Solutions · Customer support

Trained on Your Tickets.
Not the Whole Internet.

General models answer support questions like a new hire on day one: politely, slowly, and with a lot of guessing. Parsec trains models on your conversations, policies, and tools, and serves them in production. You own the models.

cost per resolved conversation: −50–60%; cost per resolved conversation
median decision latency: 69 ms; median decision latency
resolution quality: +5–15%; resolution quality

Get Started Talk to Us

The problem

A general model rediscovers your business on every call.

A frontier model knows a great deal about everything and nothing specific about you. Prompting harder does not fix it.

Guesses at your policy

Refund windows and escalation paths live in a prompt it re-reads every turn. A prompt is a request, not a guarantee.

Invents arguments

Ask it for an order ID or a SKU and it writes one that looks right. Sometimes it is not the one in the record.

Always confident

Every answer comes in the same even tone. Nothing tells you which turns to hand to a human.

Slow and expensive

Most support turns are decisions, not conversations. Generating them token by token costs seconds and dollars.

What Parsec delivers

Custom intelligence as a service.

We train the models that should handle your traffic, prove them on your evals, serve them, and keep retraining them as your business changes.

A team of models

Each job on a turn gets the smallest model that does it well. A TOD-shaped decision model ranks your allowed actions and only serves when confident. Your existing LLM handles whatever the team defers.

Your policy, in the weights

Trained on your resolved conversations and the actions your agents took. It knows your refund policy the way your senior agents do.

Evals first

Evaluations built from your real traffic before any training. You see the numbers before you commit traffic.

Post-training, not just fine-tuning

Reinforcement learning in a shadow environment, with rewards built from your outcomes, resolves tickets the way your best agent would.

Never too early

Your models run on Parsec’s shared runtime with a universal K/V cache. There is no minimum traffic that makes them worth it.

It keeps learning

New resolved tickets become new training data. Retraining is part of the service.

What support teams see · Measured on held-out conversations

Arguments resolved: 94.7%; exact match, 694 argument sets; Selected from real records. It cannot invent one.
Calibration error: 0.8%; when it says 90%, it is right ~90% of the time; That makes the handoff threshold a dial you can trust.
Reads the photo: 97%; on chart questions, held out; 87% on diagrams, 90% on science questions.
Action catalog: 300; options per decision; Most decision models cap at 26 because they read answers as letters.

Decision latency is 69 ms median and 121 ms at p95.

Benchmark and methodology

Side by side

Prompted vs trained.

Prompted vs trained.
	Custom models on Parsec	Prompted frontier model
Where your policy lives	In the weights	In a prompt, re-read every call
What is guaranteed	Behavior measured on your own evals	Nothing; you hope it follows instructions
Action arguments	Selected from real records	Generated; can be wrong
Confidence signal	Calibrated probability per decision	None you can threshold
Human handoff	Set a threshold, defer below it	Rule-based or manual
Images and screenshots	Read alongside the text	Depends on the model and prompt
Decision latency	Under 100 ms	Seconds
Cost per conversation	50 to 60% lower	Full frontier price, every turn
Who owns the model	You	The lab
Gets better over time	Every retraining pass on your tickets	When the lab ships a new model
Hard turns	Passed to your existing model, unchanged	Attempted anyway

How it works

From your tickets to production.

01
Weeks 1–2: evals
Send a slice of resolved conversations in any shape. We score your current setup against evals built from them.
02
Pick one flow
Order status, returns, billing disputes. A month of tickets in one narrow flow is enough to prove it out.
03
Weeks 2–4: training
We curate the data, train the team of models, and post-train them. First custom models serve in about four weeks.
04
Set the threshold
Every decision carries a calibrated probability. Start conservative and open it up as the evals earn it.
05
One-line switch
Point your assistant’s inference at Parsec. Deferred turns pass through unchanged. Nothing else moves.
06
Month 3 on: compounding
New tickets feed retraining. Cost per conversation keeps falling. The models are yours at any point.

When it is not the answer

Custom models earn their keep on volume and repetition.

Too few tickets

Under a few hundred resolved conversations in the flow. Route through Parsec anyway; we start training once there is enough.

Open-ended conversation

If turns are mostly prose rather than decisions, they still go to your frontier model and the savings shrink.

Weekly policy changes

A tuned model learns the policy as it stood. Fast-moving rules belong in the prompt until they settle.

Unchecked writes

We defer write actions until your evals show the models earn that trust. Decide your risk line first.

Common questions

Questions.

Do I have to replace my current assistant?

No. Your custom models sit in front of it. Confident turns never reach your frontier model. Everything else passes through unchanged.

Which models does it work with?

Any model you call over an API, as the parent. If you can point traffic at an endpoint, it works with Parsec.

What does it cost?

Usage pricing on the traffic Parsec serves. No reserved capacity, no idle GPUs, no bill that climbs with context length. Training and post-training are part of the service.

Is my data used for other customers?

No. Your data trains your models and nothing else. It is never pooled across customers.

Who owns the models?

You do. The checkpoints are yours. Run them elsewhere if you ever want to.

Can it run inside our network?

Yes. Enterprise deployments run the full stack in your cloud with licensed, containerized checkpoints. Your tickets never leave it.

What if a model is wrong?

Every decision reports a probability. Set the threshold so uncertain turns defer, and audit the served ones.

How long until production?

Evals in about two weeks. First custom models in about four. Compounding gains from month three.

Has this been done before?

Yes. The same pipeline cut time to first token by 60% and time per output token by 40% for voice agents, with quality up.

Your Agents Already Know the Answers.

Train a model that learned them from your tickets. Start with a two-week eval.