Poised is shutting down. What Deepgram's move toward Saga tells us about speech products

The tempting story is that “AI speech coaching failed.”

We do not have evidence for that story.

There is no public Poised post-mortem that gives a definitive shutdown cause. Deepgram has not published Poised revenue, retention, acquisition cost, usage frequency or unit economics. So the honest answer to “why did Poised shut down?” has two parts:

1. the shutdown and the strategic product shift are observable;
2. the exact commercial cause is not public.

The shift is still useful to study because products that all “use voice” can be solving very different jobs.

Poised, Saga and Ptichi use voice for three different outcomes

Poised publicly positioned itself as an AI communication coach. It sat beside calls, gave real-time feedback, helped with meeting notes and provided post-meeting analysis. Poised now displays the shutdown notice on its site.

Saga is built around another job. Deepgram describes it as Mac dictation and a voice interface: speak, get cleaned-up text, and place that result into the app where you are already working. That is how Saga presents itself today.

Ptichi is being built around a third job:

I want to speak better when the software is no longer watching me.

That difference is more important than which product has the nicer waveform.

Product object	Where the value appears	What should remain
Poised	During or immediately after communication	Guidance about the current interaction
Saga	While doing computer work	Useful text, command or action
Ptichi	During deliberate practice	A skill that transfers to the next situation

The last row is our product thesis, not a validated Ptichi outcome. The desktop app is still in development and public installers are not available. The current roadmap keeps product truth separate from ambition.

What actually happened after Deepgram acquired Poised

In June 2024, Deepgram announced that it had acquired Poised's assets, proprietary technology and integrations, together with key engineering talent. At that point Poised was still described as real-time communication support for online meetings. Deepgram's acquisition announcement.

Shortcut came later. Its public positioning already moved the center of gravity from “watch how I speak in a meeting” toward work at the speed of voice: turn spoken thoughts into polished writing, create documents, convert scattered ideas into plans and get quick answers. Deepgram explicitly connected Shortcut with its earlier Poised acquisition. Introducing Shortcut.

Saga makes the change even clearer. The current product page emphasizes real-time dictation, AI editing, any text field and a personal dictionary. This is voice-as-input rather than a speech-training system.

It is important not to rewrite the history more strongly than the sources do. Deepgram does not describe Saga as a literal rename of Poised. A more accurate description is a strategic product shift after the Poised acquisition.

Why the Saga direction makes product sense for Deepgram

From here, we are interpreting the public evidence. These are plausible product reasons, not disclosed shutdown causes.

1. Voice becomes an interface instead of an object being judged

With Poised, the person talks and the product comments on how they are talking.

With Saga, the person talks in order to do work: draft an email, leave a comment, create a document or formulate a task.

The second value is visible immediately. An empty field becomes usable text. You do not need a learning study to establish that the output exists.

A training product has a harder burden. A better second attempt does not prove that the speaker will perform differently tomorrow.

2. Saga maps directly onto Deepgram's core technical advantage

Deepgram builds speech recognition and voice infrastructure. Saga turns that capability directly into a user action.

Speech training needs more than accurate transcription. It has to answer uncomfortable questions:

• What should this person change next?
• Can the recording support that inference?
• Which cue helps rather than distracts?
• Does the improvement survive without the cue?
• Does it transfer to different wording and a different context?

That is not only speech AI. It is learning design.

3. Skill products have to prove transfer

A dictation loop can be short:

speak → receive text → use text.

A learning loop is longer:

speak → observe → understand → change one thing → retry → transfer → retain → use in a real situation.

If the loop stops at an attractive dashboard, the product may have measured speech without teaching anything.

That is the main warning we take from this market signal.

The problem Ptichi is choosing to solve

We do not want to build another screen that tells someone they used seven filler words, spoke at 162 words per minute and were “83% confident.”

Those numbers may sometimes be useful observations. The trouble begins when the number becomes an answer to a question it does not measure.

Ptichi is built around a smaller practice loop:

Take A → listen → choose one change → Take B → compare → use new wording → retest later.

The Ptichi technology model is being designed around that loop.

That creates several product jobs.

Find the next useful practice target, not every possible flaw

Most people do not need twelve simultaneous corrections.

They need to know: what am I changing now?

