Case study · Salesforce · Apex · LLM agents

A Salesforce AI assistant that costs about $2 per 1,000 questions, on the model you choose.

Holocron answers questions about cases, opportunities and leads in plain language, straight from live Salesforce data. I designed and built it end to end: the reasoning loop, the data tools, the model connection and the admin setup. The same two-step question costs $200 per 1,000 at Agentforce list price.

See it running in Salesforce
Holocron openai.gpt-6-luna
Replay a recorded question

Behind this answer

    These are real turns, recorded in a development org holding Salesforce sample data. The answers and timings are shown as they came back; only the pacing of the replay is shortened. M marks a model call, D a data lookup.

    The product, on screen

    See it running in Salesforce

    Real screens, captured in order during live sessions in a development org that holds Salesforce sample data, on GPT-6 Luna. Nothing is mocked up.

    The full-page app. Find a customer's cases, summarise one, then translate the summary.

    Counts, then shows"7 cases: 2 open and 5 closed", followed by a table the app builds itself, with a CSV of the full result.
    Keeps the thread"Translate that summary" names nothing. The model works from the conversation so far.
    Any language the model speaksThe same summary came back in Spanish, with the case number and dates intact.

    The floating window on a record page. One click on the bubble, then "Summarise this case".

    Opens where the work isThe bubble can sit on any record page, the home page or the utility bar. The window can be dragged larger.
    Knows what is open"Summarise this case" names no case. The assistant already knows which record is open.
    Reads criticallyIt noticed that this sample case was closed a year before it was opened, and said so.
    Typed or spokenThe microphone button turns a spoken question into text before it is sent.

    At a glance

    What it does, in six lines

    Agentic tool calling, native to SalesforceMulti-step, multi-tool reasoning runs in Apex. It needs no MCP server, no middleware and no Agentforce licence.
    Several tools in one stepWhen a question needs unrelated lookups, the model asks for them together and the loop runs them in the same step.
    The model is a setting18 models from OpenAI and Anthropic today. An administrator switches between them from a dropdown.
    The user's own permissionsEvery query and update runs as the person asking. Updates stay off until an administrator turns them on.
    Answers with their evidenceEach answer lists the records behind it, with links and a CSV export of the full result.
    Deploys without configurationWorks on standard objects out of the box. An administrator adds custom objects, fields and business terms without code.

    Cost

    The same question, priced two ways

    Agentforce bills each action an agent takes. Holocron bills only the tokens the model reads and writes, at the provider's own rate. The bars below are drawn to one scale.

    Questions per month Agentforce, 2 actions Holocron, no caching Holocron, 25–55% cached
    1,000$200$2.16$1.21–$1.73
    5,000$1,000$10.78$6.05–$8.63
    10,000$2,000$21.55$12.10–$17.26
    25,000$5,000$53.88$30.26–$43.15

    How the numbers were made: Agentforce is 20 Flex Credits per action at $500 per 100,000 credits (list prices as of 2026). Holocron uses token counts logged from 107 tested questions (two-tool-call questions averaged 19,088 input and 493 output tokens) priced on GPT-6 Luna at $0.10 per million input tokens, $0.01 cached and $0.50 output. This is usage cost only. Licences are left out on both sides, and so is the effort to build and maintain a custom assistant. Full calculation.

    Model choice

    Changing the model is a dropdown, not a deployment

    Holocron supports OpenAI and Anthropic models today, with more providers on the way. An administrator picks a model by name and the next question runs on it.

    OpenAI

    • GPT-6 Luna
    • GPT-6 Astra
    • GPT-6 Sol
    • GPT-6.1 Sol
    • GPT-5.6 Luna
    • GPT-5.6 Sol
    • GPT-5.6 Terra
    • GPT-5.5
    • GPT-5.4
    • GPT-OSS 120B
    • GPT-OSS 20B

    Anthropic

    • Claude Opus 5.5
    • Claude Fable 5
    • Claude Sonnet 5
    • Claude Opus 5
    • Claude Opus 4.8
    • Claude Opus 4.7
    • Claude Haiku 4.5

    Coming next

    • Google Gemini
    • Qwen
    • Z.ai GLM

    Models are served through AWS Bedrock today. Direct Azure AI Foundry and Google connections come next.

    In progress

    Who is spending what, to the minute

    Every answer already writes its token counts to a usage log inside Salesforce. Next, that log becomes a dashboard: tokens, cost and questions for any period, broken down by user, current as of the last refresh.

    Preview · sample data for a six-person team
    Cost
    Questions
    Tokens
    Per 1,000 questionsall users, this period

    By user

    UserQuestionsTokensCost

    The people and figures in this preview are invented, sized from the real averages in the cost section and priced at GPT-6 Luna rates. The log it will read is real: in the development org so far it holds 40 answers, 303,788 input tokens (64% read from cache) and 7,889 output tokens, which comes to about 1.7 cents.

    Where it stands

    Built and tested, and what comes next

    How it is designed, and the before-and-after measurements behind its speed and cost, are what I walk through in an interview.

    Working today

    • Answers over cases, opportunities, leads, accounts and more · search, count, summarise, export, update
    • 18 models from two providers · each tested for tool calling
    • Admin setup page · pick the model without code
    • Deploys to a clean org with no configuration
    • Usage, cost and response time logged for every answer
    • A customer-specific build · in user acceptance testing at an enterprise client

    Contact me

    I'm open to new roles.

    If your team builds on Salesforce or with AI, I'd like to hear from you, and I'm glad to demo Holocron live.

    Email rahul99to@gmail.com
    Phone +91 7416188392

    Holocron was designed and built by Rahul Ravipudi. I also built Zeus AI Accelerator, which answers questions about each new Salesforce record and fills in its fields as it arrives.