Training Simulator is Cresta’s answer to a problem every contact centre recognises: agents learn on live calls, in front of real customers, and the mistakes they make while learning are mistakes customers actually experience. Launched on 9 July 2026, the product drops agents into conversations with simulated customers that argue back, get frustrated, and escalate — customers generated, Cresta says, from the company’s own recorded conversations rather than from a script somebody wrote.
The pitch is genuinely good. It is also, on the evidence Cresta published, entirely unquantified. We counted every numeral in the two documents Cresta released on launch day — a press release and a blog post signed by CEO Ping Wu — and across 1,773 words of body copy the total is zero. Not one percentage, not one duration, not one sample size, not one customer result. The only digits on either page belong to the dateline.
That absence is worth examining, because Cresta clearly knows how to publish figures. Forty-two days earlier the same company launched Synthetic Customers, the capability that turns conversation data into customer personas, and that announcement’s blog post carries twelve numerals including seven percentages. Same company, same CEO, same underlying idea, two very different disclosure standards. This article counts what was published, quotes the places where two Cresta documents describe the same feature differently, and sets out what a buyer should ask before signing.
Table of contents
- What Cresta’s Training Simulator Actually Does
- Counting Every Figure in the Training Simulator Launch
- The Realism Test Training Simulator Never Cites
- Where the Training Simulator Story Changes Between Documents
- The Word Cresta Never Uses About Training Simulator
- Training Simulator and Synthetic Customers Never Meet
- What the Training Simulator Launch Does Not Disclose
- How to Evaluate Training Simulator Without Vendor Numbers
- What Training Simulator Means for Contact Centre Buyers
- Frequently Asked Questions About Training Simulator
- References and Further Reading
What Cresta's Training Simulator Actually Does
Before auditing the numbers, it is worth stating the product fairly, because the concept is sound and the criticism here is about evidence rather than about the idea.
Two documents, one launch day
Cresta announced the product on 9 July 2026 through two pieces of copy published the same day. The first is a press release titled “Cresta Launches Training Simulator, Using AI Agents to Provide Dynamic Training to Human Agents”. The second is a blog post under Ping Wu’s byline titled “Cresta Launches Training Simulator, Bringing AI Agent Technology to Human Agent Training”. A product page and a short video page carry the same claims in compressed form.
How a simulated customer gets built
The press release describes the mechanism plainly: Training Simulator “generates simulated customers based on a company’s actual customer conversations”, and those simulated customers “respond dynamically in real time to exactly what the human agent says, push back, show emotion, and escalate when answers fall short”. The worked example given is a cancellation save, where the simulated customer “believably objects to retention offers and escalates when answers feel generic”.
Cresta’s wider platform is a conversation intelligence suite grounded in natural language processing, and the press release is careful to say the simulated customers are “powered by the same AI Agent technology Cresta uses in customer-facing deployments” — the production stack, not a separate training-only model.
Who grades the practice conversation
Grading is the part of Training Simulator that is most obviously useful. Cresta says the same AI-powered quality criteria that evaluate live conversations also grade the simulated ones, so an agent who passes a simulation has demonstrated behaviours that quality management will look for on the floor. There is, in Cresta’s words, “no separate rubric to build, maintain, or explain”. Scenarios can be tagged to specific behaviours and assigned by managers, with results surfacing in the coaching plan they already use.
The three claims worth testing
Strip the marketing away and Training Simulator makes three testable claims. The simulated customers are derived from real conversations. They behave realistically enough for practice to transfer. And the grading matches live quality assessment. Each of those is measurable. None of them was measured in public.
Counting Every Figure in the Training Simulator Launch
The audit below is mechanical and repeatable. Anyone can run it against the same pages and get the same answer.
The counting method
We took the body copy of each document — from the headline to the closing call to action, excluding site navigation, footers and the related-articles cards that appear on every Cresta page — and counted every numeral, including years and dates. Navigation had to be excluded because Cresta’s global menu contains product names and links on every page, which would otherwise contaminate the count on all four documents equally.
