Sales teams resist AI for rational reasons. Their compensation depends on a process that currently works well enough, and every new system historically arrives with more data entry attached. Any implementation plan that ignores that is a plan to be quietly ignored in return.
The framework below is the one we use on client engagements. It is deliberately slow at the start — most of the first month produces no visible AI at all — because the failures we are asked to rescue almost always come from skipping that part.
Days 1-30: understand and baseline
Workflow discovery
Sit with three or four people doing the job and document what actually happens, not what the process document says. Record every step from lead arrival to closed opportunity, who does it, how long it takes and where the information comes from. Expect to find two or three steps that exist only because a system cannot talk to another system.
Use-case prioritisation
Score each candidate use case out of five on four dimensions and multiply. Anything under 60 goes on the backlog regardless of how interesting it is.
| Dimension | Question | Score 5 means |
|---|---|---|
| Frequency | How often does this task occur? | Multiple times daily per person |
| Time cost | How long does it take each time? | Substantial and consistently measurable |
| Tolerance for error | What happens if the output is imperfect? | A human reviews it before it matters |
| Data readiness | Is the information needed already accessible and clean? | Available in a system we already control |
The tolerance dimension is the one teams under-weight. High-frequency, high-time tasks that go straight to a customer without review are the worst possible starting point, however attractive the time saving looks.
Baseline measurement
Before anything changes, record: median time to first response, preparation time per meeting, CRM completeness, follow-up rate within 24 hours, and stage-to-stage conversion for the last two complete cohorts. Four weeks of baseline is the minimum that survives scrutiny later.
Tool selection and data access
Prefer capability inside a system the team already uses. Where a new tool is genuinely needed, confirm three things before purchase: where data is processed, whether it writes back into your CRM, and whether you can export everything if you leave. Agree data access with whoever owns information governance now, not after the pilot.
Governance
One page, four questions: what data may be used, what must be reviewed before reaching a customer, who approves a new use case, and what happens when output is wrong. Publish it where the team works, not in a policy folder.
Days 31-60: pilot with a small group
Choosing the pilot group
Three to six people, and deliberately not only your enthusiasts. Include at least one experienced sceptic. Enthusiasts tell you the ceiling; sceptics tell you the floor, and the floor determines whether the rollout survives.
Replace a step, do not add one
The single most important design rule. If the AI-assisted brief is produced in addition to the existing preparation, you have added work. Something in the documented workflow must be switched off on day one of the pilot.
Training
Sixty to ninety minutes, role-specific, using live accounts rather than examples. Cover what good output looks like, what bad output looks like, what must never be sent unreviewed, and how to report a failure in one click. Follow up with a fifteen-minute check-in each week of the pilot.
Human approval points
Define these explicitly for each output type. A workable default for a first pilot: internal briefs and CRM updates go straight through; anything customer-facing is reviewed and edited by the person whose name is on it; anything contractual or commercial is human-authored.
Adoption monitoring
Track weekly active use, not licences issued. If usage drops in week three, the cause is almost always friction in the workflow rather than scepticism — ask the sceptic first, they will usually tell you exactly what is wrong.
Days 61-90: decide, scale or stop
| Criterion | Threshold to scale | If not met |
|---|---|---|
| Primary metric versus baseline | Improved and sustained for at least three consecutive weeks | Extend the pilot once, then stop |
| Adoption | Majority of the pilot group using it weekly without prompting | Fix workflow friction before expanding |
| Quality | No rise in corrections, clarifications or disputed scope | Tighten approval points |
| Administrative load | Not increased for the pilot group | Remove a step or stop |
| Support burden | Manageable without a dedicated person | Simplify before scaling |
Failure and rollback
Write the rollback procedure before the pilot starts, and make it genuinely cheap: keep the previous workflow documented and usable for the duration, keep exports of anything the new tool holds, and name the person who can call a stop without a meeting. Teams that cannot roll back easily tend to persist with things that are not working, because reversing looks like an admission.
How to avoid creating new administration
- One tool per job. If a capability exists in your CRM, use it there even if a standalone product is slightly better.
- No parallel record keeping. If notes live in two places, salespeople will maintain neither properly.
- Automate the capture, not the judgement. Activity logging is a good automation target; deal stage is not.
- Delete a field for every field you add. CRM forms grow without limit unless someone actively prunes them.
- Measure administrative time explicitly. If it rises, the implementation has failed regardless of what else improved.
Common mistakes
- Launching to the whole team at once, so there is no comparison group and no safe way to iterate.
- Skipping the baseline, which makes the day-90 decision a matter of opinion.
- Choosing a customer-facing use case first because it is the most visible.
- Making the pilot voluntary and then reporting adoption as a success metric.
- Letting a vendor run the pilot, which reliably produces a positive result and an unusable dataset.
Key takeaways
- Spend the first month on discovery, use-case scoring and baselines rather than tools.
- Replace a workflow step on day one of the pilot; never add one.
- Include a sceptic in the pilot group and treat their friction reports as the primary signal.
- Define approval points by output type before anything reaches a customer.
- Write the rollback plan first, and use the day-90 criteria to scale or stop without negotiation.