Estimates for this kind of work miss because they price the build. The build is the predictable part.
The realistic shape
Weeks 1 to 2: access and reality. Getting seats, keys and a sandbox. Reading real records rather than documentation. Mapping the process as it actually runs, which is never how it is described. This phase looks like nothing is happening and it predicts everything that follows.
Weeks 2 to 4: first useful result. One team or one location, in production, doing something that was not happening before. Narrow on purpose. The guidance from people who build these systems is consistent: start with the simplest thing that works and add autonomy only where it earns its cost.
Weeks 4 to 10: it survives its author. Error routes, alerting, a runbook, a named owner, reporting somebody actually reads. This is the phase that gets cut under pressure, and cutting it is what turns a system into a dependency.
Quarter 2: rollout. Fast only if a location is a configuration rather than a build. If unit two requires rebuilding, this phase expands without limit.
The four things that cause delay
Access. Roughly half of all slippage in these projects is waiting for a seat, an API key or a sandbox. It is administrative and entirely avoidable.
An undocumented process. If nobody can describe what should happen, the first weeks go to discovering it. Worth doing, but it is discovery time, not build time, and it should be planned as such.
Data quality. Every plan assumes cleaner data than exists. Duplicates, inconsistent fields, contact details older than the records holding them. Pull fifty real records before the estimate, not after.
Compliance registration. In the United States, application to person messaging requires brand and campaign registration with the carriers. Until it is approved, texting is blocked. Voice usually is not blocked by the same gate, so the sequencing matters: ship voice while the registration is pending.
What to agree before anyone builds
- One result that will be true at the end, named in business terms
- Who grants access, and by when
- The definition of done: what a person opens to verify the business result, not an execution log
- The handover date and the owner
Projects with those four written down land inside their estimates far more often, because the estimate is describing the same work everyone is doing.
A note on measuring too early
You will see directional movement on speed and coverage metrics inside a month. Attribution takes a quarter, because it needs enough closed deals to be more than noise. Judging a revenue system on four weeks of attribution data produces confident conclusions from nothing, and non deterministic systems in particular have to be measured by running real data through them rather than reasoned about theoretically.
If the blocker is inside your systems
That is a specific job. MitHub publishes open profiles of engineers who have taken this work from pilot to production, with the deployments written out.