What the count returns
The Training Simulator press release runs 620 words and contains two numerals: the “9” and the “2026” of its dateline. The Training Simulator blog post runs 1,153 words and contains no numerals at all. Combined, that is 1,773 words of launch copy carrying zero substantive figures.
| Cresta launch document | Date | Body words | Numerals | Percentages |
|---|---|---|---|---|
| Training Simulator press release | 9 Jul 2026 | 620 | 2 (dateline) | 0 |
| Training Simulator blog post | 9 Jul 2026 | 1,153 | 0 | 0 |
| Synthetic Customers press release | 28 May 2026 | 534 | 2 (dateline) | 0 |
| Synthetic Customers blog post | 28 May 2026 | 1,982 | 12 | 7 |
The sibling launch that did carry numbers
The contrast is the point. Cresta’s Synthetic Customers blog post, published 42 days before Training Simulator, is 1,982 words long and carries twelve numerals. Two of those are years, so ten are substantive figures — roughly one every 198 words. The Training Simulator blog post is 1,153 words with none.
The same corpus, counted by numeral
Plot the numerals instead of the words and the shape inverts. The longest Training Simulator document is the emptiest.
What the sibling actually disclosed
The twelve numerals in the Synthetic Customers blog break down into seven percentages, three plain counts and two years. The percentages are real disclosures: personas covering 100% of an account support queue’s volume, a worked example where “these 12 personas represent ~82% of conversation volume”, a note that a typical customer survey reaches 3–5% of customers, plus cited third-party figures of 32% and 70%.
The Realism Test Training Simulator Never Cites
The most interesting gap is not that Cresta published no numbers. It is that Cresta describes a specific test for exactly the claim in question, and then never reports its result.
A methodology with no result
Both Synthetic Customers documents contain a version of the same sentence. The press release: “Cresta measures simulation realism using a blind evaluation methodology testing whether humans and AI can distinguish synthetic conversations from real ones.” The blog: “Cresta validates simulation realism using a blind evaluation methodology, testing whether humans and AI can distinguish synthetic conversations from real ones.”
The result is missing in both
Read those sentences again and notice what is not in them. There is no pass rate, no sample size, no description of who the human judges were, no confidence interval, no baseline against a generic model, and no link to a methods write-up. A blind evaluation produces a number by construction — that is the entire point of running one. Cresta names the instrument and withholds the reading.
Neither Training Simulator document mentions it
The realism sentence appears twice, and both appearances are in Synthetic Customers material. Neither the Training Simulator press release nor the Training Simulator blog post mentions blind evaluation at all. So the product whose core promise is that practice “feels like the real thing” is the one that never references the test for whether it does.
Why the missing number matters here
Realism is not a nice-to-have for this product; it is the mechanism. If a simulated customer is more agreeable than a real one, agents rehearse against an easier opponent and arrive on the floor over-confident. Cresta’s own Synthetic Customers blog makes precisely this argument against generic models, citing an arXiv preprint finding that large language model simulations “converge toward an overly cooperative, generic user that flattens individual differences”. The company identified the failure mode, built a test for it, and published no score.
Where the Training Simulator Story Changes Between Documents
Two Cresta documents released on the same day describe the same capabilities in materially different words. These are not paraphrases; the claims differ in strength.
Minutes, or an afternoon
On how fast a new scenario can be built, the press release says a training lead “can create a new scenario for an upcoming product launch in minutes, rather than waiting weeks for a vendor to build it”. The blog post, published the same day, says users “can stand up a new scenario for an upcoming product launch or policy change in an afternoon”. An afternoon is roughly 240 minutes. Both statements cannot be the operative one.
“Are based on” versus “can be built with”
On provenance, the press release is definite: simulated customers “are based on real conversations”. The blog is conditional: simulated customers “can be built with insights from real conversations surfaced by Cresta Insights”. The first says every simulated customer comes from your data. The second says some can, from insights derived from your data. For a product sold on the strength of being grounded in your own calls, that is the load-bearing sentence.
