At a glance
| Sector | Organized labor. A national union representing tens of thousands of licensed professionals |
| Structure | Members work under a separate collective agreement with each employer, negotiated and administered by that unit |
| Objective | Give every member a correct answer about their own agreement, without routing every question through a representative |
| Key use cases | Member self-service on contract questions, pay and scheduling interpretation, representative preparation, military leave rights, HR and internal governance |
| Solution | CustomGPT.ai. Twenty-one grounded assistants on one enterprise agreement, each scoped to a single contract, embedded in the union’s member site behind its own sign-on |
| Corpus | 42.5 million words across 1,132 documents, in separate libraries so no assistant can reach another unit’s terms |
| Access | The union’s existing sign-on handles authentication, and page-level rules decide which unit’s assistant a member reaches |
| Share | Anonymously, at the customer’s request |
The challenge
A member on a rest day working out what a schedule change does to next month’s pay. A local representative taking the same question from thirty people in a week, on an agreement he never sat at the table for. A staff officer at head office who knows all of them and can answer any of them, for one person at a time.
Every employer group bargains its own terms, so the answer that is correct for one member is wrong for the member sitting next to her. One assistant trained on all of the agreements would be worse than no assistant at all, because it would sound certain. The union’s problem was never a shortage of expertise. It was that the expertise lived in a few people’s heads and in documents nobody reads end to end.
The solution
- One agreement per assistant, isolated from the rest. Each bargaining unit has its own, grounded in that unit’s contract and nothing else. Narrow scope is an accuracy decision before it is a design decision.
- The member reaches it where they already are. Inside the member site, behind the sign-on they already use, on the page they already visit. A separate destination is a destination nobody goes to.
- Cite the exact clause, and know when to escalate. Routine questions get answered. The hard ones get pointed at the clause and handed to a person. A union that answers everything will be wrong in public.
- Representatives trust it before members see it. Each unit set an accuracy bar on its own agreement and had to clear it before the assistant was switched on for members.
- Agreements change, and nobody rebuilds. Amendments, side letters and new terms land constantly, so source documents sync from the union’s own document store.
The capability that made it possible
In a union, accuracy depends on who is asking. The same words about pay are correct for one member and a grievance for the member sitting next to her, and no amount of general intelligence resolves that, because the model has no way to know which contract governs the person typing.
Scoping each assistant to a single agreement removes the wrong answer from reach entirely. Putting the right assistant in front of the right member, using the access rules the union already runs on its own site, does the other half. That is why there are 21.
The results
| Measure | To Date |
| Grounded assistants in production | 21 |
| Questions asked | 8,182 |
| Questions in the busiest month | 1,991 |
| Growth, latest six months on the six before | 4.9× |
| Questions answered, six-month blend | 87.4% |
| Longest unbroken run of daily use | 71 days |
The number to read first is 4.9. Questions across the latest six months ran almost five times the six months before them, on an estate that was already live and already working. The busiest month on record is the most recent complete one.
Three quarters of questions arrive through assistants embedded in the union’s own member pages. Nobody had to be sent somewhere new.
The run of 71 straight days with at least one question is still going, which is what it looks like once representatives stop treating something as a trial.
Why it worked
- Data isolation sits at the assistant level. Each one is grounded in a single agreement, so the answer that would be wrong for this member is not reachable at all.
- It lives inside the member site. Behind the union’s own sign-on, with page rules deciding who sees what.
- Short source libraries. Every assistant is held to the few documents that decide the answer, which did more for accuracy than adding material would have.
- The difficult document became our problem. One employer’s contract arrived in a layout nothing could parse. We fixed it on the platform, and every customer has that fix now.
Frequently asked questions
What is an AI contract chatbot for a labor union?
It is an assistant grounded in a single collective bargaining agreement, so a member can ask a plain-language question about pay, scheduling or leave and get an answer drawn from their own contract with the clause cited. Here there are 21 of them, one per bargaining unit.
Why one assistant per agreement rather than one for the whole union?
Because the same question has a different correct answer in every unit. An assistant holding two contracts will blend them on the question that matters most, and it will sound certain while doing it. Scoping each to one agreement makes the wrong answer unreachable rather than merely unlikely.
How accurate is it?
This estate answers 87.4% of questions on a six-month blend. Accuracy here is as much about isolation as about retrieval: because each assistant can only reach one agreement, the failure mode where a member gets another unit’s terms cannot occur.
Does it cite the exact clause?
Yes. Routine questions are answered with the clause cited. Harder questions are pointed at the clause and handed to a person, which is the boundary that makes representatives willing to put it in front of members.
Does it cover side letters and memoranda as well as the main agreement?
Each assistant’s source library holds the few documents that decide the answer for that unit, and source documents sync from the union’s own document store. Amendments, side letters and new terms land constantly, so a corpus that updates itself is what stops the estate going stale inside a single bargaining round.
How do members reach the right assistant?
Through the union’s own member site, behind the sign-on they already use. Page-level rules decide which unit’s assistant a member gets, so nobody has to know which one is theirs. Three quarters of questions arrive this way.
Can it be used by representatives and staff as well as members?
Yes. Four of the 21 point inward rather than at members: HR, military leave rights, the union’s own constitution and admin manual, and the agreement that sits above the individual units. Internal assistants have no public downside, which makes them an easy place to start.
How did the union make sure it was safe to launch?
Each unit set an accuracy bar on its own agreement and had to clear it before the assistant was switched on for members. Units came on one at a time rather than as a single program, so nobody shipped on hope and nobody had to walk anything back.
What happens when a contract is in a format the platform cannot read?
One employer’s contract here arrived in a layout nothing could parse. We fixed that on the platform rather than asking the union to reformat anything, and every customer on the platform has that fix now.
What does it cost to add the twenty-second unit?
Nothing beyond the existing agreement. All 21 sit on one contract with us, so a new unit is a configuration rather than a procurement.
Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.
