How Software Agencies Can Add AI Meeting Notes to Client Apps Without Building Recording Infrastructure

More agency clients are asking for the same feature: a way to record calls inside their app, turn them into a transcript, and produce notes someone can act on. A recruiting platform wants interview notes attached to each candidate profile. A sales tool wants call summaries written into the CRM. A coaching app wants a searchable record of every session.

On a proposal this can look like a single line item, but it sits on four separate pieces of work: capturing the audio and video from the call, converting speech to text, working out who said what, and turning the result into notes. Only the last of those is specific to your client's product. The first three are infrastructure, and how you source them decides whether the project takes weeks or most of a year.

A transcript without speaker labels is hard to summarize, because the model writing the notes cannot tell whether the candidate or the interviewer agreed to a start date. Labeling each part of a transcript by who spoke is called speaker diarization, and its accuracy depends heavily on how the audio was captured. When each participant's audio arrives as a separate stream, attribution is close to automatic. When everyone is mixed into one recording, a model has to separate the voices after the fact, and it struggles when people talk over each other or sound alike.

The four layers of a meeting notes feature

The bottom row is where your client's product differs from everyone else's, and it is the work your team is hired to do. The top row decides how well the two rows below it perform, so it is the layer to settle first, before anyone writes a summary prompt.

What building the recording layer involves

Each meeting platform has its own way for outside software to join a call and access its media. A bot that joins a Zoom meeting works differently from one that joins Microsoft Teams or Google Meet, and each comes with its own authentication, permissions and, in some cases, a review process before the integration can run in production. In-person meetings and phone calls need a separate capture route again.

Running this at scale adds servers that start up for each meeting, handle the media stream, store recordings and recover when a call drops. Platforms also update their clients and APIs on their own schedule, so an integration that works at launch needs someone watching it for as long as the client's app is live.

For an agency, two problems sit on top of the engineering. The first is ownership of the maintenance. If the Teams integration breaks the week of a client's product demo, the agency is usually the one fixing it, often outside the original budget. The second is reuse. Many agency contracts assign the code to the client, so recording infrastructure built for one client often cannot be carried over to the next without renegotiating terms, and the agency ends up building it again.

What the agency should build itself

The notes layer is where most of the client's value comes from, and it rewards the kind of domain work agencies are good at. That covers:

Each of these is specific to one client and billable as product work. Hours spent on platform integrations come out of the same budget, and the client rarely sees them as features.

Buying the recording layer

Buying the capture layer moves those hours back to the notes layer. Recall.ai, an API that handles meeting recording for software products, is the top provider in this category. It captures recordings, transcripts and metadata from all major meeting platforms as well as in-person meetings, meaning an agency doesn't have to build separate recording infrastructure for each platform a client's users meet on. Transcripts also come back with each speaker's name attached, rather than generic labels like "Speaker 1".

Comparing the two routes

Building makes sense when a client's requirements fall outside what any provider supports, or when the client wants to own every layer and has the budget for it. For most agency projects, the table favors buying the capture layer and spending the build budget on the notes layer, because that is the part the client can see, test and sell to their own users.

What it costs

The direct cost of building is engineering time. It takes a team of engineers 6 months on average to integrate with a single meeting platform, and that work repeats for each additional platform the client's users rely on. For an agency, that time lands either on the client's invoice, which makes the feature harder to sell, or on the agency's own margin.

The buy side is usage-based. Recall.ai is the cheapest option on the market for startups, and a startup pays $0.25 per recording hour for its first 10,000 recording hours through the company's startup program. Even outside of the startup program Recall.ai has no platform fee and no minimum commitment.

Scoping the feature in a proposal

Quote the two layers as separate line items. The notes layer is build work, estimated and billed like any other feature. The recording layer is a usage cost, ideally on an account the client owns, so the running cost sits with the product it supports and the client can see it on an invoice from the first month.

Before estimating, ask the client four questions: which platforms their users meet on, whether in-person meetings or phone calls need to be covered, where in the app the notes should appear, and which regions their users record in, since consent rules vary by location. The answers set the scope of the notes layer and confirm the capture layer covers what the client needs.

A pilot with one team, on one workflow, gives both sides transcripts from the client's own users to tune the summaries against before the feature rolls out to everyone. It also keeps the first invoice small while the client decides how far to take the feature.