Not “be more confident,” but make the boundary between the decision and its condition audible. Not “speak slower,” but stop losing the last important words of a long thought.

One bounded target makes two attempts meaningfully comparable.

Separate an observation from a judgment about the person

Pace, pauses, energy, pitch and transcripts can produce useful signals. They should not automatically become conclusions about competence, personality or charisma.

Ptichi should be able to say:

this pause became shorter;

without pretending that it means:

you are now 14% more confident.

Teach without permanent dependence on the cue

This matters directly for Ptichi's real-time research.

We are not rejecting real-time feedback. We are exploring a low-latency layer for pauses, pace, energy, F0 and other observable signals.

But every live cue needs an exit path.

If someone controls pace only while watching a pace meter, we may have improved their ability to use a meter. The speech skill is still uncertain.

A stronger Ptichi progression is:

live cue → less cue → no cue → new wording → later retest.

Real-time becomes a training instrument rather than the product's final state.

Test transfer, not only a polished second take

A person can repeat one sentence until it sounds excellent.

Then the question changes and the old pattern returns.

Ptichi therefore needs a transfer attempt with different wording. If the change does not survive a new sentence, we should not call it learned yet.

Make practice concrete

An interview answer, project explanation, technical presentation and second-language assessment do not create the same listener task.

A useful programme should not merely contain content. It should organize repeated practice around what the listener needs to recover.

The warning for Ptichi

It would be easy to look at Poised and conclude that Ptichi simply needs more real-time metrics.

We take the opposite lesson.

If Ptichi becomes a permanent instrument panel beside the speaker, we risk building something a person uses instead of building the skill.

Warning signs would include:

• the user optimizes a score rather than a listener task;
• a live widget helps only while it is visible;
• the improvement disappears on new wording;
• the system produces many comments but cannot choose the next practice target;
• an acoustic change is labeled better communication without an outcome check;
• an English intonation pattern is copied mechanically into Russian or German;
• progress history becomes a collection of charts without a reproducible behavioral change.

Analytics can still be useful. The rule is that analytics must serve practice.

Our bet: a learning product should become less necessary

That is an awkward sentence for SaaS, but a useful one for skill training.

Early practice may need a lot of support: a prepared example, marked text, recording, a visual signal and one specific cue.

Later, some of that support should disappear.

If Ptichi works, the learner should increasingly be able to:

• notice the problem without the dashboard;
• choose the right technique;
• repair it with less feedback;
• transfer the technique to new wording;
• use it in a real conversation.

The goal is not maximum time inside a voice dashboard. The goal is a reproducible change.

That is harder to measure and probably harder to monetize than a simple usage loop. It is also the reason Ptichi can become a distinct product rather than a smaller meeting coach.

Try the distinction with your normal recorder

Take one real, non-confidential 30–45 second answer. For example:

Tell me about a project you are proud of.

Record Take A.

Do not search for ten flaws. Pick one listener question:

Did I understand what your personal contribution was?

Change one thing only: state your role earlier, place a pause before the result, or remove one long detour.

Record Take B.

Then do the important part: put the original wording away and answer a similar question in new words.

If the change survives, that is more interesting than a polished second take.

If it disappears, that is useful too. You found the boundary of the skill.

For one small practice target, try Two thoughts. One useful pause..

What would change our mind

This is not a victory lap around a closing competitor.

Our market interpretation would weaken if Deepgram published a post-mortem showing that Poised was shut down for reasons unrelated to product strategy or user value.

Our Ptichi thesis is testable too.

If persistent live feedback reliably transfers to unaided speech, we should not remove it for ideological reasons. If users consistently get more value from real-meeting assistance than from deliberate practice, we should revisit the boundary of the product.

For now, the useful principle is:

The cue can be temporary. The skill should survive it.

What we take from the Poised story

Poised shutting down does not prove that people do not need help with speech.

Deepgram's move through Shortcut toward Saga shows how large the neighboring opportunity is: use voice as an interface to software. That is a strong and understandable product job.

Ptichi is deliberately choosing another one.

Saga helps turn speech into output on a screen. Poised helped during the communication itself. Ptichi is being built to help a person change how they speak — and carry that change away from the product.

If we cannot demonstrate transfer, fading support and useful change beyond a rehearsed script, that distinction will remain positioning copy.

That is exactly the risk we need to test first.