How to build a CRM with AI: a step-by-step guide
Learn how to build a CRM with AI in six steps: define your data, prompt an AI app builder, test with real records and publish safely.
On this page
How to build a CRM with AI comes down to three things: write down what your team tracks, describe it to an AI app builder in plain language, then test and refine the result before anyone relies on it. This guide walks through that process for a small sales team, with example prompts you can copy and adjust. You don't need to write code, but you do need to decide what the data looks like and check what the tool produces.
What a CRM needs to do (and what you can skip)
A customer relationship manager (CRM) is a place to keep track of the people you sell to, the deals you are working on, and what happens next. Commercial CRMs can do much more, including email sync, marketing automation and reporting across dozens of fields. Most small teams use a small fraction of that.
The minimum set of records
For a focused sales CRM, you need four kinds of information:
- Contacts: name, company, email and phone.
- Deals: a name, a value, a stage and an owner.
- Notes: a running log or text field on each contact, so context isn't lost between calls.
- Follow-up dates: the next time someone should get in touch.
If you can track those four things reliably, you have a working CRM. Everything else is optional until you prove you need it.
Why a small team needs fewer features
Off-the-shelf tools are built for many teams at once, so they include fields and settings you will never use. A custom CRM can match your process exactly: if your pipeline has four stages, it shows four stages, not twelve. The trade-off is that you own the design, so you have to make the decisions a vendor would otherwise make for you.
Questions to answer before you build
Write short answers to these before you open any tool:
- Who uses the CRM? Sales reps, a founder, a client-services person?
- What does each person need to see when they open it?
- Which information do they update every day, and which do they update once a month?
- What does a deal move through, from first contact to signed or lost?
- Who should be able to see or edit each deal?
Answers to these questions become your data model and your prompt. Skipping them is the most common reason a first version doesn't fit the team.
How to build a CRM with AI in six steps
These steps apply to any AI app builder that generates an app from a description. The examples below use FromScratch, but the process is the same elsewhere.
Step 1: Write the data model in plain words
Describe each record type as a list of fields. For example:
- Contact: name, company, email, phone, notes.
- Deal: title, contact, value, stage (Lead, Proposal, Won, Lost), close date.
Keep the language plain. Don't worry about database tables or relationships yet; the builder handles that. What matters is that each field has a clear meaning and that you have chosen stage names your team already uses.
Step 2: Describe the app in one prompt
Write a single prompt that covers the screens and the core workflow. Describe what people will do, not how the software should work internally. A good first prompt says who the app is for, which records it holds, which pages it needs and how the pipeline behaves. The full example is in the next section.
Generate the first version. Don't add features yet. You want a baseline you can judge.
Step 3: Review the live preview
Open the live preview and click through it as a team member would. Check each screen against the process you wrote down in the first section:
- Does the contact list show the fields you need?
- Can you move a deal from one stage to another?
- Do the stage names match your team's language?
- Is anything missing that you expected to see on the first screen?
Write down every mismatch. A list of five specific problems is much easier to fix than a vague feeling that something is off.
Step 4: Refine with short, specific requests
Refinement is a conversation. Ask for one change at a time and check the result before asking for the next. Specific requests work better than broad ones:
- "Add a search bar to the contacts list that filters by name and company."
- "Show a task reminder list on the home screen with contacts whose follow-up date is today or earlier."
- "Rename the stage 'Proposal' to 'Quote sent'."
Each change should be small enough that you can see whether it worked. If a change breaks something, you can restore an earlier version rather than untangling it by hand.
Step 5: Test with realistic records and a second user
Before anyone depends on the CRM, test it with data that looks like your real data:
- Add 30 to 50 contacts, including messy entries: missing phone numbers, company names spelled two ways, a deal with no value yet.
- Move deals through every stage, including Lost, and confirm the dashboard totals change correctly.
- Check dates. Set a follow-up for yesterday and confirm it appears in the reminder list.
- Ask a second person to log in and complete a task without your help. Watch where they hesitate.
This step catches most of the problems that look fine in a demo: duplicate contacts, date errors and confusing stage names.
Step 6: Publish and share the link
When the test passes, publish the app to a public URL or to your own domain. Share it only with the people who need it, and make sure login is required before anyone enters customer data (more on that below). Keep a short note of who has access so you can remove people when they leave the team.
Example prompts for a sales CRM
Good prompts are concrete. They name the records, the stages, the pages and the rules. Here are three examples you can adapt.
A full starter prompt
This is a complete first prompt for a small sales team:
Build a simple CRM for a small sales team. It should have a contacts list with name, company, email and phone, a deal pipeline with stages (Lead, Proposal, Won, Lost), and a notes field on each contact. Add a search bar, a dashboard showing deal value per stage, and require login so only team members can see the data.
