Case study
August 2026
In use
An analyst for every rep.
A Claude Project per salesperson, wired to their own CRM login. They name the account and what they need, in plain English. It reads the live CRM and the call transcripts, does the work, and writes back only what they approve.
- Skills behind one prompt
- 13
- Qualification areas graded
- 8
- Project per rep
- 1
- Current stage
- In use
Context
Every rep researched accounts, qualified deals, and prepped demos their own way. The institutional knowledge that should have shaped that work, the qualification methodology, the competitor battlecards, win and loss history, pricing, the territory map, all existed. It just lived in documents nobody opens in the middle of a deal.
Meanwhile the CRM record and what actually happened on the calls had drifted apart, which is normal and also the reason a lot of forecast conversations go badly.
The problem
A general AI assistant is only as current as whatever the rep remembered to paste into it. Ask it where a deal stands and it will answer confidently from a stale fragment, which is worse than not asking.
A shared prompt library fixes none of that. It drifts, it is optional, and it has no owner, so within a month six people are running six different versions of the same play.
What I built
One Claude Project per salesperson, each connected to that person's own CRM login, so it only ever sees their accounts and writes as them. The skills and the reference library are shared and centrally maintained: individual workspaces, one common playbook. Fix a skill once and everyone gets the fix.
Thirteen skills sit behind a single plain-English prompt, covering territory planning, account research, outbound, inbound qualification, deal scoring, risk reads, demo prep, follow-up, ROI modelling, and competitor comparison. The rep never names a skill or remembers a command. They type the account and what they need.
Rules and facts are kept apart on purpose. Behavioural instructions stay short because the corpus of facts lives beside them rather than inside them. Nothing is cached between sessions, so every run reads current data, and state lives in the CRM rather than in chat history.
What happens when you hit enter
1
Identify
Work out which job the plain-English ask maps to
2
Resolve
Find the account, from a name or a pasted CRM link
3
Pull
Read the live CRM record and the call transcripts
4
Run
Execute the skill against what it actually found
5
Propose
Show changes as a table and wait for approval
6
Log
Record which job ran, on what account, and when
The guardrails are the product
Sales teams do not adopt a tool that might quietly edit their pipeline. Four behaviours make it safe enough to hand over.
- It asks before writing
- Every proposed change renders as a table: current value, proposed value, why. Approve it, take part of it, or ignore it. Do nothing and nothing happens.
- It shows its sources
- Each claim is tagged as confirmed on a call, assumed, or an internal benchmark. When it does not know, it says so rather than filling the gap with something that sounds right.
- It will disagree with the CRM
- If a contact is flagged as a strong advocate but never actually pushed internally, it says so and explains why. The rep can overrule it.
- Runs get logged
- Which job, what account, when. That feeds improvements to the tool, not performance reviews, and logging never touches deal data.
The rep-facing guide
I wrote the onboarding guide as its own artifact, with the architecture folded away behind a disclosure so reps see the usage guide and the curious can read the build. It is embedded below in full. Open "How it's built" to see the nine layers and the run pipeline.
Loading the live artifact…
Detail
A worked exchange. The rep types two words; it resolves the account, both open deals, the qualification state, the 68-day gap and the expired incentive it found in the call transcripts, then stops and offers next steps.
All thirteen skills mapped to when you would reach for them and what comes back, with the four behavioural guarantees underneath.
Outcome
It is in use with the sales team, and the reception has been the useful kind: reps asking for more skills rather than asking whether they have to use it.
The part that transfers is the shape. One maintained playbook, individual permissioned workspaces, plain English at the front, and guardrails strong enough that handing an AI system write access to the pipeline is not a leap of faith.