Case Studies
A Day in the Life of a Digital Operations Partner
Eight jobs, five of them with AI's help. Every decision in them still needed a human who knew the client and its customers.
8 Oct 2026 · 8 min read · By Sophie Kazandjian
People often ask how a digital operations partner differs from a virtual assistant. The other question this day answers is what's left for a person to do once AI can draft the emails and build the pages. It's one day of my retainer work, with the client and everyone around them left unnamed, followed by a different kind of job: a product demo video that took two sessions and eight versions.
The client sells to larger organisations, which then send their staff through its work. I look after the operations around that: sign-ups, an online learning platform, the emails people receive, reporting for each paying organisation, and the website, which I built.
The shape of the work
Seven jobs in a day, and one build
| Job | Building | Writing & creative | Coordination | Judgement | AI used |
|---|---|---|---|---|---|
| 01Sponsor dashboard My build | |||||
| 02Participant FAQ page My build | |||||
| 03One network change for a whole organisation | |||||
| 04Web page in three rounds My build | |||||
| 05A place carried forward | |||||
| 06Survey reminder email | |||||
| 07Social post | |||||
| Another day: one build, two sessions | |||||
| 08Product demo video My build | |||||
Sponsor dashboards
Every organisation that buys from this client wants to know the same thing: how are our people getting on? Each one now has a read-only dashboard showing who is signed up and how far they've got. The dashboards were my idea, and AI helped me build them. Seeing that sponsors needed them came from knowing what they ask for. They run on Cloudflare, reading straight from the client's database, so there is no per-viewer licence to pay. I've written separately about why I chose that over Airtable's paid add-on.
I sent a new one for review that day. The only change the client asked for was to the first number on the page. The headline count still included people who had withdrawn, so the true figure only appeared further down. The headline is the number a sponsor reads first, so it now counts only people still enrolled.
A coach with employer-funded clients or an agency with several accounts could use the same build. The "how are our people getting on?" email becomes a link the sponsor already has.
A participant FAQ
Participants send their questions to a shared inbox, and most of them have been asked before. The website had an FAQ, but it was written for buyers, who ask different things. I wrote one for the people taking part, covering signing in, missed sessions, technical trouble and changing dates. AI can produce a plausible FAQ in seconds. It can't know how this client handles a problem in the middle of a live session, or who has to approve a change of dates, and those were the two answers the client corrected.
Once it was approved, I added the link to the foot of every participant email, next to the inbox address, and sent it to the colleague who covers the inbox part-time.
A course that wouldn't load
Late in the afternoon a participant wrote to say parts of the course wouldn't load on their work laptop. Their IT desk hadn't found the cause.
The course platform pulls in video, interactive exercises and forms from other services. If a corporate network filter blocks any of those domains, a lesson loads only in part, which matched what they described.
Why the course wouldn't load
One change instead of a ticket per person
The problem
A lesson is assembled from several services. Block one domain and it loads only in part. Each person from that organisation hits the same problem and reports it separately.
The fix
One email for the client to send, with everything IT needs in it. One change on the organisation's network then covers every participant for as long as access is needed.
I recognised it because another organisation had hit a similar problem with a webinar platform. Handled one ticket at a time, every person from that organisation would have to report the same problem separately. So I drafted an email for the client to send to the organisation's programme lead, with the IT contact they had given us earlier copied in, asking for what had worked the first time: one named person in IT, the participant list, the dates access is needed for, and the exact domains to allow.
A web page in three rounds
A page for something new the client is launching had gone up in draft the night before. The feedback said it described too much of the wider offer, and asked whether it could show the real thing. I rewrote it and took screenshots that showed enough without giving the content away. How much of something the client sells to show on a public page is a commercial judgement, and the screenshots had to make it. Two more rounds followed, mostly wording and consistency, plus a new section near the top for buyers.
All three went live that afternoon. I built the site, so nobody else had to be booked in to change it.
A place carried forward
Someone could no longer take part, and their organisation couldn't find a replacement in time. Processing the withdrawal took minutes, and an AI could have done it.
