are most expensive. Across twelve proposals a quarter, that is around thirty-six hours returned to selling, against a verification cost of nine hours — a defensible ratio, and one you can check against your own numbers.

Measuring it

Sales has an advantage: the outcome is measurable.

• Seller time on selling, which should rise from around 40% of the week.
• Proposal cycle time, against a baseline, including review.
• Win rate over a sufficient period — not one quarter.
• Research briefs produced, and the error rate found at the check.

Win rate is the honest measure and the slowest to move. Cycle time and seller time are faster signals that the workflow is working.

Common mistakes

• Sending AI-drafted outbound unread. The most common and most damaging error.
• Letting AI set or imply commercial terms. Price and commitments are human decisions.
• Unverified claims about the client. Research is only as good as the sources that were opened.
• Automating follow-up without a human check. Committed dates and promises need a reviewer.
• Auto-updating the forecast. Stage and forecast are judgements.
• Measuring activity rather than outcome. Emails sent is not a result; win rate is.

Frequently asked questions

How do sales teams use AI?

For account research, proposal drafting against an approved library, follow-up drafting from the meeting record, and CRM hygiene — with a human verifying claims, terms and commitments in each case.

Can AI write proposals?

It can draft the standard sections. Commercial terms, capability claims and client references must be human-set and human-approved, and the final document needs a named approver.

Is it safe to let AI draft client emails?

Yes, provided every outbound message is read by a person before it sends. The risk is not the drafting; it is the unread commitment or the unverified claim about the client.

How much time can AI save a sales team?

The prize is the non-selling share of the week, which recent research puts at around 60% of a rep’s time (Salesforce, State of Sales 2026, n=4,050). Measure your own baseline: cycle time for a proposal, and hours spent on research per opportunity.

Should AI update the CRM automatically?

It can populate factual fields from notes. Stage and forecast should remain a human judgement, because they are predictions rather than transcriptions.

Does AI improve win rates?

It can improve the inputs — better research, faster responses, more consistent proposals — but win rate depends on many factors and moves slowly. Treat it as a lagging measure, and track cycle time and seller time in the meantime.

Should we disclose AI use in proposals?

Where it is material to the client, or a contract or rule requires it, yes. In practice the more useful disclosure is the control: that every figure, claim and commitment was verified by a named person before submission.

What about AI features in our CRM?

They are usually the best starting point, because the data and the workflow already live there. Apply the same rule: let it populate factual fields, and keep stage and forecast as human judgements — those are predictions, and predictions need an accountable author.

Should research briefs be shared with the client?

Only where the content has passed the check and the client is permitted to see the sources cited. A brief is an internal preparation document unless an agreement provides otherwise.

Next step

Pick one workflow — research briefs are usually the safest start — define the check, and run it for a month with the terms and claims rules applied. See Designing a Human-in-the-Loop Workflow or book an AI adoption call to build the workflow with your sales lead.

Sources

• Salesforce, State of Sales Report 2026 (n=4,050): around 60% of a sales rep’s week is non-selling activity; only 35% of teams fully trust their CRM data, and 47% say accuracy has worsened.

Figures are cited from their sources and dated. Where a source is a vendor survey, the sample size is stated where published.