Field notes
The butchering of Sales Navigator
FBED Labs is preparing for its first epoch: one complete pass of the entire training dataset through a machine learning model. What surprised us first was not the learning. It was how few prospects Shannon admitted from what should be good Sales Navigator searches.
How Shannon learns: live policy
Our models run a little differently to most attempts at commercial processes. They are live policy: our machine learning systems learn whilst they work, so a data pass happens on real prospects, for real customers, in the real world.
We do not do the work for Shannon. We approve, disapprove, add notes and monitor. Our systems traverse every decision point between buyer and seller across the whole commercial process.
Where we are: logging policy
We are in logging policy. That means testing different states and moves, and building the labelling that trains our support models.
Those support models, such as Shannon Huffman, give a quality metric and a management system to frontier models such as Claude, which turn Shannon's output into human language. The hard part is agreement: given the same context, how closely do two separate models produce the same output and the same score? The answer is not very well. That is why Shannon Huffman exists.
Why Sales Navigator searches admit so little
Our first logging process runs on LinkedIn, and Sales Navigator is its sales platform. You use its advanced search to pick the filters that match your ideal customer profile.
What we did not expect was for Shannon to admit so little volume from such large searches. It tells us there is a quality problem on LinkedIn that takes a long time to unpick when you build lists, and it could be hurting your numbers.
The numbers
- Collected since logging began: 17,362 prospects.
- Collected between 12th and 23rd October: 12,771.
- Admitted by Shannon: 1,528, the prospects she judged to be valid behavioural and deterministic ideal customers on our real world variable (RWV) scores.
- Admission rate: 12.0%.
Shannon Desktop · 26 Oct 2026
Two ways to read it
- Shannon is admitting only the people she thinks will be sincere ideal clients for our clients.
- There is something we are missing.
Our enrichment and our ability to extract from LinkedIn are very good. Over the next week we will widen the test to see whether the admission rate really stays under 12%. If it does, the people who use the platform more loosely, our competitors among them, may be losing later conversations from the very start. Shannon Desktop · 26 Oct 2026
Approval and Shannon's quality
In the first field period, Shannon presented 1,142 messages as opening moves for the prospects she admitted. Our operators approved 538, an approval rate of 47%. Shannon Desktop · 26 Oct 2026
On the surface that feels low, and frankly it is. A large share of them are still waiting for review.
This is the first true reporting week. Shannon is working her way down the decision process, and we expect to cross the scheduling barrier this week, as prospects are now pooling inside our conversation bands.
What is changing for Shannon this week
- Shannon Transition 1.0 is being installed after completed model evaluations. It is a big semantic upgrade.
- Decision infrastructure will be completed, with new RWV and rubrics that resemble the fixed weights we will deploy into the systems over the next year.
- T tests: our own concurrency and volume tests, and stronger baseline metrics for acceptance, reply and scheduling rates.
Shannon Desktop · 26 Oct 2026
We expect the first completion metrics and analysis 2 weeks from today. Shannon Desktop · 26 Oct 2026
What we'll report next Monday
Acceptance, admission, reply rates and initial scheduling data.
Take part in the research
If you would like to get involved by giving us 1-2 of your team for outbound management and sell-side planning, contact sarabrazier@fbed.io. Shannon Desktop · 26 Oct 2026