Every company works differently.Its AI should too.

Ack trains specialized models on a company's workflows, rules, and operational history, then checks their work against the system of record.

See how Ack worksRequest a workflow audit
Dithered photo of stacked CRM documents with chat prompts asking whether a lead is qualified, whether to route to enterprise, and whether to update the deal stage, beside a Deal Review form.
  1. Written rules
  2. Company-specific training
  3. Verified actions

  • Policies only tell part of the story.

    On a second company's never-seen test, Haiku 5.5 given the written policy completed 61 of 90 tasks.

  • A fine-tuned open model learned more.

    On the same test, our fine-tuned 9B with RL scored 75 of 90 vs 61 for Haiku 5.5 given the written policy.

  • Proof is in the database.

    All outcomes were checked against the actual database writes, not judged from the agent's response.

Ack CRM research · synthetic Salesforce CRMArena-Pro · 2026

How it works

  1. 1Learn

    Collect workflow examples, policies, and operational history to understand how a company works.

  2. 2Train

    Fine-tune an open-weight model for one company's workflow and the tools it needs.

  3. 3Verify

    Grade the work against the system of record, not what the model says it completed.

Dithered photo of a support agent wearing a headset, working in a CRM at a desk.

CRM workflow model · Ack 9B

Qwen3.5-9Bfine-tuned + RL90 held-out taskssecond company, held out

Verified task completion

75 / 90Ack 9B (fine-tuned + RL)
tasks completed correctly

  • Haiku 5.5 + written policy61 / 90
  • Ack 9B (fine-tuned + RL)75 / 90
  • Haiku 5.5 + rules we extracted by hand77 / 90

    Not an apples-to-apples baseline: we read the company's history and wrote the rules for Haiku by hand. That extraction is the work Ack is built to automate.

deterministic database grading0 wrong-record writes across the fine-tuned runs reported

First company · 89 tasksthis test informed our data design

  • Haiku 5.5 + written policy50 / 89
  • Sonnet 5.5 + written policy53 / 89
  • Ack 4B (fine-tuned)79 / 89
  • Haiku 5.5 + rules we extracted by hand77 / 89

    Not an apples-to-apples baseline: these rules were written by hand after looking at test failures.

Synthetic Salesforce CRMArena-Pro data (CC-BY-NC research data, credit Salesforce) · deterministic database grading · single training seed · no customers yet

See the results

Ack CRM research · Synthetic benchmark result, not customer traction

General-purpose AI knows the rules. Your company has its own.

  1. 01Company workflows + examplesInput
  2. 02Company-specific trainingAdapt
  3. 03Specialized modelExecute
  4. 04Verified workDatabase ✓

Starting with CRM

Lead qualification, ticket routing, and deal-stage correction. One narrow workflow at a time, with outcomes we can verify.

Building toward company-specific intelligence

As open-weight models improve, Ack can train and maintain specialized models for more workflows across a company's systems.

Your company already knows how to do the work.

Give its AI a way to learn.

Request a workflow audit