MethodPricingCalculatorDemoBlog

Shadow audit · Migration · Run

Measure what Jev saves your workflows, before you migrate.

A shadow-mode audit on your data, then migration of your AI routing to Jev by TypeSafe AI, with LLM fallback, a decision log and human review.

Book the free pre-qualificationEstimate my savings

Independent service, not affiliated with TypeSafe AI.

01 · The problem

LLM routing costs money, and nobody knows when to trust it.

cost

Every routing decision goes through an LLM billed per token.

trust

Your teams review everything, because they cannot tell when the model is right.

proof

Jev has just launched. Its numbers come from the vendor. Nobody has measured them on your data.

02 · The method

Measure first. Migrate only if the numbers justify it.

  1. step 1free

    Pre-qualification

    30 minutes. A written go or no-go within 48 hours.

  2. step 22 weeks

    Shadow-mode audit

    We replay a pseudonymised sample of your past decisions on Jev and on your LLM. Nothing changes in production.

  3. step 3after an audit

    Migration, then run

    Gradual rollout, with the kill switch in your hands.

03 · The verdict

An honest verdict, three options.

The audit measures all three options on your real cases. It recommends the cheapest one that holds your quality bar.

option A

Keep your current routing

No change. The baseline for the other two.

The right choice if your bill is small or reviews are already rare.

effort
none
LLM bill
unchanged
reviews
unchanged
vendor
current

option B

Calibrate your current LLM

Same model, with a confidence threshold calibrated on your cases.

The right choice if your LLM rarely gets it wrong. Only doubtful decisions go to review.

effort
a few days
LLM bill
unchanged
reviews
lower
vendor
current

option C

Move to Jev

Routing by Jev, LLM fallback, decision log.

The right choice if your LLM bill is heavy, and Jev holds up on your data.

effort
migration
LLM bill
much lower
reviews
lower
vendor
+ TypeSafe AI

The verdict can be not to migrate.

04 · Architecture

Every decision follows a branch. Every branch leaves a trace.

05 · Pricing

One entry price. Everything else is quoted.

Pre-qualification

free

30 minutes. A written go or no-go within 48 hours.

Flash Audit

€3,500

Your most expensive decision, for example triaging your tickets, measured on your data. Delivered 2 weeks after your DPO signs off in writing.

Pioneer price €2,500 for the first 3 clients, until 31 December 2026. In return, a case study is published without your name. The migration credit is then €500.

Migration and run

quoted

Priced from the audit results. LLM fallback, decision log, human review and kill switch always included.

  • Every migration starts with an audit. The Flash Audit gives a €1,000 credit on a migration signed within 60 days.
  • Your Jev and LLM usage stays on your own accounts.
  • For businesses only. Prices exclude VAT. French clients: VAT not applicable, article 293 B of the French General Tax Code. EU business clients: reverse charge, you account for VAT in your country.

06 · Calculator

Estimate your savings. Then have them measured.

Start from the flow where your LLM sorts or routes the most: tickets, emails, listings or invoices.

Tickets, emails, listings or invoices that go through your AI triage.

Every LLM call that decides counts: team, urgency, sentiment, then every new pass, for example at each follow-up. 5,000 items judged 10 times make 50,000 calls.

Depending on your quality bar: high end for high-stakes decisions, fast for large volumes with low stakes.

Advanced settings: your real numbers

Leave empty to use the estimate. Enter your bill if you use prompt caching or the Batch API.

System prompt plus the text to classify. A 300-token ticket makes about 1,000 tokens with its prompt. 50 output tokens are added.

90%

The rest falls back to your LLM, which is still billed.

Estimated API savings, per month

€4,411

that is 88% off your routing bill

decision calls per month
2,000,000
current LLM bill
€5,060
bill after migration, Jev included
€594
API savings
+ €4,466
EU hosting
− €55

What about quality? On TypeSafe’s published evals, Jev matches Claude Sonnet 5 on average and beats Opus 5 on customer service: 76.0% against 72.4%. It lags on invoices. Against a smaller model, it can therefore improve quality, depending on the job. The audit measures it on your cases.

Have your numbers measured: book 30 min

Estimate based on assumptions. Bill estimated at public prices as of 23/09/2026, per million input and output tokens: Opus 5.5 at $4 and $20, Sonnet 5 at $2 and $10, Gemini 3.8 Flash at $0.75 and $3.75 until 31/12/2026, then $1.50 and $7.50. Jev at $0.042 per million input tokens. $1.1463 to €1. Savings before our fees, which are quoted. Review time is not counted. EU hosting at €55 a month. The human gain is freed time: it only becomes money if those hours are reassigned. Your audit measures your real numbers.

07 · Examples

Three volumes. Three verdicts.

Marketplace

24 million listings a month to moderate, one call each, on Gemini 3.8 Flash.

Routing bill per month
€19,628 → €2,842
Savings per month
€16,731

Even against a fast model, volume makes the savings. The audit also compares the quality of both on your listings.

Customer service

200,000 conversations a month, judged 10 times each, on Claude Sonnet 5.

Routing bill per month
€5,060 → €594
Savings per month
€4,411

Team, urgency and sentiment, at every message: calls add up fast. And on this job, TypeSafe measures Jev above Opus 5.