Prompts or conversations
The press release also reveals that there are two build paths, joined by an “or”: leaders “can build scenarios using prompts or auto-generate them from real conversations”. The blog describes only the prompt path, saying leaders “can create new scenarios from a prompt, informed by insights from real-world conversations”. A prompt informed by insights is a different artefact from a scenario generated from a transcript, and only one of them is what the headline claim implies.
| Claim | Press release, 9 Jul | Blog post, 9 Jul |
|---|---|---|
| Scenario build time | “in minutes” | “in an afternoon” |
| Provenance of simulated customers | “are based on real conversations” | “can be built with insights from real conversations” |
| Build paths described | Prompts or auto-generate from conversations | Prompt only, “informed by insights” |
| Grading authority | “the same AI-powered quality criteria” | “the same AI-powered quality criteria” |
| Data source named | None named | “Cresta Insights” |
The Word Cresta Never Uses About Training Simulator
The popular shorthand for this product — including in the framing that prompted this article — is that it builds AI customers from transcripts. Cresta does not use that word.
Zero transcripts in the corpus
Across all four launch documents plus both product pages, the string “transcript” appears zero times in the body copy. Cresta consistently says “conversations”, “conversation data”, or “real customer conversations”. The one place the word does surface on the site is the transcript control on a video player, which is a user-interface element rather than a claim.
Conversations, insights, and Opera rules
The vocabulary also drifts on grading. The press release credits “the same AI-powered quality criteria”. The blog credits “Cresta Insights” for the conversation grounding. The video page says simulations are “graded by the same Opera rules that grade live calls”, introducing a third named component. Three descriptions, one feature, and no architecture diagram reconciling them.
Why the vocabulary matters
“Transcript” implies a specific, auditable artefact: this call, that customer, those words. “Conversation data” and “insights derived from conversations” are broader and can accommodate aggregate statistics, extracted behaviour patterns, or a prompt written by a human who read a summary. The looser term is not evidence of anything improper. It does mean a buyer should not assume the tighter one.
Training Simulator and Synthetic Customers Never Meet
The strangest editorial fact in this launch is how thoroughly the two related products avoid each other.
42 days between two launches
Cresta launched Synthetic Customers on 28 May 2026 and Training Simulator on 9 July 2026 — 42 days apart. The Synthetic Customers announcement explicitly lists “human agent training” as one of four applications, saying the capability enables “dynamic role play training to onboard agents faster”. That is Training Simulator’s entire job description, published six weeks early.
Nav-menu mentions only
Yet in the Training Simulator press release, blog post and product page, the phrase “Synthetic Customers” appears exactly once each — every time inside the site’s global navigation menu, which renders identically on every page of the domain. In the body copy of all three Training Simulator documents, the count is zero. The blog credits Cresta Insights instead.
The four personas that did get named
The contrast in specificity is stark. The Synthetic Customers blog names four personas from a worked example built on 90 days of closed conversations — the Persistent Challenger, the Cautious Verifier, the Efficient Transactor and the Anxious First-Timer — and describes how each behaves. It states that those four represent 100% of that queue’s real volume. Training Simulator’s launch names no personas and quantifies no coverage.
What a reader is left to infer
None of this is evidence of a problem with the product. Marketing teams brief separately, and product boundaries shift. But a reader trying to work out whether Training Simulator’s simulated customers are Synthetic Customers under another name cannot answer that question from Cresta’s published material, and the answer changes what the buyer is purchasing.
What the Training Simulator Launch Does Not Disclose
Pulling the gaps together gives a checklist that is easier to act on than to argue with.
No pricing, no availability
Neither Training Simulator document mentions price, list rate, packaging, or how the product is metered. Neither states whether it is generally available, in beta, in limited release, or gated behind an existing platform tier. The strings “pricing”, “availability”, “generally available”, “beta” and “early access” appear zero times across both. Every call to action is a demo request.
No sample size, no baseline
There is no named launch customer, no pilot result, no ramp-time reduction, no attrition figure, and no before-and-after on quality scores — despite ramp time, attrition and quality being the three outcomes the copy repeatedly promises to improve.
The disclosure table
| What a buyer would want | Disclosed for Training Simulator? | Disclosed for Synthetic Customers? |
|---|---|---|
| Realism score from the blind evaluation | No — not mentioned | Methodology named, result withheld |
| Persona coverage of real volume | No | Yes — 100% and ~82% examples |
| Conversation history required | No | Yes — 90-day worked example |
| Ramp-time or attrition result | No | No |
| Pricing or packaging | No | No |
| Availability stage | No | No |
How to Evaluate Training Simulator Without Vendor Numbers
The absence of published figures is not a reason to dismiss the product. It is a reason to generate your own figures during evaluation, which is better evidence anyway.
