THE SHORT VERSION
Fireflies AI verdict: useful notes, with details to check
In our first Fireflies AI test, the notes preserved the final launch date, a corrected copy deadline, and an explicitly unapproved advertising budget. All eight expected tasks appeared. Yet two tasks had an incorrect or contradictory owner, and two deadlines were missing.
Our provisional take: this output could help someone reconstruct a call, but we would check it against the conversation before assigning work or sending a client follow-up.
Changed dates, budget boundaries, and unresolved questions.
Task ownership, completeness of deadlines, and source timestamps.
First-run scope: one 3:38 synthetic recording, one voice reading two roles, one upload. The separate second recording uses two synthetic voices and is reported below. Neither is a general accuracy benchmark.
01 / THE SETUP
How we tested Fireflies AI with a recorded client call
We wrote a fictional conversation between Emma, a consultant, and Liam, her client. Before using Fireflies, we recorded the expected answers: eight tasks, three decisions, two open questions, and a set of commitments the tool must not invent.
The script moves a launch from October 1 to October 8, replaces a September 14 copy deadline with September 16, and explicitly rejects a possible $500 advertising spend. Liam owns the client feedback and subscription-plan check. Olivia can advise, but owns neither the feedback nor a calendar invitation.
- Input
- WAV upload · 3 min 38 sec
- Voice
- One synthetic voice, two spoken roles
- Settings
- English (Global) · General Summary
- Account
- Free account · trial declined
The same synthetic voice reads both roles and their names. This makes it a limited test of content preservation. It cannot tell us how Fireflies separates two real speakers, handles background noise, or performs across a working week.
Read the exact script used in our first test
For the first test, we uploaded a WAV recording of the script below and selected English (Global) and General Summary. We did not enter a custom summary prompt. One synthetic voice read both roles, including their names.
All people and commitments are fictional. September 7 is the date spoken in the scenario; we ran the test on September 6, 2026. The answer key was kept outside Fireflies.
Emma: This is a fictional client meeting for testing AI meeting notes. I am Emma, the consultant, and Liam is the client at Northstar Studio. Today is September seventh, twenty twenty-six. We are discussing the October launch of your new booking website.
Liam: Correct. The project budget is three thousand dollars. I originally suggested launching on October first, but our photographs will not be ready. Let us move the launch to October eighth, twenty twenty-six. Please treat October first as cancelled, not as an extra deadline.
Emma: Agreed. The confirmed launch date is October eighth. I will send the revised homepage wireframe by September ninth. You will review that wireframe and send one consolidated set of comments by September eleventh. Is that right?
Liam: Yes, I will collect the feedback myself. Please do not assign that task to Olivia. She can advise me, but I remain responsible. I will also upload our logo and the final brand colours by September tenth.
Emma: Great. We discussed a paid advertising campaign last week. Are we including that in this project?
Liam: No. Advertising is outside the three-thousand-dollar budget. I mentioned five hundred dollars for ads as a possibility, but I have not approved that amount or any campaign. Do not order ads or treat the extra money as committed.
Emma: Understood. I will check whether the booking tool can connect to your existing calendar by September twelfth. That is a feasibility check, not a promise that the integration is supported. If it is not supported, we will discuss alternatives before selecting one.
Liam: That works. I also want a Spanish version eventually, but we have not decided whether to include it. There is no owner or deadline for translation yet. Keep it as an open question.
Emma: One more correction. Earlier I said I would send the copy draft on September fourteenth. Please replace that date with September sixteenth. I am still the owner of the copy draft, and September sixteenth is the only agreed deadline for it.
Liam: Agreed. I will confirm the booking tool subscription plan by September fifteenth. I will send you the plan name and its limits. I am not buying an annual subscription during this meeting.
Emma: For the next check-in, I suggest September seventeenth at ten in the morning, London time. Does that suit you?
Liam: Yes, that is confirmed. Emma, please send the calendar invitation by September eighth. Invite me only; Olivia does not need an invitation. Please keep the internal discussion about possible advertising spending out of the short client-facing follow-up email. It can stay in our internal notes.
Emma: I will send that invitation by September eighth. In the follow-up email, I will include the confirmed October eighth launch, the three-thousand-dollar project budget, our agreed tasks and deadlines, and the open question about Spanish. I will not present the calendar connection as already working or the advertising budget as approved.
Liam: That is everything. Thank you.
02 / WHAT CAME BACK
Meeting notes and action items: our test results
*A task counts as fully correct only when its owner, agreed deadline, and important constraints are present without contradictions across Notes and Action items. This definition was operationalized during assessment. It is not a product-wide score.
