Analysis Paths

An analysis path is a friendly name for a table you analyze often — like “Operations Results” or “the P&L”. Register it once, and from then on you (and your connected AI assistant) can ask questions about it by that name, without repeating which project and which table you mean every single time.

Pick one as the workspace default, and plain questions like “why did this go up last quarter?” just work — the assistant already knows what “this” refers to.

Note

Analysis paths are a shared workspace convenience: an administrator sets them up, and then anyone on the workspace can use them in their questions — through Microsoft 365 Copilot or a coding agent connected over MCP. Every question still runs under each person’s own permissions.

Why Use Them

Without an analysis path, every question has to spell out exactly where the data lives — the project and the table — which gets tedious fast and is easy to get wrong. With one, you talk about your data the way you actually think about it:

• Ask by name. “Summarize the Operations Results.” instead of “summarize table ops_summary_v3 in project Enterprise Profitability.”
• Set a default. Mark your most-used table as the default and drop the name entirely — “what changed since last month?” lands on the right data automatically.
• Keep everyone consistent. When the whole team refers to “the P&L,” they all reach the same table — no guessing, no stale copies.

Register an Analysis Path

If you’re a workspace administrator, you don’t need a form or a settings screen — just ask your connected AI assistant, in plain language, to register the path for you:

“Register an analysis path named Operations Results that points at the ops_summary table in the Enterprise Profitability project.”

The assistant confirms the path and it’s ready to use immediately. You can register as many as you like — one per table you analyze regularly.

Tip

Choose names the way your team talks: “Operations Results,” “the Forecast,” “Headcount.” The friendlier the name, the more natural your questions become.

Set a Default Path

As an administrator, mark one path as the default so questions that don’t name a table land on it automatically:

“Make Operations Results the default analysis path.”

From then on, “what were the top movers this quarter?” or “why did the total drop?” run against that table with nothing else to specify. There’s a single default for the whole workspace, so everyone’s untargeted questions resolve to the same table. Setting a new default simply moves the crown — there’s only ever one, and the assistant tells you which it is.

Making a Path Useful to AI Assistants

A name and a table tell an assistant where the data is. They don’t tell it what the data means: which filters always apply, how your fiscal year works, what your team’s words refer to. Without that, an assistant guesses, and a confident wrong answer is worse than a slow one. This section covers the two places you put that knowledge, and why both matter.

In this guide, AI assistants means any LLM connected through PlaidCloud MCP (Claude, Microsoft Copilot, ChatGPT and others) plus PlaidCloud’s built-in AI.

Prompt Guidance decides how good the answers are

A path with good Prompt Guidance and column memos on its table gets correct, consistent answers from every assistant. A path with neither gets whatever the assistant infers from column names. Ten minutes spent on these two things is the highest-value setup you can do for a path.

Add Prompt Guidance

Each analysis path has Prompt Guidance: free text, written by you, that tells assistants how to use the path. Write one directive per line.

To add it, open the Tools menu and choose Analysis Paths, then select a path and choose Edit (or choose New to create one). The Prompt Guidance field follows Table, just before Set as workspace default. An assistant connected with permission to manage analysis paths can also set it. MCP and API callers see this field as description, for example: “Set the description of the Operations Results path to …”

Assistants use the Prompt Guidance in two ways:

1. To choose the path. When a question doesn’t name a path, the assistant compares it with each path’s name and Prompt Guidance and picks the best match.
2. As standing guidance. Once it has chosen a path, the assistant applies the Prompt Guidance to every query against that path: the filters to include, the terminology to translate. For an assistant connected over MCP, the business numbers themselves come from the project’s metric definitions, not from the guidance.

The Prompt Guidance is stored with the path and is visible to everyone in the workspace who can see analysis paths. It reaches assistants that work through MCP and PlaidCloud’s built-in AI. It is not passed to the AI assistance inside a workflow’s LLM step.

What to Write

Include what a new analyst would need to be told on their first day with this table. Give each directive its own line.

Write this	Example
Questions the path answers, and what it doesn’t.	“Monthly operating results by business unit. Use for revenue, cost and EBITDA questions. Not for headcount or forecasts; use the Forecast path.”
Which metric a business word means, for assistants connected over MCP.	“‘Revenue’ is the Net Revenue metric. billed_amount is invoicing, not revenue.”
Filters that always apply.	“Always exclude scenario = 'Budget' unless asked about budget. Intercompany rows have entity_type = 'IC'; leave them out of totals.”
Sign conventions, calendar and grain.	“Costs are stored as negatives. Fiscal year starts in July, so FY26 is Jul 2025 to Jun 2026. One row per business unit per month.”
Business terms mapped to data.	“‘The Midwest’ means region IN ('IL','IN','OH','MI'). ‘Core products’ means product_tier = 'A'.”

Keep it short and specific. One directive per line beats a paragraph, and the limit is 1,000 characters. Leading and trailing spaces are trimmed.

Note

For an assistant connected over MCP, how a business number is calculated — which column, added up how, over which rows — belongs in a metric definition, where a second Architect approves it, not in Prompt Guidance. Guidance says which filters and terms apply to the path; a metric says what the number is.