Questions to put to the vendor
Ask for the blind evaluation result and its sample size, since the methodology is already public. Ask whether your simulated customers are generated from your transcripts or from prompts informed by aggregate insights, and ask to see the source conversations behind a given scenario. Ask whether Training Simulator uses Synthetic Customers underneath, and whether that is a separate line item.
Metrics to demand in a pilot
Insist that a Training Simulator pilot reports the same three numbers the marketing implies: days to proficiency for a cohort against a matched control, ninety-day attrition for both groups, and correlation between simulation grades and live quality scores. That last one is the cheapest and the most revealing, because Cresta’s own design claim is that the two use identical criteria.
What good evidence looks like
A simulation is working when an agent’s simulated grade predicts their live grade, and when practice on a flagged behaviour moves that behaviour on real calls. Both are measurable inside a single quarter with data the platform already collects. Neither requires trusting a vendor figure.
Where an internal comparison helps
Run the same cohort against your existing role-play programme for one intake. If Training Simulator only matches supervisor-led role-play, the case rests on scale and cost rather than on quality — which may still be a strong case, since supervisor calendars are the constraint the product is explicitly designed to remove.
What Training Simulator Means for Contact Centre Buyers
The claim is plausible and unmeasured
Nothing here suggests Training Simulator does not work. Cresta runs production voice AI for large enterprises, the underlying technique is well understood, and the coaching integration is a genuinely smart piece of design. The criticism is narrow and specific: the company published 1,773 words of launch copy containing no measurement of a claim that it has already built an instrument to measure.
Where the risk actually sits
The risk is not that the simulated customers are bad. It is that “realistic” is doing procurement work without a number behind it, and that the provenance language weakens between the press release and the blog. Buyers evaluating autonomous AI agents and agent-facing tooling should treat unquantified realism claims the way they treat unquantified accuracy claims — as a hypothesis to test in a pilot.
A reasonable position to take
Shortlist it. The problem it targets is real, the coaching loop is well designed, and the alternative — scarce role-play gated behind a supervisor’s calendar — is genuinely worse. Then make the pilot produce the numbers the launch did not, and make the contract depend on them.
What would change this assessment
One published blind-evaluation result with a sample size would change it substantially. So would a single named customer with a ramp-time figure. Both are within Cresta’s gift, and the Synthetic Customers announcement proves the company is willing to publish numbers when it has them.
Frequently Asked Questions About Training Simulator
What is Cresta Training Simulator?
It is an agentic training product launched on 9 July 2026 that lets contact centre agents practise against AI-simulated customers which respond dynamically, show emotion, push back on weak answers, and escalate — replacing scripted role-play with adaptive conversation.
Does Training Simulator really build customers from transcripts?
Cresta says simulated customers “are based on real conversations” in its press release, and “can be built with insights from real conversations” in its blog post. The word “transcript” does not appear in either. There are two documented build paths: prompts, or auto-generation from real conversations.
How realistic are the simulated customers?
Unknown from published material. Cresta states it runs a blind evaluation testing whether humans and AI can distinguish synthetic conversations from real ones, but has published no result, sample size or baseline for that test in any of the four launch documents.
How is Training Simulator different from Synthetic Customers?
Cresta has not said. Synthetic Customers launched 42 days earlier and lists human agent training among its applications, but the Training Simulator launch documents never mention it outside the site navigation menu.
What does Training Simulator cost?
Cresta has published no pricing, packaging or availability information for the product. Every call to action on the launch material is a demo request.
Is the missing data a red flag?
Not on its own — unquantified launch copy is normal in enterprise software. It is notable here only because Cresta describes a specific realism test it has already run, and because its own sibling launch six weeks earlier disclosed ten substantive figures in a comparable blog post.
References and Further Reading
Cresta — Training Simulator product page
Cresta — Cresta Launches Training Simulator, Bringing AI Agent Technology to Human Agent Training
Cresta — Introducing Cresta Training Simulator (video)
Cresta — Synthetic Customers product page
Cresta — Introducing Synthetic Customers: A Living Model of Your Customer Base
LangChain — State of Agent Engineering
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.