Invoices

90,000 supplier invoices a month at a large group, 3 checks each, on Claude Sonnet 5.

Routing bill per month
€1,295 → €154
Savings per month
€1,086

Not enough calls against Sonnet. And on invoices, TypeSafe measures Jev behind: 61.8% against 72.9% for Sonnet 5. Here, the audit will likely conclude not to migrate.

Savings before our fees, EU hosting deducted. 1,000 to 2,500 input tokens per call depending on the text, Jev decides 90% of cases. Public orders of magnitude: 800,000 listings published a day on Leboncoin, 2.3 million conversations a month for Klarna’s assistant in 2024. Your audit measures your numbers.

08 · Eligibility

Who it is for, and who it is not for.

Who it is for

  • A large volume to triage: mass B2C customer service, a marketplace, or a platform that processes items for its clients.
  • At least €1,700 a month of LLM spend on this triage. Depending on your model, that means 0.3 to 1.8 million decision calls a month, for example 200,000 tickets judged 10 times.
  • An LLM in production to classify, route or triage text, in your code or your automation tools.
  • Jev access in your own name. TypeSafe AI has temporarily paused sign-ups. We check your access during pre-qualification.
  • Around 400 past examples, and your DPO’s written approval before any data is sent.

Who it is not for

  • Fewer than a few hundred thousand decision calls a month, such as a small internal invoice flow. The audit may conclude that calibrating your LLM is enough.
  • Data that cannot leave the EU, even pseudonymised.
  • Health data or other special-category data under GDPR Article 9.
  • High-risk cases under the AI Act, such as hiring or credit.
  • Decisions mainly about amounts or dates.
  • Requests for a “GDPR compliant” label.

09 · Guarantees

What protects you.

  1. A refusal before any invoice if you are not eligible.
  2. The audit balance is due only when a report measured on your data is delivered.
  3. No commission from TypeSafe AI. A no-go is a delivered result.
  4. 20% of the migration is due only if the acceptance criteria are met. Otherwise, we fix it for 30 days at our cost. Excluding Jev outages or changes, and scope changes on your side.
  5. Kill switch in your hands, original workflows kept for 90 days, code in your repository.

10 · Exclusions

What is not included.

  • Legal advice.
  • Jev and LLM usage.
  • EU hosting, €30 to €80 a month.
  • Continuity of the Jev service, which depends on TypeSafe AI.
  • 24/7 on-call.

11 · FAQ

Frequently asked questions.

What is Jev?

Jev is a decision model from TypeSafe AI, launched in September 2026. It does not generate text. It returns a typed choice and a probability that your code can use directly. TypeSafe AI says the probability is calibrated. The audit checks it on your data. To try it, follow our Jev in n8n guide.

Are you a TypeSafe AI partner?

No. Jev Migration is an independent service. We take no commission from TypeSafe AI, and your Jev access is in your own name.

Jev is brand new. What if we depend on it?

On error, timeout or rate limit, the decision falls back to the LLM automatically. The kill switch sends all traffic back to the LLM in under 5 minutes, without us. TypeSafe AI may retire a Jev version, in which case you must switch versions. Jev sign-ups are also temporarily paused, so we check your access before any commitment.

Our data goes to the United States. What about GDPR?

Yes, it goes to the United States. TypeSafe AI does not appear on the Data Privacy Framework list, so the transfer relies on the standard contractual clauses in its DPA. Data is pseudonymised in the EU before sending, which does not make it anonymous. Your DPO gets a data-flow sheet with the Flash Audit, and a full technical file with the migration. We sign an Article 28 processing agreement. We promise neither compliance nor a risk-free transfer.

Does it work in languages other than English?

Jev targets English first. So we calibrate per language, on your data, and every threshold ships with its measured error rate.

Is Jev as good as a large model?

It depends on the job. On TypeSafe’s published evals, Jev agrees with the reference 67.8% of the time on average, as much as Claude Sonnet 5, against 73.1% for Claude Opus 5. On customer service, it beats Opus 5: 76.0% against 72.4%. On invoice processing, it lags: 61.8% against 72.9% for Sonnet 5. The reference is the average of two frontier models, on workflows built by TypeSafe: this is agreement, not ground truth. Jev makes no format errors, by design, but it can pick the wrong option confidently. The audit measures the gap on your cases.

What if Jev changes its prices?

You pay Jev directly. Even at 10 times the price, the customer-service example still saves €3,619 a month.

You work alone. What if you are unavailable?

The system runs without us. The code, configuration and runbook are in your repository. The kill switch is in your teams’ hands. On larger projects, a senior freelancer is named in the contract.

We can do this in-house.

Yes. The open-source building blocks exist. The audit may even conclude that calibrating your current LLM is enough.

4 audits per quarter.

Book 30 minutes.

Have four numbers ready:

  1. Items processed per month
  2. LLM in use
  3. API invoices for the last 3 months
  4. Review hours

Étienne Lescot

Independent developer, n8n specialist, based in France.

Book the free pre-qualification

Or by email: contact@jevmigration.com

Independent service, not affiliated with TypeSafe AI.