What it couldn't have done was weigh what the place was worth. The organisation was new to the client, had handled its registrations well and had lost the place through no fault of its own. I suggested to the client that it carry the place forward to a later cohort, because how a new customer is treated the first time something goes wrong shapes whether it comes back. The client agreed and told the organisation's sponsor, and I recorded the arrangement on its dashboard, where both sides could see it.
Smaller jobs
A reminder email asked people to finish a short survey before their first session and told them why. The survey records where each person starts, and answered later it can't. Last came a draft social post that read as more alarming than it needed to. I rewrote it from a calmer angle, because judging how a post will read depends on knowing who reads it, and set it up in the client's publishing workflow for approval.
A product demo video
The voice asking the questions in this demo was generated in ElevenLabs. It stands in for a user speaking to the app, and the words on screen were typed by the app's own dictation as it listened. Speech also suits a demo, because the room hears the questions instead of having to read them.
A client's product lead needed a demo video of their app prototype for a sales presentation to a prospective buyer. Her script went screen by screen: what to type, which options to pick, where to scroll, which AI responses to show. The video had to look like someone using the app on a phone and talking to it through speech-to-text, and it had to leave room for a director to narrate live.
I had the exported prototype, five short voiceovers I'd made in ElevenLabs, and two rough screen recordings of me clicking through the whole sequence with the app's AI switched on. The recordings were no use as footage, with my typing, a visible cursor and uneven scrolling. What they held was the AI's real wording for every response, which the video couldn't do without.
So I didn't edit them. I ran the prototype in an automated browser and replayed the script frame by frame. The voiceovers fed the app's dictation, so the words filled each box in time with the voice, and every AI response was copied word for word from the rough recordings. A buyer watching sees the app's real output.
The rest was a matter of making it look right. I chose the oak-desk background and refined it until the phone sat on it as it should. The camera moves in where there's text to read, a small circle marks each touch, and the voice is mixed underneath.
The demo video
Real output, replayed frame by frame
What I had
Replay
Render
The AI reproduced the app faithfully. It couldn't judge how the video would feel to someone watching it in a room, and every round of changes came from imagining the buyer watching.
The human in the loop
Nine things the AI missed
- The opening launched into speech before anyone could read the screen.
- The text lagged the voice too much, then once tightened followed it too closely.
- Viewers needed time to read each question before the answer appeared, as a real user would.
- Some zooms were quick and served no purpose.
- Scrolls were too fast to show the viewer was still on the same page.
- The processing screen needed to be longer, then shorter with a clearer pulse, then finer rings kept off the text.
- The touch circle and text cursor belonged only where the action was.
- A pulsing microphone after dictation suggested the app was still listening when it wasn't.
- A faint line flickered on the edge of one screen.
Each version was saved under its own name. That paid off when a media player kept serving a cached copy of an older cut and it looked as if the fixes hadn't been made. Version eight, after two working sessions, runs 5 minutes 17 seconds at 1080p, with a smaller copy on a shareable canvas for quick review.
FAQs
- What does a digital operations partner do day to day?
Runs the operations and builds the systems underneath them. On this day, for one client, that meant a sponsor dashboard, two website pages, an IT escalation, a carried-forward place, a reminder email and a social post. Other days go almost entirely on one larger build, such as a database, an automation, a website or a product demo video.
- How is that different from a virtual assistant?
A virtual assistant takes tasks off your list. A digital operations partner also asks why the tasks keep arriving and builds what stops them, such as an FAQ page that answers people before they write in, or a dashboard that answers a sponsor without an email. Most of my retainers include both kinds of work.
- Where does AI fit in?
It helped with five of the eight jobs here, the ones that produced something: the dashboard, the web pages, the social post and the demo video, where ElevenLabs supplied the voice. It can't weigh a client relationship or know how a particular client runs its work. Carrying a place forward for a new customer, deciding how much of a product to show and catching the nine problems in the demo were all judgements a person had to make.
- Does this only apply to businesses that run training?
No. The same pieces turn up in coaching practices with employer-funded clients, in agencies with several accounts and in any business that reports to the people who pay for it. I also take consultancies through Cyber Essentials, which many larger buyers now ask their suppliers for.
- How would this work for my business?
Book a free 45-minute consultation. We talk through what's working and what's stuck, and if it's a fit I send a tailored proposal with options and pricing within a few days.