A 14-day plan to put AI to work in one department

Only 22% of organizations have scaled AI across more than one business unit, according to a Gartner survey of 1,303 functional leaders published September 1. The same survey found that about 11% didn’t know what their own function had spent on AI in 2025.

The more useful number is further down. Gartner’s best performers reported positive returns on 81% of their AI initiatives, while the weakest group could not say what return they got on 29% of theirs, as CXO Today summarized. Gartner describes its high performers as tracking returns constantly and cutting initiatives that underperform. Neither habit needs a bigger budget.

Smaller companies run into the same wall. When the San Francisco Fed asked small businesses about AI, the barriers they named included the time and capacity to train staff and gaps in knowing how to implement it, according to its September 1 summary.

A fixed two-week plan answers both problems. It starts with a baseline and ends with a written decision, and two weeks is short enough that a busy team can protect the time. This is the plan I would run for one department.

By the numbers

22%

of organizations have scaled AI across more than one business unit

Days 1 to 4: pick the work and measure it

Day 1. Pick one department and one workflow. Choose a department where the work arrives as text and repeats every week: support inboxes, sales follow-up, invoice matching, client reporting. Then pick one workflow inside it. “Answer order-status emails” is a workflow. “Improve customer service” is a wish. Name one person who owns the outcome. That person is not IT. It is whoever runs the department.

Day 2. Time the task as it runs today. Have two or three people log every instance of the task for the full day: when it started, when it ended, what they had to look up. You want a real minutes-per-task number and a real count per week. This is your baseline. Without it, day 14 becomes an argument about feelings.

Day 3. Write the procedure down. Sit with the person who does the task best and write out every step, every exception and every place they check something. This document becomes the agent’s instructions. You may find that three people do the task three different ways. Settle on one.

Day 4. Collect 20 real examples and set the rules. Pull 20 recent cases with the correct outcome for each. Include a few ugly ones. Then decide what the agent may read, what it may change and what it may never touch. Write that down too. Customer payment data, HR records and anything that sends money out should start in the “never” column.

Days 5 to 9: build it, test it, run it in the shadows

Day 5. Build the first version with tools you already pay for. You may already pay for an AI assistant inside your email, help desk or office suite. Start there. Load the written procedure as the instructions and connect only the data you allowed on day 4.

Day 6. Run the 20 examples. Score each output as right, fixable or wrong. Count them. If fewer than half are right, look at the instructions before you blame the model. Go back to the procedure and find the step you left out.

Day 7. Fix and rerun. Tighten the instructions, add the missing exceptions, and run the same 20 again. Do not add new examples yet. You are checking that your fixes worked.

Day 8. Design the review step. Decide who looks at the agent’s output before it reaches a customer or a system, and how. The simplest version: the agent drafts, a named person approves, edits or rejects with a one-line reason. Those reasons become your next round of instruction fixes.

Day 9. Shadow mode. The agent works on live cases, but its output goes only to the reviewer. The team still does the work the old way. Compare the two before everyone leaves each evening. This is the cheapest way to find the cases your 20 examples missed.

TimelineThe 14 days at a glance

1. Days 1-2Pick one workflow, time it
2. Days 3-4Write the procedure, set rules
3. Days 5-7Build, test on 20 real cases
4. Days 8-9Add review, run in shadow mode
5. Days 10-12Train the team, go live
6. Days 13-14Measure, then decide

Days 10 to 12: train the team and go live

Day 10. Train the people who will use it, on their own work. Skip the slide deck. Each person runs the agent on five of their own cases from the past week and reviews the output with the department owner. They learn what it gets wrong in their corner of the job. They also see that the review step is real.

Day 11. Go live with review on everything. The agent handles every incoming case in the workflow. Every output passes a person before it goes out. Keep a simple log: case, agent output, approved or changed, reason if changed.

Day 12. Read the log. Look at every change and rejection from day 11. Group them and find the two or three patterns that account for most of them. Fix the instructions for those patterns that afternoon.

Days 13 and 14: measure and decide

Day 13. Time the new version. Use the same method as day 2. Count minutes per task including review time, because review is real work. Compare against the baseline. Also count the error rate from the log.

Day 14. Decide, in writing. Write one page: the workflow, baseline minutes, new minutes, weekly volume, hours saved per week, error rate and what still goes wrong. Then make one of three calls. Expand it to the rest of the team. Keep it running and fix the top problem. Or stop and pick a different workflow. All three are fine outcomes. Not deciding is the only bad one.

That one page is the thing Gartner’s weakest performers don’t have. It also becomes the template for the second department.

A few rules hold the whole two weeks together. Keep it to one workflow, even when someone on day 6 suggests adding a second because “it’s basically the same.” It never is, and the extra scope is how two weeks becomes two months. Keep the department owner in charge of the decision on day 14, with you in the room. And protect the calendar. If timing, testing or the go-live slips by a day, slip the end date by a day instead of skipping a step. The baseline and the log are what make the result believable to everyone who wasn’t in the room.

What to do this week

ChecklistBefore day 1

• One workflow in one sentence
• One owner from the department
• Two days blocked for timing
• Twenty past cases pulled
• Data the agent can't touch

1. Pick the department and write the workflow in one sentence.
2. Name the owner and block day 1 and day 2 on their calendar now.
3. Ask the owner to pull 20 recent cases with the right answer for each.
4. Check which AI tools you already pay for that touch that department’s work.
5. Put day 14 on the calendar as a decision meeting with you in the room.

Sources

• Gartner finds only 22% have successfully scaled AI (TechInformed, September 4, 2026)
• Only 22% of Enterprises Have Scaled AI Across Business Units: Gartner (CXO Today, September 1, 2026)
• AI Adoption Among Small Businesses: Qualitative Insights from the Small Business Credit Survey (Federal Reserve Bank of San Francisco, September 1, 2026)