Data 360 costs for AI agents: catalogues and manuals

Data 360 costs for AI agents come mostly from documents that change, not from the agent asking. When an Agentforce agent reads spare part catalogues, manuals and service bulletins, Data 360 charges 150 Flex Credits per MB to index a document with the default parser and 600 per MB with LLM-based parsing. Searches cost 3 credits per million rows. In the worked example below, the agent’s searches cost about $72 a month, while re-indexing changed documents costs almost $1,000. The levers are the parser, the document scope and where the files are stored.

What does Data 360 charge when an agent reads documents?

An agent does not read a PDF directly. According to Trailhead1, Data 360 splits the document into chunks, turns each chunk into a vector embedding and stores it in a search index. At runtime a retriever fetches the matching chunks for the prompt. Salesforce’s billing notes for Agentforce Coworker2 put it in one line: Data 360 consumes credits “to process, semantically index, and query content”. How Data 360 assembles context beyond documents is covered in the Agent Context Engine explained.

Three meters follow from this:

• Processing per MB whenever a document is indexed or indexed again, split into Unstructured Processing and Intelligent Processing. According to the help article on unstructured data billing3, chunking and vectorising count only once, so a 100 MB PDF is billed as 100 MB.
• Queries per million rows each time the agent searches.
• Storage. Every ingested file counts toward Data 360 storage, and so do the chunk and index objects created for the search index.

150

credits per MB of Unstructured Processing, base tier

600

credits per MB of Intelligent Processing, base tier

3

credits per million rows queried, base tier

Rates fall with volume: above 300,000 credits a month they drop in three steps to a fifth, for unstructured processing 30 credits per MB above 12.5 million credits. Tiers apply per usage type and reset on the first day of each calendar month. Flex Credits cost $500 per 100,0004, and the same page notes for its Agentforce examples that Data 360 credits are not included. That is why these numbers rarely make it into the first estimate.

Default parser or LLM parsing: why do catalogues cost four times as much?

A spare part catalogue is mostly exploded drawings with position numbers and tables of part numbers. The default parser extracts text. For multimodal documents Trailhead5 describes an LLM-based parser, a Docling parser for nested tables and image processing for diagrams. The help article sets out the billing:

• LLM-based parsing: the entire document goes to the LLM, and its full size is billed as Intelligent Processing.
• LLM-based visual preprocessing: only documents that contain visual elements or tables are billed as Intelligent Processing, the rest as Unstructured Processing.
• Image processing: the entire document is billed as Intelligent Processing.

The list of Data 360 usage types6 also names Intelligent Context, the workspace for trying parsing and chunking settings, among the features billed as Intelligent Processing. For catalogues, “only documents with visual elements” means almost every document, so visual preprocessing saves little there. For operating manuals that are mostly text, the default parser is often enough. I would decide per document type and test retrieval quality on a sample, because 600 against 150 credits per MB is the biggest single lever in the estimate.

Worked example: after-sales documents at a machine builder

Assume a machine builder whose agent answers dealers and service reps on spare parts and maintenance. All volumes below are assumptions to show the method; the rates come from the rate card at list price.

• 6,000 MB of spare part catalogues as PDF, parsed with LLM-based parsing
• 3,000 MB of operating and maintenance manuals plus 200 MB of service bulletins, default parser
• about 300,000 chunks in a hybrid search index
• 4,000 conversations a month with two searches each
• 300 MB of changed catalogues and 100 MB of other changed documents per month

Item	Calculation	Credits
Initial load, catalogues	500 MB × 600, 2,500 × 480, 3,000 × 240	2,220,000
Initial load, manuals and bulletins	2,000 MB × 150, 1,200 × 120	444,000
Initial load in one month		2,664,000 ($13,320)
Monthly updates	300 MB × 600 + 100 MB × 150	195,000
Monthly searches	8,000 × 600,000 rows × 3 per million	14,400
Running cost per month		209,400 ($1,047)

For the searches I assume each query counts the whole index. The help article says that for hybrid search “the number of vectors and keyword records in the search index are counted”, and that a typical index has as many keyword records as vectors. Salesforce’s own pricing example7 assumes a query scans 20 percent of the unstructured data, so my figure errs on the safe side.