The standalone action list contained no calendar deadlines. The times in brackets are recording references, not due dates. Five of the seven agreed deadlines appeared elsewhere in Notes; September 10 for the brand assets and September 15 for the subscription check were absent.
| Expected task | What Fireflies returned | Result |
|---|---|---|
| Emma · Wireframe · Sep 9 | Owner in action list; deadline in Notes. | Preserved |
| Liam · Feedback · Sep 11 | Liam in the action list, but Emma in the ownership notes. | Owner conflict |
| Liam · Brand assets · Sep 10 | Correct task and owner; deadline missing. | Incomplete |
| Emma · Integration check · Sep 12 | Task, owner, deadline, and uncertainty retained. | Preserved |
| Emma · Copy draft · Sep 16 | Correct owner and replacement deadline. | Preserved |
| Liam · Plan check · Sep 15 | Assigned to Emma; deadline missing. | Owner + date |
| Emma · Invite · Sep 8 | Owner, deadline, and Liam-only restriction retained. | Preserved |
| Emma · Follow-up · No due date | Task and content restrictions retained; no deadline invented. | Preserved |
The five fully correct tasks come from reading the notes and action list together. They should not be interpreted as an accuracy percentage for the transcription or for Fireflies generally.
03 / LOOK AT THE OUTPUT
Original screenshots: two task ownership errors
The notes say Emma is responsible for collecting feedback. In the script, Liam explicitly takes that responsibility. The action list later gives feedback to Liam correctly, so the output contradicts itself.
The subscription-plan check is also listed under Emma. It belongs to Liam, who promises to send the plan name and its limits by September 15. That deadline is missing from both the notes and action list.
The transcript places the whole recording under “Speaker 1,” consistent with our single-voice input. Several spoken role names are missing from the transcript. That may have made ownership harder to recover; we cannot infer how the same exchange would behave with two real voices.
What Fireflies kept correctly
The October 8 launch replaces October 1, the copy deadline is September 16, and the $3,000 budget excludes advertising. The notes explicitly say the possible $500 ad spend is not approved. Spanish remains undecided, and the calendar connection remains a feasibility check rather than a working integration.
Some source references appear early: the copy correction is marked 01:30 in Notes, while it appears in the transcript block starting at 01:58. We have not aligned the audio word by word, so we treat this as a reason to check navigation rather than a measured timestamp error.
For a working task list, try our meeting action items template and checked example. It keeps missing deadlines and source details visible.
04 / A SEPARATE FOLLOW-UP CHECK
AskFred follow-up email test: dates and task owners
We also asked AskFred for a client-facing follow-up using the same first recording. This was one separate prompt and its first answer, with no corrections or answer key supplied. It was not a second recording or a two-speaker test.
Read our exact follow-up prompt
Using only this meeting, draft a client-facing follow-up email from Emma to Liam. Include agreed actions, owners, final due dates and open questions. Respect any instructions in the conversation about what should be excluded. Do not send the email.
The draft included all seven agreed task dates, including September 10 for the brand assets and September 15 for the subscription-plan check. Those two dates were absent from the original notes. It also kept the final launch date, project budget, next check-in, and undecided Spanish scope.
However, five of its six deliverables did not explicitly name an owner. Feedback was assigned to “you” in the email addressed to Liam, and Emma separately promised to send the calendar invitation. The remaining deliverable dates still needed a clear person beside them.
The email body omitted the internal advertising discussion. Below the signature and a separator, AskFred added an explanation mentioning that discussion. Copying the whole answer would include that extra note. We did not send the email.
This is a useful example of different outputs preserving different details. It does not establish that follow-up drafts are generally more accurate than meeting notes.
For a checked rewrite and reusable structure, see our meeting follow-up email template and worked example.
SECOND RECORDING / TWO SYNTHETIC VOICES
Two-voice test: more deadlines, but still conflicting owners
We repeated the same fictional script with two synthetic voices on 6 September 2026. Emma used Microsoft Zira Desktop (en-US); Liam used Microsoft Frank (nl-NL) reading English. This 4:01 WAV included a 350 ms pause between turns. Fireflies displayed English (Global) and General Summary. We assessed the first output without editing or regenerating it.
This was not a real two-person meeting. Voice identity, pronunciation and pauses changed together. There was only one upload per condition, so this comparison cannot isolate why an output changed or measure typical performance with human speakers.
| Check | First: one voice | Second: two voices |
|---|---|---|
| Expected tasks mentioned | 8 of 8 | 8 of 8 |
| Agreed task deadlines present | 5 of 7 | 7 of 7 |
| Fully correct tasks* | 5 of 8 | 4 of 8 |
| Tasks with owner conflicts/errors | 2 | 2 |
| Key decisions preserved | 3 of 3 | 3 of 3 |
| Transcript speaker labels observed | Speaker 1 | Speaker 1 and Speaker 2 |
*The same completeness rule requires a consistent owner, final due date where agreed, and material restrictions. These counts describe two specific outputs, not product-wide accuracy or a statistically reliable improvement or decline.
