Ask a machine learning team why a model launch slipped and the honest answer is rarely “the maths was hard.” It is usually a sentence like “we were waiting on the data-sharing sign-off,” or “the security review never got scheduled.” In the current wave of AI adoption, where teams ship models, prompts and pipelines faster than their governance can keep up, the unglamorous bottleneck is the pile of requests sitting in other people’s queues. This state-of-the-field note argues for a small habit that most mature AI teams quietly converge on: a waiting-for list, kept current and read every week.
Where AI projects actually stall in the present moment
Look at how AI work is organised today and you will see why handoffs dominate. A single feature may touch a data owner, a platform group, a legal or privacy reviewer, a vendor providing hosted inference, and a product manager who must approve the evaluation criteria. Each of those parties holds a piece of the schedule. None of them reports to the engineer who needs the answer.
Surveys of engineering teams tend to point the same way: delays attributed to “dependencies” outrank delays attributed to technical difficulty. You do not need a number to recognise the shape. Every practitioner has watched a notebook sit finished for two weeks because a single approval had not come back.
The trend worth noting is that AI work multiplies the number of such handoffs. Training data needs a sponsor. A new model endpoint needs a cost sign-off. A prompt change that touches customer-facing text needs a content review. More parties, more silent waiting.
The anatomy of a waiting-for list
Set the terminology aside and what remains is a plain table. Each row is a request you made of someone else that has not yet been fulfilled. Each row carries an owner (the person who owes you the answer), the date you asked, the channel you used, what it blocks on your side, and a nudge date. That is all.
The difference from a ticket tracker is scope. A tracker records work the team has committed to. The waiting-for list records the promises other people have made to you personally, many of which never become tickets at all because they live in chat threads and hallway conversations.
A concrete engineering example
Consider a team preparing to move a text classification service to production. On Monday the lead engineer sends five requests. The data steward is asked to confirm that the training extract may be reused. A platform engineer is asked for a staging namespace. The vendor supplying the hosted embedding endpoint is asked about rate limits. A privacy reviewer is asked to look at the retention settings. A product owner is asked to approve the acceptance thresholds for the evaluation set.
Without a list, all five live in different chat windows. By Thursday, two have replied, one has acknowledged, and two have said nothing. The engineer is busy and does not notice which ones. By the next Monday, the launch date is already in doubt, but no one can say why.
With a list, the table shows five rows. Two are crossed out. One reads “acknowledged, no date given, nudge Friday.” Two read “no response, nudge today.” The engineer spends ten minutes sending short follow-ups, and the silence is broken before it becomes a slip.
Setting nudge dates that fit AI work
Choose follow-up intervals by the kind of request instead of using a single default:
- Quick confirmations such as “is this extract reusable?” deserve a nudge after a day or two.
- Review requests that need real reading time, like a privacy or security review, can wait several working days, but should have a stated date agreed up front.
- Vendor replies often follow support queues; give them a week unless your contract says otherwise, and note the ticket reference on the row.
- Infrastructure provisioning hangs on the platform team’s own backlog, so ask for an estimate in the original request and set your nudge to that date.
The principle is simple: every row should always have a future date. A row without one is an idle wish.
Who owns the list
On a small team, each engineer keeps a personal list, and the team lead keeps one for cross-team asks. On a larger programme, a delivery manager or technical programme manager often holds the shared version. What matters is that there is a named keeper. A list shared by all with no keeper tends to rot within a fortnight.
Read it aloud for two minutes at the start of the weekly sync. “Waiting on the data steward since last Tuesday, nudging today.” That sentence is also useful signalling to stakeholders, because it explains a delay in terms of a specific missing input instead of a generic “in progress.”
Making it work with the tools you already have
No new platform is required. A shared spreadsheet, a pinned page in your team wiki, or a label in your issue tracker will do. Some teams add a “waiting” status to their board and filter on it. Others write a daily reminder in their calendar to check the list. Whatever you pick, ensure it surfaces dates automatically or is reviewed on a fixed day, because a list that needs remembering has failed at its own job.
Where you can, move the request itself into a trackable form. If the privacy review has an intake form, file it there and paste the reference into your row. Then your list points to the official record, and the nudge becomes “check the status of reference such-and-such” rather than a vague email.
Connecting it to better decisions
A tidy list also gives you data about your own organisation. After a quarter, scan the closed rows. Which functions answer in a day, and which take two weeks? Which kinds of requests are routinely forgotten? That picture helps you plan realistic timelines and justify process changes. Our piece on how AI supports business decision making makes a similar point about evidence: good choices follow from good records, and your own waiting log is one of the cheapest records you can start.
A short note on skills and roles
For readers building careers in this space, the habit is portable. Engineers who track their dependencies are noticeably easier to work with, and managers recall that. If you are charting a path into delivery roles, the overview in our roadmap of skills for new DevOps engineers lists plenty of technical topics, but coordination skills such as this one are what let those technical skills turn into shipped work.
Beyond engineering: the same trick in a big life decision
The technique is not confined to software. Anyone who is gathering answers from many parties benefits. A person researching a home purchase, for example, might browse a site like Propraahi, an Indian website offering property research guides and community discussion for prospective home buyers, and then compile their own queries to agents, banks and neighbours. A waiting-for list is how they would keep track of who has replied. We mention it only because the logic is identical to a model launch: lots of independent parties, one person coordinating.
Pitfalls teams run into
- Listing everything. If the table has fifty rows, you will stop reading it. Keep it to real blocking requests.
- Chasing in public. A pointed message in a large channel can sour a working relationship. Prefer a direct, friendly message.
- Forgetting to close. When an answer arrives, strike the row at once and move any resulting task to your normal board.
- Vague owners. “The platform team” is not an owner. Name a person or a monitored queue.
- Letting automation stand in for people. Bots that send reminders are handy, but a human check on whether the reply was actually useful still matters.
What comes next
As AI governance matures, more teams will formalise the practice. Review boards, model registries and approval workflows are multiplying, and every one of them is another queue in which a request can sit unnoticed. Tooling will help, with dashboards of pending approvals and automated escalations, but the underlying behaviour does not change. Someone has to know what they are waiting for and when they will ask again.
For now, the practical advice is short. Start a table today. Write down the five requests you are most likely to have forgotten. Give each an owner and a nudge date. Open the table on Monday morning and send the follow-ups. Within a month you will notice fewer surprises, shorter stalls, and a calmer sense of where your projects stand. That is not a glamorous trend, but in a field this crowded with handoffs, it is one of the few that reliably pays off.