A CRM that keeps itself, and calls that write themselves up
An in-house build: call recording that starts itself, AI summaries written without anyone typing them, and a CRM record that stays current without manual entry.
This work is being carried out by Bipasha Das inside her current full-time employer, not under OPERIYA and not for an OPERIYA client. It is published here as implementation experience, described in her own words and generalised: the organisation is not named, and its commercial data is not disclosed.
- Administrative time recovered
- 325hours a monthModelled, not timed: 3 minutes of logging per call × 50 calls a day × 5 reps × 26 working days. Every input is stated so the figure can be recalculated with different ones.Model built by Bipasha Das from the team's own call volumes.
- Net saving each month
- $474$677 of administrative time at a $2.08/hour payroll rate, less $203 of telecom cost (13,000 minutes at $0.016). About $5,690 a year.Model built by Bipasha Das. Figures are converted from rupees at 96 to the dollar, the rate in September 2026.
The problem
A sales call produces two things: a conversation, and a record of the conversation. The first is the job. The second is the tax — starting a recording, remembering what was agreed, typing it into the CRM afterwards, and keeping the record current as the deal moves. It is done by the people whose time is worth the most, at the end of the day when they have the least of it, and it is the first thing to be dropped when the pipeline is busy. What survives is a CRM that is confidently out of date.
At this team's volumes the tax is not a rounding error. Five reps, fifty calls a day each, twenty-six working days, and about three minutes of logging and status-updating per call. That is the arithmetic printed at the top of this page, and it is the reason the work was worth doing: not because writing notes is unpleasant, but because at that volume it is a full-time job nobody was hired to do.
The state of this organisation's process before the build is not published here, and neither is the volume of calls it handles.
What changed
The system was built so that no step of the record-keeping asks a person to remember it. Calls are recorded automatically rather than on demand. The recording is summarised by a model rather than written up by the person who made the call. The summary lands against the right record, so the CRM is updated as a consequence of the work rather than as a separate task afterwards.
“Zero-touch” is the design constraint, not a description of the outcome: every part of the loop that could have been a prompt to a human — press record, write the note, update the stage — was built as something the system does on its own.
A sales call feeds an automatic recording, the recording feeds an AI summary, and the summary writes the CRM record. On its own baseline below, the manual path it replaced — write the call up from memory, then type it into the CRM — is drawn dashed and struck through.
Before — the process as it stood
- Sales call takes placeManual
- Recording started by hand, if at allManual
Depends on memory, so sometimes skipped.
- Call written up afterwardsManual
From memory, at the end of the day.
- CRM record updated by handManual
After — the process as rebuilt
- Sales call takes placeManual
- Recording starts on its ownAutomated
- Call summarised by the modelAutomated
- CRM record written and kept currentAutomated
A by-product of the call, not a task after it.
Systems involved
- SIP softphone on the rep's own handset
- Twilio — call routing and cloud recording
- Whisper — transcription
- A language model — note and status extraction
- GitHub Actions — scheduling and orchestration
- The sales CRM — write-back target
How it was implemented
Reps dial from a free SIP softphone on their own phones, set as the default dialler, so nothing about how they work changes. The call routes through Twilio, which records it. A backend job transcribes the audio, a model extracts the meeting notes and the new lead status from the transcript, and the result is written straight into the CRM against the right record.
The cost shape was a design decision, not an accident. Orchestration runs on GitHub Actions and the transcription and extraction run on free-tier AI, so the only per-minute cost in the system is telecom — the one part that genuinely cannot be free. That is why the running cost is roughly a third of the labour it displaces rather than a licence fee that scales with headcount.
The build was validated on a pilot first — one number, one rep — before it went to the team. A workflow that touches live customer calls is not something to roll out on the strength of a diagram.
What it produced
The upkeep left the salespeople. Recording, writing up and CRM entry are now done by the system, which is the whole of what “zero-touch” claims — not that the calls happen by themselves, but that keeping the record of them no longer competes with making them.
The interesting half of that is not the saving. Time recovered from administration is time available for the work the team is actually paid to do: on the same model, the recovered hours are enough for several thousand more outbound calls a month without hiring anyone. Whether the team chooses to spend it that way is their decision, and it is the reason the figures here are framed as capacity rather than as revenue.
The two figures at the top of this page are a model, not a measurement. Nobody sat with a stopwatch: the hours come from multiplying a per-call admin time by the team's real call volumes, and the money from multiplying those hours by a payroll rate. Every input is printed so a reader can put their own numbers in and get their own answer. The step count below is different in kind — it can be counted off the two columns above it.
- Steps in the loop that wait on a person
- 1 stepfrom 4 stepsCounted from the two process columns published on this page. The only remaining human step is the sales call itself.
- Recording no longer depends on anyone remembering to start it.
- Call summaries are produced without a person typing them.
- The CRM record is current because updating it is not a separate job.
What it says about operations generally
The expensive part of CRM hygiene is never the typing. It is the decision, dozens of times a week, about whether the typing is worth doing right now — and that decision reliably goes the wrong way when the person making it has a pipeline to work. Automating the record removes the decision rather than the keystrokes.
Nothing derived from the content of a recorded call appears on this site: no transcript, no summary, no customer, no deal. The people on those calls agreed to a conversation, not to being used as marketing material.
Related work
Is something like this running in your business?
The audit starts by finding out. One hour, no charge, and you leave knowing whether it is worth buying.