September 10 and 15 returned; “Rima” appeared
The brand-assets deadline of September 10 and plan-check deadline of September 15 were now in Notes. Liam correctly owned the plan check. However, Notes assigned the brand assets to “Rima,” while the action list assigned them to Liam. The action list also added a separate “Rima” task to manage wireframe feedback, alongside Liam’s feedback task. There is no Rima in our script.
Which four tasks passed the full check?
The wireframe, integration feasibility check, copy draft and subscription-plan check had the required details without conflicts. Feedback and brand assets failed the ownership check. The invitation kept Emma, September 8 and the exclusion of Olivia, but did not explicitly preserve Liam as the only recipient. The follow-up task omitted the instruction to keep the internal advertising discussion out of the client email: saying that spending is not finalized is a different boundary.
That leaves four fully correct tasks under our strict completeness rule. The invitation and email findings are omissions in the generated notes; we did not send an invitation or generate a second-test email to establish what would actually be sent.
Two speaker labels did not guarantee accurate wording
The transcript alternated Speaker 1 and Speaker 2 across the fourteen displayed turns. It also contained unusual wording in Liam’s turns: the script’s “I remain responsible” appeared as “e Rima in responsible,” and “our photographs will not be ready” included the extra word “helium.” The notes then used both “Rima” and “helium.” This suggests a possible path from transcription errors to summary errors, but does not prove the internal cause.
Our practical takeaway remains to check the source recording before assigning work. In this second output, more complete dates coexisted with an unsupported name and missing instructions. A real-speaker recording and repeat runs remain the next steps.
05 / THE SETUP DETAILS
Fireflies privacy settings: recap emails and recording access
During onboarding, we selected manual meeting invitations and recap emails to “Only me.” After uploading the test file, its access setting still read “Teammates & Anyone with Link.” We changed this recording to “Only Owner.” The second upload also initially showed “Teammates & Anyone with Link”; we changed it to “Only Owner” and reopened the access dialog to verify the setting.
Choosing your own address as the recap recipient did not make this upload owner-only. Inspect recording access as well as email recipients before using a similar setup for client work.
The upload supported our WAV file. Automatic file selection through our browser-control extension was blocked; a manual upload worked. That was a limitation of our test setup, not evidence that Fireflies cannot upload recordings.
06 / COST & WORKFLOW
Pricing and the export limit we encountered
We used a free account. Selecting PDF in the first recording’s download dialog produced an upgrade prompt; no PDF downloaded. The provider’s download guide places downloads on Pro and above. We have not tested paid export quality.
For plan rates, annual commitments, one- and five-seat examples, and conflicting credit allowances, see our Fireflies AI pricing guide and cost calculator. It separates researched plan conditions from our free-account test results.
Check export entitlement if you need files outside Fireflies. Upgrading for access does not establish that the generated notes will be more accurate.
07 / OUR PROVISIONAL TAKE
Who should consider Fireflies AI?
The useful part of this output was the preservation of changed decisions and boundaries. If you need a starting point for reviewing a conversation, that is promising. But in this pilot, a complete-looking action list still needed corrections.
If you intend to turn notes directly into assigned tasks or a client email, check each owner and due date first. We have tested two synthetic recordings and one unsent follow-up draft on the first recording, but not autonomous sending, integrations, real multi-speaker meetings, or long-term use, so we are not recommending unattended use or declaring a category winner.
Fireflies AI review: frequently asked questions
Did Fireflies invent an approved advertising budget?
No. In this output, it explicitly kept the $500 ad spend unapproved and advertising outside the $3,000 project budget.
Why is this not an overall accuracy score?
These are two short synthetic recordings of the same script. We counted expected tasks and checked their details for each run separately. We did not measure word error rate, real-speaker separation, repeatability, or general performance.
Can I use these notes as a ready-to-send follow-up?
Check the owners and dates first. Our separate AskFred draft recovered all seven task dates but left five deliverables without explicit owners. Select only the email body; an explanatory note beneath it mentions the internal advertising discussion. Nothing was sent.
Is there a Fathom comparison?
Not yet. Fathom authentication did not complete in our Windows test environment. We have no equivalent output to compare and cannot attribute that individual login issue to a general product failure.
EVIDENCE & DISCLOSURE
Based on our 6 September 2026 fictional upload, original Notes and Transcript, and the answer key prepared before the test. A separate AskFred draft and PDF-export check used the same recording on that date. Screenshots show the provider homepage and pricing, our setup screens, and original output. Output screenshots are cropped to omit account details; the upload preview trims blank space visually, with the full original available. No paid plan was used. The provider link on this draft is a standard link with no affiliate tracking.
Read our complete test method →