positioning · how the TD learns
How the TD learns.
Every call dispatched, every override entered, every question the TD answered — captured, scoped to your org, and fed back into the system so the next dispatch is sharper than the last. Here’s exactly what gets learned, what stays where, and what the contributor sees.
the principle
The TD that runs your office in year three is meaningfully smarter than the one you started with in year one.
Not because the underlying AI got better. Because the corpus did.
Every confirmed call adds to the office’s understanding of itself. Every BA override teaches the system one specific exception. Every TD answer that gets a correction or a thumbs-down flags a misused pattern. Every advance, rider, gear sheet, and post-mortem feeds the next answer.
The TD doesn’t get smarter by training on the world. It gets smarter by reading what your office has done.
what gets captured
Four categories of signal.
Lifecycle
Every transition in the dispatch loop.
When a request hits the office. When the TD parses it. When the BA approves it. When offers go out. When crew accepts. When the day is worked. When the show settles.
example
612 days dispatched this quarter is the sum of every day-worked row in the system. The TD knows which days were filled cleanly and which required overrides.
TD interactions
Every question, every answer, every correction.
What the BA asked. What documents the TD pulled to answer. Whether the BA accepted, corrected, or ignored the answer. The corrections are the most valuable signal — they teach the TD what it got wrong.
example
BA asks "who FOH’d the Steyer event last August?" The TD answers with a citation. BA corrects with "no, that was the broadcast op, the FOH was someone else." The correction trains future answers.
Behavior
How each role uses the system.
Which surface the crew prefers (text, email, voice, portal). Which time of day the BA opens the queue. Which dashboards get viewed and which get ignored. The patterns of use that shape product priority.
example
87% of crew at this local accept offers via text within 8 minutes. The system optimizes the default channel order accordingly.
Outcomes
What actually happened.
Did the crew show up. Did the show go well. Were there grievances. Did the call run the CBA minimum or longer. What the steward wrote.
example
A 14-day install had a clean day 1 and a chaotic day 7. The post-mortem capture flags day 7 patterns the TD will surface on similar future calls.
the scope wall
Your data trains your TD. Not anyone else’s.
Org-scoped at the database level
Your local's dispatch history, override reasons, TD interactions, and post-mortems train the TD that serves your local. They are not visible to any other org, not used to improve any other org's TD, and not exported for cross-org pattern learning unless your local explicitly opts in. The boundary is enforced by Postgres row-level security — not by application code, not by policy. The database refuses to return cross-org rows. Full stop.
User-scoped for personal records
Pay stubs, W-2s, 1099s, mileage logs, and personal call sheets in xHand are scoped to the individual crew member. Org admins cannot read them. The TD cannot read them. The boundary is Postgres-enforced, not policy-enforced.
No model training by upstream AI
The TD uses Anthropic Claude for reasoning and Voyage AI for embeddings. Both are contracted to not train on customer data. Your corpus is queried at request time; it never improves anyone else's model.
Opt-in anonymized aggregates only
If your org opts in, anonymized aggregate observations contribute to industry-wide intelligence — gear pricing ranges, time-to-fill benchmarks, classification-mix patterns. Floored at n ≥ 5 per range. No individual call, member, vendor, or show is identifiable. Opt-out is one toggle, effective immediately.
the differentiator
You’re not the product. You’re the contributor.
Most software surveils silently. Dispatch surfaces the contribution back to the contributor. Every role sees their own fingerprint on the system.
Crew see
- calls accepted
- days worked
- times their history was cited in TD answers
- response rate to offers
Dispatchers see
- override rate trending down as the TD learns
- TD answer quality improving as the corpus deepens
- office turnaround compressing as workflows tighten
Employers see
- their booking history
- their re-book templates
- patterns the TD has learned about their shows
- the pre-fill on next year's gala
The fingerprint is visible on purpose. Contribution should be felt, not extracted.
the loop
How the loop closes.
- 01
Action
A call gets dispatched. An override gets entered. A TD answer gets corrected. A post-mortem gets captured.
- 02
Capture
The event is written to the dispatch event log — append-only, timestamped, attributed, org-scoped.
- 03
Embed
Relevant events feed the corpus. Confirmed calls become queryable patterns. Corrections flag misused references. Overrides teach exceptions.
- 04
Recall
Next time the TD is asked a related question, it pulls from the deepened corpus. The answer cites the dispatch history that taught it.
- 05
Improvement, visible
Office turnaround compresses. Override rate drops. TD answer quality climbs. The dashboards show it. The contributor's fingerprint shows it.
honest limits
What the TD can’t do.
It can't read what you didn't capture
If a grievance happened and never got documented, the TD doesn't know about it. The corpus reflects what you put in.
It can't fix bad inputs
If an advance has a wrong venue or a misspelled classification, the TD inherits the error. Garbage in, garbage out applies the same way it does anywhere else.
It can't replace the BA's judgment
The TD surfaces patterns and recommends. The dispatcher decides. Every recommendation can be overridden, and every override teaches the system one more thing about how your local actually runs.
It can't predict what hasn't happened
Year one of using Dispatch, the TD is reading from your first batch of contributions. Patterns sharpen with volume. The system gets meaningfully better at year two. Year five is a different system entirely.
next up
How we measure the improvement — days, calls, office turnaround, baselines. The methodology behind every number.