The only automation maths that matters
Most automation business cases are written to be approved, not to be true. Here is the calculation we run before quoting, including the two costs everyone leaves out.
Blog
Written by the engineers doing the work, from real engagements. No listicles, no predictions, no vendor marketing dressed as analysis.
Latest
Most automation business cases are written to be approved, not to be true. Here is the calculation we run before quoting, including the two costs everyone leaves out.
Ranked from the diagnostic work we have done on other people's broken workflows. The causes are boringly consistent, and almost all of them are preventable at build time.
A demo agent needs to work once, in front of you. A production agent needs to work ten thousand times, unsupervised, including on inputs nobody anticipated. The gap is mostly engineering, not prompting.
Ranked from the diagnostic work we have done on other people's broken workflows. The causes are boringly consistent, and almost all of them are preventable at build time.
The developer left, the agency was replaced, and now a set of workflows nobody understands is running the business. A method for getting control back without breaking anything.
A demo agent needs to work once, in front of you. A production agent needs to work ten thousand times, unsupervised, including on inputs nobody anticipated. The gap is mostly engineering, not prompting.
Most retrieval builds fail at chunking and evaluation, not at the model. What we do differently and how we measure whether it worked.
Response time beats almost every other conversion optimization available to a service business, and it is mostly a plumbing problem.
Automating an undocumented process encodes its accidents permanently. The mapping session is the cheapest part of the project and the one that changes the outcome most.
The first automation should be small, dull, and finished in a week. Here is why the ambitious first project fails so consistently.
Most automation business cases are written to be approved, not to be true. Here is the calculation we run before quoting, including the two costs everyone leaves out.
Most production AI workloads are paying for capability they do not use on the majority of their traffic. Six changes that reduce spend materially and are measurable.
Real situations where the honest recommendation was to do nothing, fix something else, or hire a person instead.
The decision comes down to volume, control, and who maintains it. Here is the framework we use, including the cost crossover points where the answer changes.
WhatsApp is the primary business channel across much of the world and the rules for automating it are strict. What is allowed, what is not, and how accounts get restricted.
Send it over. We come back within 24 hours with an honest read on whether it is worth automating, and what it would return.