Tokenomics In Commercial Real Estate AI: 4 Questions To Ask Before Starting Each Task
Wesley D. Snow, Co-Founder and President of Ascendix Technologies. 30+ years in software development; a pioneer in AI for Real Estate.
gettyβFor many firms, AI cost management begins with an unexpected bill.
For a long time, the cost of each AI task was hard to track, partly because the tools didnβt show it and partly because there is no clear understanding of how many tokens each task will take.
As a solo worker who uses AI for tasks it can facilitate, and not every little thing in my to-dos, I rarely hit the token wall. But as the head of a company full of workers who are curious about exploring AI and implementing it into their workflows, an overcharge is actually something I expect.
AI companies priced their plans to win market share, and many users, including me, built daily habits around those tools. The next invoice is often an unpleasant reality check. Mine was $1,500, and it came from a single automated job. β
βAI providers bill each request in tokens, and a token is roughly three-quarters of a word. Each request is billed for what AI reads (cheaper input tokens) and, separately, for what it writes (more expensive output tokens). For example, according to Anthropicβs current pricing, Fable 5.1 charges $10 per million input tokens and $50 per million output tokens it writes. So, each written word costs five times as much as a read one.
The choice of model is an even larger factor than the read/write price split. The smallest Claude model, Haiku 5.5, charges $0.10 per million input tokens, 100 times less than Fable. By my rough math, a typical deal summary costs about 15 cents on the top model and less than a penny on the smallest one.
A difference of a few cents matters little for one request and adds up quickly for automated jobs that run many times a day. ββ
Looking up a contact or sorting emails by sender takes a person a few seconds. An AI assistant, given the same task, searches your connected systems and reads through records to find a phone number you could have looked up yourself, and you pay for every record it reads.
Summarizing a 60-page lease or rebuilding a deal history from scattered emails takes a personβs hours, and AI does it in minutes, so the time saved easily covers the token cost.
Because output tokens cost more than input ones, tasks with a lot of reading and a short answer are the cheapest way to use AI.
A lease summary fits that pattern, since the AI reads the full lease and writes a page or less. An interactive dashboard or a deck works the other way, because nearly every token in it is generated output. Generating an email is still relatively cheap, especially when you do not use the most advanced model on such a trivial task.
Visual, dynamic, interactive and self-updating output is the most expensive kind.
A mistake in a rough first draft of an email takes a minute to fix, while a lease abstract with the wrong rent escalation can lead to a billing dispute and a lost client.
AI tools make errors more often than their confident tone suggests. Invented details and outdated information are not as common in AI output as it used to be, but the time spent checking should count toward the task cost.
A question you ask once is billed once, while a job that runs on a schedule is billed on every run for as long as it stays switched on.
My $1,500 bill came from a scheduled job. A while back, I set up a job that reads my inbox every three hours, matches senders to CRM records, and logs the emails. It reads every message, including unqualified cold pitches from data sellers and follow-ups from a webinar I attended once. It ran on the most expensive model because that was the default setting, and once it was running, I stopped checking it. β
Large companies run into the same problem on a larger scale. Uber reportedly used up its 2026 AI budget in four months and then capped employees at $1,500 per tool per month, roughly the amount my inbox job cost me. ββ
Run a week of your tasks through those four questions and sort them into the ones worth sending to AI and the rest.
Lease summaries, meeting prep from emails and notes scattered across systems, deal histories and first drafts can usually be sent to AI, because each one replaces hours of reading with a short answer that is quick to check.
Contact lookups, inbox sorting and dashboards nobody opens can usually be done without AI, because a person can do them faster or nobody uses the result.
Repeating this as your tasks, the available models and their prices change will lead you to effective AI cost management instead of slowing down AI adoption or stopping its use at all. β
It also helps to tell the AI that youβre working within a budget and ask it to flag tasks where a cheaper model or a shorter answer would do. In my experience, it suggests useful savings once it knows that cost matters. β
I still use AI every day, but I now check what a task will cost before I hand it over. The inbox job behind my $1,500 bill has been replaced by an AI email parser built for that task. It operates on a cheaper model and can process every incoming email automatically, but I keep that setting switched off and send it only the emails I want logged in the CRM, so webinar follow-ups and cold pitches never reach it.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?