Serious AI with Chris Duff
October 19, 2026
Compare what the work cost before the pilot with what it costs now, measured the same way, and count everything the new version costs: the model bill, the subscriptions, and the hours people still put into checking its output. If the before number was never written down, you can’t tell yet, and rebuilding that baseline is the first job.
That answer comes from Serious AI Partners, an enterprise AI implementation firm run by Chris Duff, M.D., who has spent 10+ years directing engineering teams and built his judgment in a business where being wrong cost him money. The receipts below come from Serious Land Capital, his own land investment firm and the place every method here was tested.
Serious Land Capital buys raw land with its own money at title and resells it. Before its automations, overhead ran about $32.7K a year, and a full-time operations manager at roughly $1,576 a month, base plus deal bonuses, was about 70% of ongoing overhead. The role’s routines were rebuilt in roughly 48 hours and run on a single Claude Max subscription. Overhead now runs $13,754 a year, $1,146 a month all-in with tax preparation and entity filings included, and the contractor line is $0. Roughly $500 a month of software was cancelled along the way, including an SEO retainer at $200 a month and the CRM at $74.
That is a gap of roughly $18.9K a year. I trust it because both sides come out of the same general ledger, and the cost stack behind it was built on August 25, 2026 from actual card charges. As the Serious AI Proof page puts it, this is the same arithmetic we would run on a hire before anyone signs an offer.
Yes, and ours did on its first measured run. Underwriting (working out what a parcel is actually worth and whether the deal clears at that price) is the highest-value hour in a land business, and it took 45 to 60 minutes per parcel by hand. The first fully measured automated run, on August 18, 2026, took 3.5 hours and burned 10 to 15% of a week’s model allowance. The Proof page sums it up as correct answer, unsustainable economics.
Judge a pilot only on whether the output was right and that run passes. Judge it on its economics and it fails. What fixed it was moving the repeatable parts into code. The procedure became software, 40,190 lines of it with 10,036 of those being tests, so the deterministic parts stopped being model calls at all, and what was left got matched to the model each task needs, with cheaper models on extraction and formatting and the top model held back for judgment and the written read. It now runs in under 30 minutes unattended, at about 1% of the weekly allowance.
No, so run the old way and the new way on the same inputs and compare the answers. When the Deal Engine (the software that now runs that underwriting every week) disagreed with the manual read by roughly $19K on a test deal, the cause surfaced the same day, and the re-run landed at $87,992 against a hand-built $88,000. My read is that without the hand-built number on that same deal, a $19K miss would have looked like a faster, cheaper answer.
Set the accuracy bar in writing before the pilot starts, too. For SLC Chat, an internal assistant built on the firm’s underwriting, Chris’s April 13, 2026 newsletter issue set it down in writing: “We are not going live until SLC Chat is answering correctly ~85-90% of the time minimum.” My bet is that a bar written down in advance is hard to quietly lower once the tool is installed and people like it.
In our view, medical practices show the pattern most clearly. In a May 12, 2026 MGMA Stat poll of medical practice leaders, with 257 applicable responses, 46% said new AI tools had made their providers more productive in the past two years, 27% reported no productivity gains and 14% were unsure. MGMA, the Medical Group Management Association, calls the unsure group largely a measurement problem: uneven use, paired with limited tracking and feedback, makes it hard to establish a productivity baseline, let alone measure change.
The same write-up describes a February 2026 study at a large academic medical center where nearly all ambulatory physicians using an AI scribe perceived that their documentation time had decreased, while the measured savings were modest on average. MGMA’s advice to practices is to ask vendors for pre-post evidence using objective measures, not just satisfaction surveys. MGMA Stat polls run weekly among practice leaders who sign up for them, so read these as what those leaders report rather than a random sample.
Watch for work that moves rather than disappears. In MGMA’s August 4, 2026 poll, leaders whose groups use AI described what happened to staff workload, mostly in their own words, which MGMA then coded: 30% saw no real change, 27% said work shifted to other tasks and 19% said workload had been reduced. I’d only count shifted work as a saving once the desk it landed on gets cheaper too.
Against the cheapest way to get the same result today, as well as the way you used to do it. Chris’s May 18, 2026 issue, $30K A Year Ago. $160 Today., puts numbers on how fast that moves. In early 2025 the web app for Land Pricer, the firm’s earlier AI underwriting tool, cost about $25K to $30K out of pocket, with 4 to 5 engineers in rotation and a senior product manager. The same issue reports that a comparable web app build for SLC Chat cost about $160 in subscriptions and API tokens and half a week of Chris’s time, with him as the only human in the loop. The issue’s own caveat: Land Pricer was more complex in some ways, though not by orders of magnitude.
The same thing happens on the buying side. A few months before that April 13 issue, the firm seriously considered paying a third-party AI service a discounted $6K a year for what Chris called “basically a watered-down version of what SLC Chat is becoming,” and by the time of that issue, Chris wrote, Cowork let the firm build a better version itself, on a faster timeline, for pennies on the dollar. That story is in I Almost Paid $6K/year For A Worse Version. We think a pilot that beats last year’s quote can still lose to what a small team could build this quarter, so price both before you call it a win.
Write the before number down now. Pick the unit the business feels (hours per order, cost per claim, error rate per hundred) and pull it from the same source you will use afterward, so the comparison is like for like.
Count the full after cost. Model and subscription bills, the hours someone still puts into checking output, and any work that moved to another desk all belong on the after side.
Set the bar before go-live. Put the accuracy you will accept and the cost you will accept in writing, then run old and new side by side on the same inputs until the gap is clear.
Price today’s alternative. Ask what the same result would cost to build or buy now, rather than what it cost when the pilot was approved.
If a pilot can’t produce those numbers, the missing baseline is the finding. The receipts behind the Serious Land Capital overhead and underwriting numbers above are on the Proof page.
AI can produce the analysis. It cannot be wrong, because being wrong requires having something at stake, and that part stays with a person. Serious means underwritten.
=====
If you have an AI pilot running, or one that wrapped up, and the team can’t yet point to the number that moved, that is the place to start. The engagement opens with diagnosing the constraint gating the business and where AI is falling short on it, then building with verification sitting where a person still owns the outcome. You leave the first call knowing what your constraint is whether or not you ever hire us.
Fill out a short form, and we read every answer ourselves. If there is a fit we will reach out and find a time that works for you, and if there is not, we will say so and point you toward someone better suited. Either way, the moat question is worth answering now, in a market AI has already reset.