R
Before agreeing to a customer's ask, a sales rep needs to know things like “Can we offer this client a 15% discount?”, “Has legal approved this kind of contract before?”, and “Is there a template for this deal size?” Each answer lives somewhere different.
The questions pile up exactly when the pipeline is heaviest. Approvals chased manually push signatures into the next quarter.
Without precedent in front of them, reps discount defensively to be safe. Inconsistent discounting erodes margin one deal at a time.
Precedent and the ‘who to ask’ live in tenured reps' heads and old email. New hires take months to learn what a 30-second lookup could tell them.
Selling time recovered = reps × deals per quarter × hours lost searching per deal × time the agent removes. Plugging in a 40-rep team working 6 deals each per quarter, with 4 hours of search per deal and an assumed 70% recovery rate gives:
One approval chased by email can take days. The agent identifies who signs off and drafts the request in minutes.
Reps often discount high to be safe. With precedent in front of them, they price to the range similar deals actually closed at.
The rep asks. The agent searches all three systems, pulls precedent, template, and approval, and replies in minutes — every fact linked to its source.
“Has a deal like this been approved before? What discount did we give, and who signs off?”
Searches all three systems at once — only what the rep is already allowed to see.
One recommendation: precedent, discount range, the template, and who must approve.
The whole problem is scattered knowledge, and connecting it is what Glean exists to do. The job is to assemble a smart assistant on top of Glean, not build a data pipeline from scratch.
Glean indexes Salesforce, SharePoint, and email out of the box. No new database to build.
Glean inherits each source's access controls, so a rep only sees what they could already open themselves.
Every answer comes with its source, so the rep can trust it and verify in one click.
Glean's workflow steps, search actions, and per-step model choices let us assemble this fast.
Four layers, one direction of flow. Glean sits in the middle as the unified, permission-aware search layer.
Built in Glean Workflow mode so steps run in a fixed, auditable order. Each step has a clear job, a Glean step type, the sources it touches, and a model chosen for that job.
Rep asks in plain language. The question and the linked deal are pulled into agent memory as starting context.
Reads intent, works out what “a deal like this” means, then pulls deal size, segment, industry, discount, and terms from Salesforce — rep types nothing extra.
One question fans out into three searches at once: similar past deals, the right contract template, the matching approval thread.
Opens the exact template and approval email that matter, so the answer quotes the real content rather than a snippet.
The judgement step: rank comparable deals, compute the discount range, apply the approval rule. Rules are applied exactly — never guessed. Runs on prior-step memory only.
Turns the decision into one sourced answer card — precedent, discount range, template, and the approval note with links. Formats only; it does no math.
If sign-off is needed, the agent asks the rep to confirm before doing anything. Nothing is sent without a person saying go.
On confirmation, drafts the approval request — pre-filled with deal details and precedent as justification — and routes it to the right approver. A person still approves.
A rep is working a $300K mid-market retail deal and asks the agent a single question. Here's what the agent's answer would look like, based on this design.
Existing permissions are respected. Glean inherits the access controls each source enforces, so a rep only sees deals and documents they can already open. The agent never grants or widens access.
Every fact is cited. Each number ties back to a real deal, template, or email. No source, no answer shown.
Rules are applied, not guessed. The decision step states the approval rule; the model only writes the wording — it can't invent a price or an approver.
It admits gaps. If there's no truly similar deal, it says “no clear precedent” instead of making one up.
A human still decides. The agent informs and drafts; it never sends or commits anything on its own.
Saving time on each deal is the start. The compound effect is what makes this worth building.
The real setup work is making sure the connectors are indexed and the pricing and approval rules are findable. Rollout leads with approvals — the biggest time sink — not just precedent lookup.
Launch with the questions reps ask most — what did we give, and who has to approve it, surfaced from day one.
Add template selection so the answer also returns the right contract for the deal size and type.
Let the agent draft and route the approval request to the approver — still with a human signing off.
Same agent, more questions: terms, clauses, SLAs — and more segments over time.
Measure it simply: time-to-answer before vs after, and days-earlier a deal closes. Those two numbers are the whole business case.
Retrieval and formatting use fast, precise models. The one step that weighs several conditions at once gets the most capable model — speed where it's a lookup, depth where it's a decision.
| Step | Glean step type | Model | Why this model |
|---|---|---|---|
| Understand the deal | Plan & Execute (Thinking) | Claude Sonnet 4.6 | Reads a messy question and fills the gaps from context |
| Search 3 systems | Company Search | GPT-5.4 Fast | Fast retrieval, runs the three searches in parallel |
| Read documents | Read Document | GPT-5.1 | Precise single-document extraction |
| Reason & decide | Think (Deep Reasoning) | Claude Opus 4.6 | Weighs several conditions and precedent at once — the critical step |
| Compose the answer | Respond | GPT-5.1 | Clean, faithful formatting of the decided answer |
| Draft & route | Respond → Write action | GPT-5.1 | Reliable, instruction-following draft for the request |
Solution design · Deal Desk Acceleration Agent on Glean. This is design and architecture work, not a delivered program — the agent platform on the homepage is where the production numbers live.
← Back to portfolio