Dizzzank (@dank0296) on X

What it claims

This is a thread summarising a talk by a forward-deployed engineer (FDE). The speaker is not named. The thread adds detail the catalog summary leaves out.

Where the time goes. Speeding up every step of a workflow often does not shorten cycle time, because "the days are hiding in the handoffs between steps." If you only optimise the individual tasks, you may see no change in end-to-end time.

More of the case study. The $5B company's process was mapped as seven steps. Cloning the CRM in a sandbox showed twenty steps and seven loops:

• Deals that had died came back.
• Legal sent contracts back 12% of the time.
• Customers changed specs 30% of the time. The rest of this sentence is cut off on the page.

The full "CFO slide." These are the before and after figures for the rework:

The slide then compares the cost of running the agent with the cost of the human labour it replaces. That comparison is also truncated.

Note that the slide says 17 steps, while the discovery phase found twenty. The thread does not explain the difference.

How it works

Discovery takes about a month.

• Interview the longest-tenured staff. The speaker's reason is that "process maps go stale within months."
• Embed 3–4 engineers in the system of record for a month and watch where work loops.
• Mine the leftovers: SharePoint, Slack and email.

The speaker's warning: "there is no second brain." Do not expect to plug into an existing knowledge store such as Obsidian. Building that store, one department at a time, is the actual work.

Program structure for companies:

1. Baseline the KPIs before starting.
2. Pick processes that have a named owner.
3. Sort every step into the four buckets.
4. Build once and deploy everywhere.
5. Measure constantly.

The timeline is four weeks of audit, four weeks of build, then proof at "three and six." The unit is cut off; months is likely but not stated.

Why not migrate to new tools. One customer spent years and $10M migrating their ERP. Pitching a new AI-native CRM to such a customer "kills the deal." The pattern instead is to read data from Salesforce, take actions in Salesforce and notify people in Slack.

How to use it

The speaker gives a week-one playbook for individuals who want to get into this work:

• Monday: List every app that holds your data. Note which one is authoritative and where they disagree.
• Tuesday: Take 20 things you did last week.
• Wednesday: Write one process out step by step.
• Thursday: Sort every step into the four buckets.
• Friday: Reach out to SMBs. The rest of this item is truncated.

Limits

• This is a second-hand recap of a talk. The speaker, the company and the source video are not identified.
• Several posts are cut off at "Show more," including:
• the loop statistics
• the agent-versus-human cost comparison
• the human-in-the-loop rationale
• the playbook's Friday step
• the proof timeline
• The model-benchmarking advice names model families (GLM, Kimi, Qwen, Grok) but gives no evals, scores or costs.
• None of the figures can be independently checked, and the 17-versus-20 step count is not reconciled.