Notice what it does: it names every record, lists the stages in order, specifies the dashboard, and sets an access rule. Anything you leave out, the builder will guess, so be explicit about the things that matter to you.
Follow-up prompts
After the first version, use short follow-ups to build out the workflow:
- "Add a task reminder list showing contacts with a follow-up date on or before today."
- "Let each deal have an owner, and show only the owner's deals when someone filters by 'My deals'."
- "Add a 'Last contacted' date to each contact and update it when a note is saved."
Keep a running list of changes you have made. If something stops working later, the list tells you where to look.
Importing existing contacts
If you already have a spreadsheet of customers, attach it so the app works with your real content rather than sample rows. You can attach files such as spreadsheets, PDFs and images. Then ask the builder to create contacts from the spreadsheet's columns, and check a sample of the imported records against the original. Don't assume every row landed correctly; spot-check the first and last ten.
Keep your customer data safe and correct
A CRM holds personal information about your customers, so treat it as sensitive from the first day.
- Require login. Any app that holds customer data should ask people to sign in before they see anything. Test this from a private browser window to confirm that the pages are actually protected.
- Check permissions. Decide who can see every deal and who can edit it. A salesperson may need their own pipeline but not a colleague's notes. Test each role.
- Keep a copy of your contact list outside the app. Export or save a spreadsheet snapshot on a schedule. This protects you against mistakes, not just against outages.
- Review the data regularly. Once a month, scan for duplicates, stale stages and missing owners.
- Limit what goes into prompts. Don't paste sensitive personal details into a prompt when a sample record would do. Use fake names and numbers for examples. For a broader look at prompting safely, read Vibe Coding Legal Risks: Prompt Safely.
Common mistakes when building a CRM with AI
Most problems with a first custom CRM come from a few predictable habits.
Describing too many features at once. A prompt with twenty requirements produces a muddled result that is hard to check. Build the core workflow first: contacts, deals, stages, notes. Add reminders, dashboards and imports afterward, one at a time.
Skipping the test with real records. Sample data hides the problems that matter: duplicate contacts, dates in the wrong format, deals with no value. Use messy data on purpose.
Changing the data model late. Once people have entered hundreds of contacts, renaming or restructuring fields is risky. Settle the fields and stage names in the first week, and make later changes deliberately.
Treating the first version as final. The first generated app is a draft. Review every screen, click every button and check every stage transition before you call it done.
Expecting the CRM to replace judgment. The app stores what your team enters and displays it. It won't decide which lead matters or write a good follow-up email for you.
Build it with FromScratch
FromScratch is an AI app builder. You describe the app in plain language, and it generates a React front-end, a Python (FastAPI) back-end and a database. Here is the process, step by step:
- Write your data model in plain words. List the contact fields, the deal fields and the stages before you open the builder. Use the answers from the first section.
- Start a new project and paste the starter prompt. Copy the prompt below, adjust the stage names or fields to match your team, and send it.
- Wait for the live preview. The first working version can appear within minutes. Don't judge it yet; click through each page first.
- Test against your checklist. Add messy sample contacts, move deals through every stage, and check the dashboard totals.
- Refine with one request at a time. Use short follow-ups such as the search bar or reminder list examples above. If a change goes wrong, restore an earlier version, since every change is saved.
- Require login and check roles. Ask for login, then test it from a private window with a second account.
- Publish to a public URL or your own custom domain once a second person has completed a task without help.
Here is the exact prompt to start with:
Build a simple CRM for a small sales team. It should have a contacts list with name, company, email and phone, a deal pipeline with stages (Lead, Proposal, Won, Lost), and a notes field on each contact. Add a search bar, a dashboard showing deal value per stage, and require login so only team members can see the data.
You can start on the free plan. Paid plans add credits and projects, so check pricing before you move a team onto the app. If you want to see how a dashboard-style screen is built in general, Build a dashboard with AI covers the same approach.
FAQ
Can a non-developer build a CRM with AI?
Yes, for a focused sales or client-tracking workflow. You need to define the data, describe the screens clearly and review every version before relying on it. Non-developers struggle most with testing, so schedule that step rather than skipping it.
How long does it take to build a CRM this way?
The first working preview can appear within minutes of your prompt. Testing with realistic records, refining the workflow and getting a second person to use it usually takes longer, and that is where most of the quality comes from.
Do I need to replace my existing CRM?
Not necessarily. Many teams start with a small tool for one process, such as tracking proposals, and keep their current system for everything else. You can move records over once you're confident the custom app fits your work.
Can AI build a CRM for my team if we have different roles?
It can build role-based access if you describe the roles and what each one should see. Test every role with a separate login, because permissions are the easiest part to get wrong. Review access regularly as people join or leave.