AI rollout playbook: month-by-month for a 10-50 person team

Every team we talk to has the same instinct: "let's try AI across customer support, sales, and finance at the same time." Three months later they have three half-built AI deployments, none in production. Below is the playbook that works — a focused four-month rhythm starting with exactly one workflow.

Month 1 — Pick one workflow + design the AI

Week 1 — Survey the team

• Ask every team lead: "What repetitive task eats five-plus hours of your week?"
• Compile a list of 15-20 candidate workflows.
• For each, note the team size, the hours-per-week, the rough error cost.

Week 2 — Pick one

Pick based on this scorecard. Score each candidate 1-5:

• Time eaten — high score = more hours per week
• Reversibility — high score = mistakes are easy to fix
• Boundedness — high score = task fits in a clear input/output shape
• Owner motivation — high score = team lead actively wants this
• Data availability — high score = past examples + clear templates exist

Total each. Pick the highest. If two tie, pick the one with the most motivated owner — that person will carry the rollout.

Week 3-4 — Design the AI

• Define the input shape (what does the AI see?)
• Define the output shape (what does it produce?)
• Define the human approval point (who signs off, on what fraction of outputs?)
• Define the success metric (one number — see our AI ROI post)
• Pick build vs buy: vertical product, custom build, or off-the-shelf tool. See our framework.
• Measure the baseline for the success metric — two weeks of pre-AI data.

Month 2 — Ship the MVP

Week 5-6 — Build (or configure)

Build the simplest thing that could work. Resist scope creep. Three rules:

• Single workflow. The MVP solves one task end-to-end. Not five tasks partially.
• Human approval gate. Every AI output goes through human approval in the MVP. Loosen this only after week 8 when quality is measured.
• Observability from day one. Every AI action logged with the 8 minimum fields (see our production AI properties).

Week 7 — Internal pilot

• One team member uses the AI on real work for one week.
• Collect feedback daily — what's wrong, what's surprising, what's slow.
• Fix the worst three issues. Ship a v0.2.

Week 8 — Soft launch to the team

• The full target team starts using the AI.
• Daily 15-minute standup for two weeks specifically about the AI.
• Track adoption rate — what % of the team is using it daily.

Month 3 — Measure + iterate

Week 9-12 — Run the experiment

Continue using the AI. Do not change anything major. The point is to see real production behaviour without confounders.

• Weekly: review the outcome metric vs baseline
• Weekly: review adoption rate
• Weekly: review the human approval queue — what % of AI outputs are rejected?
• Weekly: review failure modes — which inputs does the AI handle badly?

Three patches are allowed during this period — for clear quality problems, not feature requests. Resist scope creep.

Month 4 — Decide

Week 13 — Run the numbers

With 6 weeks of post-launch data, compute the ROI:

1. Outcome metric delta (post-launch average minus baseline)
2. Convert to value (hours × cost / units × revenue / errors × cost-per-error)
3. Subtract AI cost + internal time
4. Annualise

Week 14 — The three-option decision

Based on the ROI and adoption data, pick one of three paths:

• Expand — ROI is positive, adoption is high. Roll the AI out to adjacent teams, OR pick a second workflow with the same playbook.
• Tune — ROI is positive but adoption is low, OR adoption is high but ROI is borderline. Spend month 5 on adoption + quality. Re-measure.
• Kill — ROI is negative after honest attempt, OR the team genuinely does not want it. Shut it down. Free the team. Try a different workflow.

Killing is the underrated option. A failed pilot that gets shut down cleanly is better than a half-alive deployment that drags on for a year.

Common failure modes

"Let's start with three workflows at once"

The most common failure mode. Three half-built deployments are worse than one production-quality one. Pick one.

"We will figure out approval later"

Translation: nobody owns the AI's output quality. Define the human approval point in week 4. No exceptions for the MVP.

"We will measure when we have time"

Translation: we will never measure. Set the baseline in week 4. Set up the dashboard in week 5. Both are non-negotiable.

"The AI works; the team is just resistant"

Maybe. More likely: the AI does not fit the actual workflow as well as the demo suggested. Spend a day shadowing the team. Watch where it breaks.

"Let's launch big"

Internal pilot in week 7. Team soft launch in week 8. Don't go company-wide until at least week 12.

What this looks like at a 25-person company

A 25-person services company we have talked with has this exact shape:

• Month 1 — surveyed 4 team leads. Picked client-update messaging (used to take 4 hours/week per project manager × 6 PMs = 24 hours/week).
• Month 2 — shipped an AI drafter integrated with their project tool. Internal pilot in week 7. Team launch in week 8.
• Month 3 — measured: hours dropped from 24 to ~8. Adoption ~85%. Approval rate ~75% (PMs edit before sending).
• Month 4 — decision: expand to a second workflow (proposal drafts). Same playbook, same owner.

This is the rhythm. It is not flashy. It works.

What this means for you

• Block 4 months on the calendar. Resist the urge to do three workflows at once.
• The hardest week is week 2 — picking. Use the 5-criterion scorecard, score honestly, pick the highest.
• Week 4 is when most teams fail to set up measurement. Do it.
• Read the readiness checklist before week 1 to confirm you are ready.
• Read the ROI post before week 4 to pick your metric.

Want a second pair of eyes on the workflow you are picking, or the rollout plan? Book a 30-minute call. We will walk through it with you.