Corporate banking email · 3 observations: vague action ("I'd like to understand if we can proceed"), opening repeats last week's call, closing tone more cautious than usual.
AGENT · SALES COACH
Support on the sales team's emails, customer by customer.
Sales Coach reads the sales team's outbound emails, recognises the customer context in shared memory, and suggests improvements on tone, quality of the requested action, and consistency with the ongoing conversation. The sales rep stays the author; the agent is the quiet reviewer who sees what the individual cannot. Integration with the corporate email system is delivered during the project.
Sales Coach at work.
Accepting 1 and 3. Ignoring 2: I want to reinforce the point from the previous call for the customer.
Corrections applied. Pattern "concrete reformulation + two options" saved in shared memory.
Why it exists.
The sales rep's outbound email is the most-read document in the customer relationship — and the one where no structured coaching happens. Manager review comes on a few cases a month, usually after the customer has already complained. Internal training runs through semi-annual workshops. In between, the rep writes dozens of emails every week where tone, clarity of the requested action, consistency with the previous conversation all shape the outcome.
How it accompanies the email.
The agent activates on the team's outbound emails. It reads them before sending (review mode) or shortly after (periodic coaching mode), compares the content against the shared customer memory (history, support tickets, active deals, previous conversations), produces a structured note. The note recognises where the email works, flags where it does not, proposes ready-to-use text alternatives.
The decision stays with the rep.
The rep chooses what to apply, what to ignore, what to save as a pattern. The agent does not send emails autonomously, does not silently correct, does not replace the rep's judgment. The default is self-coaching or team-aggregate, outside the AI Act Annex III 4(b) perimeter on worker monitoring.
Who it serves and where it applies.
Sales manager
Recognises the team's recurring writing patterns. The team-aggregate summary shows what works and what needs correction at the team level, and becomes the basis for targeted training rather than subjective perception. The data is aggregated, not individual profiling.
87 Tone too formal — try a direct opener
71 Long message: summary in the first 2 lines
94 Excellent — consistent with the customer history
Junior sales rep
Accelerates the learning curve. The revision arrives on the current case, alongside the message being drafted — not in a workshop three months after the mistake. Shared memory surfaces patterns the junior rep cannot see for lack of customer history.
RB RevOps Brief
Idritech SpA silent 8d · proposal expiring
Nexar Group proposal sent · no reply
Aludra Srl signature expected this week
Senior sales rep
Recovers the time spent on self-review. The email is the same, the writing pattern is established; the agent prepares the final check and frees up the half-hour of re-reading that today goes into every important proposal.
CustomerSignalRisk
Idritech SpA Silent 12 days · 3 emails unanswered high
Nexar Group Proposal expired · no extension med
Orbis Srl Meeting cancelled 2× med
Proxa Corp Active contact · reply < 24h low
A concrete example.
Three emails to send tomorrow morning.
For a corporate banking sales rep, the end of the day is re-reading three emails to send the next morning. The corporate email integration, delivered during the project by the Exelab team, activates Sales Coach in pre-send review mode. The agent reads the first message, recognises the customer in the CRM, retrieves from shared memory the last six months of context.
Action, opening, closing tone.
The requested action is too vague ("I'd like to understand if we can proceed"): the agent proposes a concrete reformulation with two operational options and a follow-up date. The opening paragraph repeats information the customer already knows from last week's call: the agent suggests removing it. The closing tone is more cautious than the same rep's previous emails to similar customers: informational note.
The rep chooses in two minutes.
The rep accepts the first and third observations, ignores the second because they want to reinforce the point from the previous call. They apply the corrections in two minutes. The event — rules triggered, alternatives chosen, handling time — remains in the runtime audit registry, queryable by the sales manager with a standard SQL client.
Configuration and technical resources.
Sales Coach's rules are declarative. The customer's Sales Operations team defines, in a readable format, what "consistent tone" means for the company (deal stage, interlocutor type, customer sector), which writing patterns to reinforce and which to correct, in which order observations appear in the revision note. The rules live in the customer's repository, versioned, validated at agent startup. The individual evaluative mode is an explicit opt-in with documented AI Act Annex III 4(b) obligations; the default is self-coaching or team-aggregate.
- Language
- TypeScript (Node.js)
- LLM model
- customer's choice: Anthropic, OpenAI, Mistral, open source models hosted internally, AWS Bedrock for a private model
- Built-in controls used
- pii-detector, credential-detector, topic-guardrail, message-length-limit
- Native delivery channels
- Telegram, Slack, WhatsApp, HTTP OpenAI-compatible
- Corporate email integration
- delivered during the project by the Exelab team; not a native channel from day one
- Worker monitoring mode
- default self-coaching or team-aggregate; individual evaluative mode is an explicit opt-in with documented AI Act Annex III 4(b) obligations
- Memory
- persistent per instance, pgvector + PostgreSQL FTS on customer context
- Working modes
- pre-send review, periodic coaching (configurable per instance)
- Registry
- append-only, queryable with a standard SQL client
Frequently asked questions about the agent.
No. The rep stays the author of the email. Sales Coach reads the text, recognises the customer context in shared memory, proposes observations and text alternatives. The decision on what to apply, what to ignore, what to save as a pattern stays with the rep. The agent does not send emails autonomously and does not silently correct.
The native channels from day one are Telegram, Slack, WhatsApp, and HTTP OpenAI-compatible. Integration with the corporate email system (needed for pre-send review mode) is delivered during the project by the Exelab team. The options depend on the customer's system: IMAP forward, interception webhook, dedicated adapter.
The typical pattern for periodic coaching mode is 3-5 weeks. Discovery one week, declarative rules configuration two weeks, email read integration one week, hand-off to the sales manager. For pre-send review mode, add 1-2 weeks for the depth of corporate email integration.
Yes. The agent runs on a dedicated customer instance, with per-instance encrypted secrets, and the audit registry lives in the customer's database. Sales team email data does not leave the customer's infrastructure (self-host) or the Polyant-managed infrastructure in the chosen managed profile.
From a 30-minute conversation to the squad in production.
A 30-45 minute conversation to understand how Sales Coach would configure to the customer's case. Which mode (pre-send review or periodic coaching), which writing patterns to configure, which corporate email integration.