> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.pivotal.app/llms.txt.
> For full documentation content, see https://docs.pivotal.app/llms-full.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.pivotal.app/_mcp/server.

# Limitations

Pi is a drafting and summarizing assistant. It's good at the work humans hate (reading a long thread, comparing 30 customers, writing a first draft) and weaker at work that needs precision (counting, citing a number, anything outside Pivotal's data). The honest list is below so you know when to trust the answer.

## What Pi does vs what Pi won't

| Pi does                                                            | Pi won't                                                      |
| ------------------------------------------------------------------ | ------------------------------------------------------------- |
| Read customers, contacts, onboardings, tasks, comments you can see | Send an email, post a comment, or mark a task done on its own |
| Draft a comment, email, or summary in your voice                   | Move an onboarding to a new phase                             |
| Forecast a launch date from task pace and phase velocity           | Change a customer's owner, status, or tags                    |
| Rank customers by risk signals                                     | Delete a record, even if you ask                              |
| Explain why it flagged a customer at-risk                          | Read data from another workspace                              |
| Quote a comment or email back to you with the source chip          | Read connected systems Pivotal hasn't synced                  |

Every action lands in your composer as a draft. You confirm. Pi does not have a write path to your records.

## Where Pi gets things wrong

### Counting

Pi reads tables but is not a calculator. Numeric claims ("Acme has 17 open tasks") are usually right and sometimes off by one or two. Treat numbers as a starting point and click through to the source view if you're going to act on the figure.

### Long-tail customer history

Pi reads up to the most recent 500 events per customer. Customers older than 18 months can have history Pi summarizes incompletely. Ask for sources and you'll see which window Pi used. For full history, open the customer and use the [Customer history](/product/customers/customer-history) tab.

### Anything outside Pivotal

Pi only reads Pivotal's database. Slack threads Pivotal hasn't synced, HubSpot fields you haven't mapped, Stripe invoices you haven't connected. Pi can't see them. If a prompt depends on data from a connected system, check [Field mapping](/product/integrations/field-mapping) and confirm the sync is current.

### Fresh records

Records created in the last few minutes can lag the data Pi reads from. If you just added a customer and Pi says "I can't find Acme", wait two minutes and ask again. The indexer catches up on a short interval.

### Forecasts in the first 30 days

Pi's forecast model needs at least four closed tasks to reach high confidence. New onboardings show **Confidence: Low** on the chip. Read the slippage as directional, not precise, until confidence climbs.

## What to do when Pi is wrong

1. Ask for sources. Pi attaches a chip listing the records it read. If a record is missing, the answer was built from less context than you'd want.
2. Open the source. Confirm against the underlying customer or onboarding.
3. Tell Pi. Reply in the same thread with "this is wrong because..." and Pi rewrites. Pi doesn't learn workspace-wide from a single correction, but it follows the thread.
4. Report it. If Pi consistently misreads a kind of question, email `help@pivotal.app` with the prompt and what you expected.

## Related

* [What Pi can do](/product/ask-pi/what-pi-can-do)
* [Privacy](/product/ask-pi/privacy)
* [Forecast launches](/product/ask-pi/forecast-launches)

Email **[help@pivotal.app](mailto:help@pivotal.app)** with a screenshot of where you got stuck and the customer or onboarding id from the URL.