THE BOTTLENECK
It is September, so somewhere in your building there is a spreadsheet called Budget FY27 with a line on it that says something like "AI / automation." Somebody typed that description in about four seconds and has been avoiding the amount column ever since.
You have one number to put there. It came from a vendor, and it is the build.
That number is real, and it is the smallest one in the deal. Underneath it sit four more: what it costs to run every month, what it costs when volume goes up, what it costs when something has to change, and what it costs to leave. A proposal usually prices the first, gestures at the second, quotes the third at "our then-current professional services rate," and says nothing at all about the fourth.
So your controller asks a fair question, which is what this costs over three years, and nobody in the room can answer it. Someone will also cite the figure that 95 percent of AI pilots fail. That one traces to an unreviewed MIT NANDA working paper built on 52 interviews, where the 95 percent appears in a single sentence with no data behind it, and a Wharton professor who went looking could not find its basis. It is not an argument for spending and it is not an argument against it. It is noise in a room that already cannot price the thing.
When a number cannot be defended, the safe move is always the same. Defer it. Revisit in Q2. That is not a no, it is a not now, and it holds.
Here is what the deferral actually costs. The process you were going to automate does not pause while you decide. It keeps running the way it runs, which is on people. So the next increment of volume, the second location, the account you have chased for two years, arrives at a process that only scales by adding somebody. That is the growth ceiling, and it is not made of technology. It is made of a budget line nobody could defend. Which means the real choice was never automation versus no automation. It was automation versus a hire, and you already know how that hire is going.
For what it is worth, this decision is live at roughly a third of your peers. Census Business Trends and Outlook Survey data through May 2026 put AI use at 32 percent of firms with 100 to 249 employees and 37 percent at 250 and up.
THE FIX
Make the proposal produce the number instead of arguing about it.
Five buckets. Build, once. Run, monthly, which is hosting plus model usage plus somebody watching it. Volume, the per-document or per-seat rate that only shows up when the thing works and you push more through it. Change, every workflow adjustment after go-live. Exit, what you keep and what you can actually move.
My own numbers, so this is not a free lecture. Hosting and operations start at $600 a month, and a typical first year for a build client lands around $32,000 once twelve months of that sits on top of the build. I would rather you have that figure in hand when you read anyone's proposal, mine included.
The prompt below takes a proposal and returns the five buckets, a first-year total, and a list of what it could not total because the vendor never wrote a number.
It took three versions. Version one told the model never to supply a figure the vendor did not write, which was correct, and which also stopped it from adding $9,500 and $38,000 and $6,000 together. It came back with a careful, honest analysis containing no total, which is the one thing a budget line actually is. Version two let it do arithmetic on the vendor's own figures and made it label every calculation. Version three fixed the renewal deadline, which it could not compute because the notice clock runs from go-live and the proposal never dates go-live.
Where it breaks: it reads the document, not the deal. Anything agreed on a call is invisible to it.
When not to bother: if the proposal is two pages and you have read them twice, you do not need this. Use it on the forty-page one with an exhibit B.
I am an operations leader, not a developer, and I have to put a number in
next year's budget for this. Below is a vendor proposal exactly as it
arrived.
Every fact must come from the proposal. Never supply one it does not
contain, and do not fill a gap with what is typical or likely. Where the
proposal does not answer something, write "Not addressed" and move on.
Arithmetic is the exception. You may add, multiply and annualize the
vendor's own stated figures, and you must show the arithmetic when you do.
Label anything you calculated as "calculated." A number the vendor did not
write and that does not follow from their figures is off limits.
Quote the vendor's own words for every figure, in quotation marks. If an
item fits two sections, put it in the first one it fits and reference it in
the second.
ONE-TIME COSTS - Every charge that happens once. Amount and what it buys.
Total them.
RECURRING COSTS - Every charge that repeats. Amount, how often, and what
triggers it. For anything priced per user, per document, per transaction or
per unit of usage, give both the extra charge alone and the resulting total,
at the included volume and at double it, using their rate.
BUDGET TOTAL - Calculated. First twelve months, and the full committed term
if one is stated. State in one line the usage assumption the total rests on,
and that the proposal does not state our actual volume. Show the arithmetic.
Then list every item excluded because the proposal gives no number for it,
and say plainly that the real figure is higher by the value of that list.
COSTS WITH NO NUMBER - Anything the proposal says we will pay for without
saying how much.
PRICE CHANGES - Anything letting the vendor change pricing during the term.
Quote it. Note whether it states a cap or a method.
PAYMENT TERMS - When money is due, deposits, invoice timing. "Not addressed"
if silent.
WHAT WE OWN - What we own or keep at the end. Keep software, configuration,
and data separate. Where we get a license rather than ownership, say so in
plain words.
IF WE LEAVE - What happens to our data and our access if we cancel or do not
renew.
TERM AND RENEWAL - Length, notice required, and whether it renews
automatically. Name the event the notice deadline is measured from, and if
that event has no date in the proposal, say which date I need in order to
calculate the deadline.
WHO RUNS IT - Who is responsible after go-live, and what is included versus
billed separately.
DEPENDENCIES - Any third-party service, model, platform or vendor this
requires to keep working. For each, note whether we or the vendor controls
it.
QUESTIONS TO ASK BEFORE SIGNING - The gaps above that would move the budget
number, largest first. One line each, phrased so I can paste it into an
email.
Do not rate, score, or recommend. Do not tell me whether this is a good deal.THE OPERATOR'S READ
On August 26 OpenAI scheduled its transcription models for shutdown, including whisper-1 and the gpt-4o transcribe family, on February 26, 2027. If anyone at your company built something that turns recorded calls, dictated field notes or safety walkthroughs into text, that is now a dated maintenance job, and the date did not come from your side.
This is the cost that never appears on a quote. Models retire on the provider's schedule, and the re-test lands on whoever owns the system. The notice you get varies more than you would expect. OpenAI commits to at least six months for generally available models and as little as two weeks for previews. Anthropic commits to at least 60 days.
So add one question to any AI proposal you are handed this fall. Which model does this run on, what is its announced retirement date, and who pays for the re-test when it moves? A vendor who already manages that is offering you something real. A vendor who has not considered it is also telling you something.
ONE MOVE THIS WEEK
Pull the most recent automation or software proposal sitting on your desk. Before you read another page, find three things in it: the last date you can give notice to cancel, the rate for changes after go-live, and what you keep if you leave. Twenty minutes.
If any of the three is not in the document, you do not have a proposal yet. You have a build estimate.
— Rich
P.S. If you run that prompt on one of my proposals and it comes back with gaps, send them to me. I would rather fix the proposal than win the argument.