Compare/Wait for the models
Wait eighteen months and let the models absorb it.
This is the alternative nobody puts on a shortlist and everybody considers. It deserves a straight answer, and part of that answer is that you are right about most of it.
Stated fairly
Most of what you would buy today does get cheaper and better if you wait.
What waiting genuinely gets you
- Models get cheaper per token and better at the reasoning your workflows need
- Tooling matures, and the version you buy in a year is better than the one you buy today
- Somebody else pays to discover which approaches do not work
- No integration work, no security review, no budget spent this year
- If your critical work is genuinely simple, a general model may reach it without any of this
What does not arrive on its own
- A model that knows which of your tables is authoritative for revenue, or which exceptions your team learned the hard way
- Permission to act. Capability is not authorisation, and a better model still cannot decide who may write to your ledger
- The record of how your company decides, which only exists if something was recording while your people worked
- The queue between your experts and production, which is an organisational problem rather than a model one
Waiting makes the tools cheaper. It does not make your company legible to them.
the gap is not capability, and it does not close by itselfSuppose the models get much better
Three things happen, and none of them is this layer disappearing.
More surface
You hand over more, not less
Better models mean more agents doing more consequential work, and every one still needs to know how your company operates and what it may touch. A more capable model asking for a wider permission is a governance problem arriving sooner.
Sooner
Your own models arrive earlier
Everything on the phase two page gets closer as models improve. Tuning something on your own operations gets cheaper and easier every year, and it still needs the material only your own work produces.
The test
Take any one model away
If your company keeps its capability when a provider is removed, that capability was never inside the model. The brain, the rules, the corrections and the traces stay where they are.
The one thing waiting costs you
You cannot buy six months of your own corrections in a quarter.
Almost everything here can be adopted late and the late version is better. The exception is the record: traced runs and reviewed corrections are made of work your people have already done, so a budget cannot produce them on demand.
That is the honest reason to start with one workflow rather than a programme. The workflow pays for itself on its own terms, and the material accumulating underneath it is the part you would otherwise be starting from zero on in two years.
What that material becomes →Argue the other side with us.
If you think the models will absorb this, that is a conversation worth having properly, and it is more useful to us than agreement.