Three results stand out. Searches are 7 percent of the running cost. With the default parser for the catalogues too, the initial load would cost 1,164,000 credits instead of 2,664,000. And the first year comes to about 5 million credits, more than the 2.75 million Flex Credits per org per year that Salesforce lists for Max Edition, before a single agent action. The agent side of the bill is in what a dealer service agent costs.

The hidden Data 360 costs

Re-indexing whole records

Picture catalogues attached to product records that the SAP interface updates with prices or stock every night. If that reprocessed all catalogues just once a week, the month would carry 24,000 MB of Intelligent Processing: about 6.5 million credits or $32,700 instead of the planned 195,000. I would keep documents on objects or in a Data Library that no interface touches, and test this in the sandbox before go-live.

Tiers reset every month

Because tiers restart each month, timing matters. The initial load above costs 2,664,000 credits in one month. Spread over three months, the same documents cost about 3.54 million, $4,380 more. I plan the initial load as one batch in one calendar month.

Sandbox without tiers

In a sandbox, Data 360 has no volume tiers: 120 credits per MB unstructured, 480 intelligent. Indexing all 6,000 MB of catalogues there costs 2.88 million credits, more than the 2.22 million in production. Intelligent Context8 lets you test a configuration with up to five files, and a 150 MB sample in the sandbox costs 72,000 credits. That is enough to compare parsers.

Enriched indexing

Enriched indexes use an LLM to generate additional chunks, and Trailhead9 says an enriched index creates three new chunks. The help article bills these LLM calls as Standard Prompts at 4 credits per 2,000 tokens. If each of the 300,000 chunks needs one such call, enrichment adds 1.2 million credits, and the larger index raises the rows each search counts. Trailhead ties it to measurable value from better retrieval.

Hybrid search

Trailhead recommends hybrid search for long queries with specific search terms, and part numbers and error codes are exactly that. Building the index costs the same for vector and hybrid search, but each hybrid query counts vectors and keyword records, so roughly twice the rows. At 14,400 credits a month I would pay that for correct part numbers.

Which content does not belong in the index?

The cheapest megabyte is the one you never process.

• Stock, prices and delivery dates. These live in the ERP and change daily. The agent should read them through an API action, as described in Agentforce and SAP order status, not from an indexed price list. Why the ERP owns spare parts stock is explained in spare parts between ERP and Field Service.
• Superseded catalogue versions. With both versions in the index you pay twice, and the agent may quote an old part number.
• Files above the Data Library limits. Trailhead lists 100 MB per PDF and 4 MB per text or HTML file10, and the data source of a library cannot be changed later. Large catalogues need splitting, ideally by product line.
• Documents nobody asks about. Start with the catalogues behind the most frequent dealer enquiries.

Measure the documents in MB per type

1. Measure the documents in MB per type, after removing old versions and duplicates.
2. Choose the parser per type using a sample in Intelligent Context or the sandbox.
3. Check where the files hang and whether an interface updates those records.
4. Plan the initial load as one batch in one calendar month.
5. Estimate monthly changes, not only the initial volume.
6. Calculate queries as searches × rows in the index, doubled for hybrid search.
7. Set alerts in Digital Wallet11, which tracks consumption down to agent, action and channel, and compare with the estimate after four weeks.

How to plan credits across all meters is covered in Flex Credits and Agentforce cost per action. Whether this budget belongs in Salesforce at all is part of my make-or-buy and licence planning.

Sources11

1. 1Trailhead (opens in a new tab)trailhead.salesforce.com
2. 2billing notes for Agentforce Coworker (opens in a new tab)developer.salesforce.com
3. 3help article on unstructured data billing (opens in a new tab)help.salesforce.com
4. 4$500 per 100,000 (opens in a new tab)salesforce.com
5. 5Trailhead (opens in a new tab)trailhead.salesforce.com
6. 6list of Data 360 usage types (opens in a new tab)help.salesforce.com
7. 7pricing example (opens in a new tab)salesforce.com
8. 8Intelligent Context (opens in a new tab)developer.salesforce.com
9. 9Trailhead (opens in a new tab)trailhead.salesforce.com
10. 10100 MB per PDF and 4 MB per text or HTML file (opens in a new tab)trailhead.salesforce.com
11. 11Digital Wallet (opens in a new tab)salesforce.com