Most of the AI projects I see get priced before anyone knows what they are worth. That is backwards, and it explains a lot of the ones that die.
Here is why it happens. Agent tooling got useful fast, and it got useful after this year's budgets were already set. Almost nobody has a line for it in the plan, so every AI project is competing for money that was allocated to something else. The first question leadership asks is what it costs, because that is the question you ask when the money has to come from somewhere.
So a vendor gives a number, or the internal team estimates one, and it lands in a meeting with nothing to compare it to. A six-figure quote to build a system that drafts your proposals looks either reasonable or insane depending entirely on what proposals cost you today, and nobody in the room knows that number.
Two things happen from there, and both are bad. Either the project gets killed because the number looked big, or it gets shrunk to fit whatever budget is left over, and a shrunk project turns into a pilot that cannot prove anything, so it dies more slowly.
The fix is to do the value work first, and I mean something specific by that. Pick one workflow and count what it costs you today — in hours, in cycle time, in errors, in deals you answered too late. Turn that into a dollar figure your CFO would sign off on, not a number from a vendor slide. Then write down what would have to change for that value to actually show up, because the software is not the outcome, the rollout is. A drafting tool nobody adopts is worth zero no matter what it could do.
In practice this means sitting with the two people who run the process and timing it for two weeks. The numbers you are looking for are not complicated.
None of that needs a consultant — just a spreadsheet and two weeks. The two weeks matter, by the way. Measuring can turn into its own project if you let it. Set the window, get a number that is roughly right, and move.
Once you have it, the pricing conversation is different. Instead of asking whether you can afford the build, you are comparing the build to one year of what the process costs you now, and most of the time that comparison is not close. Sometimes it goes the other way and the value is not there — and that is a good outcome too, because you found out for the cost of two weeks of measuring instead of a failed project.
The value work also fixes scope. When you know the money is in the hours estimators spend finding comparable jobs, you do not build the tool that does everything. You build the thing that goes after those hours, and you can tell in month one whether it is working, because you measured the baseline. It is also how a project gets funded mid-year: it pulls budget from the process it replaces, and it can only do that with a number attached.
We do a paid version of this for companies that want a number they can take to a board, but the rough version is something you can do yourself, and you should do it before you talk to anyone about price, including us.
Pick the one AI idea in your company with the most momentum right now. Ask whoever is championing it what it is worth per year, and see if they can answer with a number and a method.
What is the most expensive process in your company that nobody has ever measured?