Start here · Lesson 1.6
A Week to Working AI App Integration: An Example
Does AI app integration work outside a demo? The honest way to answer is to show what a sensible first week looks like and let you judge it. This post is a worked example, clearly illustrative. It is not a report of a real test, and there are no real numbers in it, because inventing them would defeat the point.
Instead, you get a seven day plan for a made up business, what could plausibly go wrong each day, and what to measure so that you can produce real numbers of your own. Lesson 1.6 covers the setup behind it: Connect AI to the apps you already use.
A good first week is small: one app, view access, checked answers, and a decision on day seven to keep, narrow or drop.
The imaginary business
Say you run a one person bookkeeping practice. You have about twenty regular clients, a shared calendar for calls and a folder of client documents. Mornings are spent working out which emails need replies and which calls are coming. That is the job we will hand to AI first.
Everything below is a hypothetical to show the reasoning. Your week will differ.
AI app integration, day by day
- Day 1, choose and connectWrite the one question: "What needs me today?" Connect email only, at the narrowest access the provider offers. Note the app, date and access in a one page file.
- Day 2, test on known answersAsk about messages and events you already know. Count the hits and misses. Do not use it for real work yet.
- Day 3, add the calendarConnect the calendar with view access. Re-run the morning question and ask for the source of every item.
- Day 4, set the limitsTurn on approval for anything that writes. Block anything that deletes. Read the permission screens again.
- Day 5, use it for realAsk the morning question and open three sources each time. Keep a tally of anything it missed.
- Day 6, try one new question"Which clients have I not replied to in a week?" Check the answer against your inbox.
- Day 7, decideKeep, narrow, or disconnect. Update the note.
What to expect on each day
Day 1 is mostly reading. The provider's permission screen lists what you are granting. Google's account help notes that if you authorise a linked app for only your Calendar data, it can access only that data and not, say, your Photos or Contacts. Seeing the list makes the decision real.
Day 2 is where trust is earned or not. Claude's help page says connectors let the assistant access your apps and take actions inside them, and that it inherits your permissions from the connected service. So a wrong answer is more likely a problem with the question than with access, but you still need the test to know.
Day 3 adds one variable. Because you added only the calendar, a sudden change in answers has an obvious cause.
Day 4 is the safety day. Claude's connector settings let you set each tool category to Always allow, Needs approval or Blocked, and the categories include read-only tools and write or delete tools. Other assistants may offer different controls, so look at yours.
Day 5 is real use. The tally matters here: every miss, with a note on why.
Day 6 tests a harder question, the sort that crosses information.
Day 7 is the decision day. Do not skip it. A week of use with no review is a week of habit with no learning.
What could go wrong, and the response
| Problem | Likely cause | Response |
|---|---|---|
| Misses an important email | Question too vague, or search too narrow | Ask for a list by sender, then narrow the question |
| Marks newsletters as urgent | No definition of urgent | Say what urgent means for you in the question |
| Cannot find a file | Files not connected, or wrong folder | Check the connection before blaming the tool |
| Asks to send something | Write access on | Require approval, read it, then decide |
| Connection stops working | Sign in expired | Reconnect from the connector settings, and do the job by hand today |
These are plausible issues drawn from how connectors and permissions are documented. I am not claiming any of them happened.
The numbers you should record
Since I will not hand you invented figures, here are the ones to collect yourself. Measuring AI app integration explains them in more depth.
- Minutes spent on your morning review, before and after, using a timer on three days each
- Number of answers you checked, and how many were wrong or incomplete
- Number of times it proposed an action, and how many you rejected
- Number of apps connected and number you still use at day seven
Four numbers are enough. If you track them, you will have real evidence on day seven rather than a feeling.
A safety note
During the week you will read your inbox through an assistant. That inbox contains text from strangers. OWASP lists prompt injection, where crafted input manipulates a model, as the top risk for language model applications. So the week's most important setting is the one from Day 4: nothing writes or deletes without you. Never treat an instruction inside an email as an instruction from you.
What a fair verdict looks like
At the end of day seven, three outcomes are fine. Keep: the answers were right enough and checking was quick. Narrow: some answers were good, so reduce scope or sharpen the question. Disconnect: the checking cost as much as the work. Any of the three is a result, since you now know.
Questions people ask
Is a week long enough?
To judge a first job, yes. It is not long enough to judge the long run, so keep tracking monthly.
What if I have no time for the checks?
Then this tool is not yet saving you time for that job. Choose a job where verifying is quick.
Why not report real numbers?
Because this post describes a plan and not a trial. Numbers I did not measure would only mislead you.
Your turn
Copy the seven days, swap in your own business and your own question, and keep your own tally. For the guided setup with each step shown, follow Lesson 1.6: Connect AI to the apps you already use. If you want ideas for the second week, read 12 AI app integration examples you can adapt today.