AI Implementation: A Practical First 90 Days
Successful AI implementation starts with one business process, not a company-wide AI strategy. Pick a task that happens often and costs real time, measure it, improve it with AI and then decide what to do next based on evidence.
Here is a 90-day plan you can adapt. The timings are a guide, and larger organizations with more approvals may need longer.
Days 1 to 15: Choose one process
List the tasks that take up your team's time. Then choose one that is:
- Frequent. It happens often enough for savings to matter.
- Measurable. You can count time, errors or cost.
- Safe to test. A mistake can be caught before it reaches a customer.
- Owned. One person in the business can make decisions about it.
Reading incoming requests and drafting replies is a common type of starting point. Avoid your most sensitive process for the first project.
Name a business owner and agree on how the work will be reviewed. Check any data rules at this stage, including what can be sent to an outside model provider.
Days 16 to 30: Measure and collect
Record how the process performs today: time per task, error rate, cost and volume. Without this baseline, you won't be able to prove anything later.
Collect 50 to 100 real examples, including difficult ones, along with what a good result looks like for each. These become your test set.
Before building, try the best available ready-made tool on some of those examples. If it does the job, you may be close to finished.
Days 31 to 60: Build and test
Build the narrowest version that could work. Keep the AI in a drafting role, with no ability to change records or contact customers.
Score it against your test set. Look closely at the failures, since they tell you more than the average score does. Fix the common ones through better data, instructions or design, then run the full set again.
Set a pass mark before you start. Otherwise, it's tempting to move the target once you see the results.
Days 61 to 90: Run it with real work
Give it to a small group for real work, with a person checking every output. Track:
- Time per task, against your baseline
- How often people change the AI's work, and how much
- Errors caught in review
- Cost per completed task
- What the team says about using it
Listen to the people doing the work. If they're quietly going back to the old method, the numbers won't show why.
Day 90: Make a decision
You should now be able to choose one of these with evidence:
- Expand. The results beat the baseline, so roll it out further and reduce review where it's safe.
- Improve. The results are promising, but a specific problem needs fixing first.
- Stop. The value isn't there. This is a useful result, and it costs far less to learn now than after a large rollout.
Common reasons this goes wrong
- Starting with the technology rather than a problem
- Skipping the baseline
- Testing only on clean, easy examples
- Having no business owner
- Announcing a broad transformation before anything works
- Leaving out the staff who do the work
If you aren't sure you're ready, begin with an AI readiness assessment.
Talk to us
DPR Software Labs can help you choose your first process and get it working in production. Bring a process you'd like to improve.