AI can help a sales team prepare for conversations, organise enquiries and draft follow-ups. Its value depends on whether those outputs lead to better work, not on how many messages the system generates.
A claim that AI increases conversion by a fixed percentage is not a substitute for an evaluation in your own sales process. Lead quality, deal size, sales-cycle length and changes in the market all influence results.
Begin with the sales process and CRM records
Define the stages and the conditions for moving between them. Distinguish a new contact, a qualified opportunity, a proposal and a signed agreement. If teams use those labels inconsistently, a model will learn or reproduce that inconsistency.
Record the source of each lead, the timing of activity and the eventual outcome. Decide which fields are reliable enough to use. Missing CRM updates often require a better workflow before they require predictive modelling.
Lead qualification: recommendations need reasons
A model can suggest a priority or identify missing information. It should not silently remove opportunities from a salesperson’s view. Provide a reason, access to relevant evidence and a route to correct a recommendation.
Historical won deals do not provide an unbiased picture of every potential customer. Salespeople may previously have pursued only certain industries or account sizes. A model trained on those outcomes can reproduce that selection rather than identify underlying potential.
Test against a simple rule-based baseline. Compare performance across relevant segments and review opportunities that were ranked low but later converted. Avoid using information that becomes available only after the outcome you are trying to predict.
Drafting and research: verify before sending
Useful applications include a briefing from approved account records, a draft response to a stated requirement or a reminder based on a documented conversation. Keep facts distinguishable from suggestions.
Check names, dates, prices, contractual promises and references before a message reaches a prospect. Do not invent personal details or imply that a human has researched something they have not reviewed. Respect the permissions and communication preferences recorded for the contact.
| Task | Appropriate control |
|---|---|
| Meeting preparation | Links to source records and a check of recent changes |
| Draft follow-up | Human approval of commitments and factual statements |
| Lead routing | Visible rule or reason, with reassignment available |
| Quote preparation | Approved product and pricing data; separate commercial sign-off |
Forecasting: evaluate on future periods
A sales forecast should be compared with a simple historical or pipeline-based method. Use data that was actually available at the forecast date. Including the final close date or later CRM changes in a training example can create misleadingly good results.
Track forecast error by horizon and segment, not just one aggregate score. Explain major assumptions and review unexpected changes. A point forecast is not a guarantee; a displayed range is meaningful only if its uncertainty has been evaluated.
Measure incremental value
Define the intended outcome before deployment: better qualified meetings, shorter preparation time, fewer incomplete handovers or more reliable forecasts. For outreach, consider a suitable comparison group and the effect of seasonality, staffing and campaign changes.
Monitor rejection, complaints, opt-outs and editing effort alongside positive outcomes. More activity can increase cost or damage relationships if the messages are irrelevant.
For revenue-related calculations, use a defensible contribution measure rather than treating all additional revenue as profit. Do not count both the same employee time saving and an overlapping staffing reduction.
An illustrative business case
Assume an implementation costs PLN 80,000. Suppose the business estimates PLN 140,000 of annual benefits and PLN 20,000 of annual operating costs. These are hypothetical inputs, not QA10 results or a quotation.
The resulting annual operating benefit is PLN 120,000. Simplified first-year ROI is (120,000 − 80,000) ÷ 80,000 × 100 = 50%. Simple payback is 80,000 ÷ (120,000 ÷ 12) = 8 months.
That payback assumes benefits start immediately and remain constant. A real model must account for ramp-up, uncertainty, support, tax and the timing of cash flows. Validate the assumptions with observed results before expanding the project.
Questions before using AI in sales
Can AI send proposals without review?
Commercial commitments need an authorised approval process. A system can prepare a draft from approved data, but permissions and checks should determine who may approve or send it.
What should we measure first?
Choose a measure tied to the actual task, such as accepted qualification decisions, preparation time or forecast error. Record a baseline and include correction effort and unwanted outcomes.
Choose a starting point
If the work starts with sales processes and organisational change, Q-Journey provides that broader route. A defined agent or CRM-connected application can be discussed through Q-Dev Agentic. Neither route assumes a fixed conversion uplift.