What Not to Write

• What assistants already see. Column types, keys and row counts are returned with the table. Repeating them uses up your 1,000 characters.
• What one column means. Put that in a column memo so it stays with the column, wherever the table is used. Prompt Guidance is for how to use the path as a whole.
• Anything confidential. Prompt Guidance is visible across the workspace, not just to people with access to the table’s project. Don’t include passwords or keys, personal data, or figures and commentary you wouldn’t show every workspace member.

Good and Weak Prompt Guidance

Both examples below are for the same path, and the same question is asked: “What was EBITDA for the Midwest last year?”

	Weak guidance	Good guidance
Text	“Operations data.”	“Monthly operating results by business unit. EBITDA = ebitda_adj; ignore ebitda. Exclude scenario = 'Budget' and intercompany rows (entity_type = 'IC'). Fiscal year starts in July. Midwest = IL, IN, OH, MI.”
What the assistant does	Picks the plain ebitda column, includes budget and intercompany rows, and reads “last year” as the calendar year.	Uses ebitda_adj, leaves out budget and intercompany rows, uses the last completed fiscal year, and filters to the four Midwest states.
The answer	A number that is plausible, wrong, and different from the finance team’s.	The number the finance team reports.

The weak guidance isn’t false. It just leaves every decision to the assistant’s guess.

Add Column Memos

Prompt Guidance covers the path as a whole. A column memo explains one column, and it is how assistants learn what a column actually holds. Where a column’s purpose isn’t clear from its name, a memo is the difference between the right column and a plausible wrong one.

Memos on the path’s table are returned to AI assistants with the column list, in table_describe with include=['columns'], as memo on each column that has one. Columns without a memo are returned as before. The same memos reach the AI assistance in a workflow’s LLM step for the tables it uses.

When a Memo Is Needed

Add a memo for	Example memo
Codes and abbreviations	seg_cd: “Customer segment code: E = Enterprise, M = Mid-market, S = SMB.”
Look-alike measures	ebitda_adj: “EBITDA after one-time items are removed. Use this for reporting; ebitda is unadjusted.”
Flags	is_ic: “1 if the row is an intercompany transaction. Exclude from consolidated totals.”
Units or currency	amt_local: “Amount in the entity’s local currency, in thousands.”
Derived or allocated values	alloc_cost: “Shared cost allocated to this unit by headcount. Not directly incurred.”

Skip columns whose names say it all, such as invoice_date or customer_name. A memo that repeats the name adds nothing.

How to Add a Memo

1. Open the path’s table in Table Explorer.
2. On the toolbar, turn on Column Properties.
3. Select the column in the column list.
4. Enter the text in Memo.

A memo set on the table applies wherever the table is used, including every analysis path that points at it.

Prompt Guidance or Memo

Put each fact in one place.

	Prompt Guidance	Column memo
Covers	How to use the path as a whole	What one column means
Typical content	Questions it answers, standing filters, fiscal calendar, business terms	Codes, units, flags, which look-alike to use
Example	“Exclude intercompany rows from totals.”	is_ic: “1 if the row is intercompany.”

Here the memo says what is_ic means, and the Prompt Guidance says what to do with it.

Note

If your workspace uses Microsoft 365 Copilot: once your workspace is on this release, reinstall the updated Copilot package to pick up the new instructions. See Microsoft 365 Copilot.

Using Them

Once your paths are registered, just ask:

1. By name — “Show me the first rows of the Operations Results.”

2. By the default — “What changed versus last quarter?” (no table named — it uses the workspace default path).

3. Switch freely — “Now do the same for the Forecast.” The assistant follows along, path to path.

This pairs naturally with allocation questions — see Tracing Allocations — so you can go from “why did the Operations Results move?” straight into the accounts and drivers behind it.

Using Them Through MCP

When an assistant connected over MCP passes an analysis path to a tool, it can use either the path’s name (not case-sensitive) or its id. A path updated with an id that doesn’t exist is refused; to create a new path, leave the id out.

query_data_read accepts analysis_path in place of a project and table for table, json_paths, dimension_value and column_values.

Preview only some columns. Pass columns=[...] to query_data_read(action='table') to return just the columns you name. Row access rules still apply in full. It can’t be combined with version_id.

Get a filtered total in one call. Ask column_values for the measure and filter on the column you care about:

query_data_read(
  action='column_values',
  project_id=<project>,
  table_id=<table>,
  dimension_column_name=<filter column>,
  value_column_name=<measure>,
  filters={<filter column>: <value>}
)

It returns one row for that value. When a column name is misspelled, the error lists up to five close matches so the assistant can correct it and retry.

Who Can Set Them Up

Setting up analysis paths — registering them and choosing the default — is an administrator action, so the shared names stay clean and trustworthy for everyone. Once they exist, every member of the workspace can use them in their questions, each still seeing only the data their own permissions allow.

Related

• Metric Definitions — agree once how a business number is calculated
• Answers You Can Trust — how PlaidCloud grades and caveats every AI answer
• Tracing Allocations — explain why a named result changed
• Retrieval, Grounding, and Analysis Paths — where analysis paths sit in the wider retrieval architecture
• Getting Started with AI Coding Agents — connect an assistant over MCP
• Microsoft 365 Copilot — use analysis paths from Teams and Outlook