The 30-Day Agent Training Loop
Ingest your docs, then correct the agent an hour a day for 30 days until it performs like your best rep.
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 95%
A concrete daily routine for turning a raw AI agent that 'kind of knows it and isn't great' into a version of your best salesperson. It demystifies the jargon (ingestion, training, QA) into: upload data, answer the agent's questions, then correct its live mistakes daily for a month. Use it whenever you deploy any GTM agent.
Origin
Lemkin argues the 2024 vendor lie was 'just turn it on, no training needed.' Having done it ~20 times, he reframes 'training' as simply answering questions and correcting daily mistakes over 30 days — something anyone with B2B experience can do.
Core principles
- 01Ingestion just means uploading and processing your data (website, wiki, docs) — the RAG/vector mechanics don't matter to you.
- 02Training is nothing more than answering the agent's questions and correcting it repeatedly.
- 03Agents ship imperfect out of the box; the daily correction cadence is what makes them good.
- 04A month of consistent correction beats years of accumulated content — you need a little good data, not a huge archive.
How to run it
- 1
Feed it your URLs and documents
Give the agent your website URL, wiki URL, training docs, and upload a few key documents (prospectus, best scripts). It ingests and processes them.
Pro tip Lean on the vendor's forward-deployed engineer to help with this first upload.
Watch out You may instantly break it with too much data — SaaStr overloaded Delphi and it took about a week to get going.
- 2
Answer its training questions
The agent turns your data into questions. Answer them — the more you answer, the better it gets. This IS the training, nothing more mysterious.
- 3
QA test before going live
Have it send practice emails and check them. It will say dumb things (hallucinations, wrong facts). Make sure it's right before real prospects see it.
Watch out Early on it may state wrong dates or facts confidently — catch these in QA, not in front of customers.
- 4
Correct daily for 30 days
Each day the agent sends emails and makes mistakes; spend an hour or two correcting them. By day 30 it will be 'pretty good' — a facsimile of your best person.
Pro tip Budget 50-60 hours total plus vendor qualification time for your first agent; the second is far easier.
Watch out If you buy a product and disappear, you'll get zero ROI. The daily oversight is non-negotiable.
In the wild
When SaaStr first used its clone for support, it told people the wrong event dates. Lemkin spent almost an hour every morning firing it up, reviewing issues, and answering them.
→ After enough correction it became well-trained; he no longer has to touch it.
After months training three sales agents down to one refined prompt, SaaStr took that single prompt and gave it to Agentforce.
→ Agentforce was 'pretty good' within a day — the second agent is dramatically faster than the first.
Common mistakes
Believing 'no training needed'
Vendors in 2024 claimed the product would generate revenue on turn-on. It fails without training and daily correction.
Assuming you need years of data
Lemkin thought 12 years of content mattered; in fact a couple of months of really good content plus a long tail is enough.
Is it for you?
Best for
Anyone with B2B/SaaS experience deploying their first GTM agent who is intimidated by the jargon
Not ideal for
Teams unwilling to commit daily hands-on time to correction and oversight
From the transcript
“if you do this for 30 days and every day you spend an hour or two correcting those mistakes by the 30th day it's going…”
“Training is just answering questions and getting better and better.”
“it might take you 50 or 60 hours plus qualifying the vendor”
From the episode
We replaced our sales team with 20 AI agents—here’s what happened
Jason Lemkin (SaaStr)