Most AI consultants have never had a dollar at risk. Not one.
AI implementation for operators who carry real financial consequence.
We find the constraint that's actually gating the business, build AI to fix it (in-house where that's cheaper, the right vendor where it's better), and stay accountable for whether the numbers move.
Show me what this looks like10+ years directing engineering teams before the first prompt.
Since 2015, managing developers without doing the development, which means specifying the work, breaking it down, and sending it back when it came back wrong. That is the entire job with agentic systems, and I had run that loop for 10+ years before there was a model to point it at.
Success Advisor at The Uncommon Business
One of the fastest-growing AI education companies in the world, founded by Callan Faulkner. As Senior AI Advisor in 2026 I advised on Automate to Accelerate (A2A), their flagship AI implementation program, working 1:1 with ~500 students and handling the most technically demanding builds across the cohorts (agentic workflows, Claude Cowork and Code systems, MCP integrations, custom AI tooling). Since October 2026 I am back as a Success Advisor, working with roughly 2000 students.
There is one constraint gating your company, and AI isn't working on it as well as it could.
You probably know which one it is already. More often than not it's the thing people bring to you because nobody else has the reps, and no document you've written has ever quite transferred it.
Sharran Srivatsaa (CEO of Acquisition.com) calls the underlying pattern the curse of capability. Talented founders are good at so many things that they build complexity nobody needed, then stay trapped as the only operator who understands how the pieces fit. I've lived that curse for most of my professional life, including in the land business, which is my most successful venture to date. We succeeded in spite of deficiencies I brought as the primary operator.
So the business grows until it reaches the edge of your calendar, and then it stops with demand still there...judgment doesn't delegate through a process doc, and the part of the work that actually carries consequence still routes through one person.
Most operators have some version of this, and it's almost never the thing they were planning to automate.
The evidence
The tool was never the constraint.
In a randomized trial, physicians given a top model didn't outperform physicians working without it, even though that same model on its own outperformed both groups. In a follow-up study, twenty hours of AI training didn't protect them from following bad AI advice. (That research is clinical diagnosis, not business workflow, so the transfer to your operation is an analogy rather than evidence.)
The structure still holds, though. The model was already better than the experts. What failed was the handoff between a capable person and a capable tool, and better models widen that gap rather than closing it, because the more often a tool is right, the faster people stop checking it.
That pattern is called automation bias, and it costs you judgment in two directions at once. You accept a recommendation you'd have pushed back on if a person had made it, and separately, you stop looking for the things that needed your judgment in the first place, because nothing got flagged and checking stopped feeling necessary. Both of those are your expertise going quiet. Neither one announces itself.
Twenty hours of training didn't do it, so the fix has to be structural. Build the verification into the workflow itself, where it happens whether or not anyone remembers to run it.
And that verification has to be yours. A generic check, the kind that ships with the tool, is the same mistake one level up, confidence with nothing behind it. Here is what ours actually catches, in our own system, every day. This comp doesn't belong, the Redfin view-map is set too wide, the subject property characteristics weren't read correctly, the child parcel subdivide math is off. Every one of those corrections came out of thousands of deals reviewed nationally and millions of dollars deployed and realized.
Your list will look nothing like that one, and that's the entire point. It has to come out of your own repetitions and the things that have already cost you money once, which is why the diagnosis comes before the build.
The four columns
A demo proves the idea is possible. It doesn't prove the thing survives your business.
Sharran sorts every task in a business into four columns.
Decide
Weighing the data and making the call.
Build
Whatever it takes to see that decision through.
Check
Validating that what got built actually works, judged against your experience and your data, and sending it back when it doesn't.
Run
How long the thing operates, automated accordingly.
For anything digital, Build has quietly left the owner's job description, and so has Run. Pre-2026, Build meant serious money and specialists. Today, once the decision is spec'd, Build is routinely the fastest part of the job.
Which is exactly where it goes wrong. It's easy to get swept up in the speed of that Build phase (I can smell a 10-minute Replit-built landing page a mile away), feeling like you're making massive progress. Then a real edge case hits. You describe the fix, the model patches it, the patch breaks something upstream, and you patch that. Six months later nobody can explain why the thing works, everybody is nervous about touching it, and the person who built it is the only one keeping it alive...which is your capital and your reputation riding on a system nobody fully understands anymore.
That's vibe coding, and it's excellent for finding out whether an idea is worth building at all. What it skips is the column that was always yours.
Check is where expertise and judgment live, and I'd argue it's currently the most valuable skill in this economy, precisely because so many people now offload it entirely (shipping whatever a model returns and assuming it works) or they perform a Check, but the expertise behind it was never earned through an established track record.
Generally, the systems that last get designed before they get written, with the failure modes named up front, the boundaries drawn so one broken piece can't take the rest down with it, the verification sitting where a human still signs, and the whole thing documented well enough that it outlives whoever built it.
What does not change
Your AI bill is going up. Is anything else?
The failure in this market is no longer refusal to adopt. Companies are adopting fast and aiming AI at work that was never the constraint, which raises output capacity without raising throughput, because the verification step never moved. It just relocated onto whoever owns the result.
That is work slop, and it quietly eats the gains before they reach the bottom line. Unchecked output ships downstream, so the capable person whose name is on the outcome (usually whoever is responsible for money changing hands) has to decode it, correct it, or rebuild it. The organization produced more, and the one person who can sign off on it is still the bottleneck, now reading somebody else's first draft on top of their own job. The token bill lands immediately and the margin never catches up.
Flat revenue is the obvious version of this. The more expensive version is revenue holding steady while margin compresses, because the new spend landed on top of the old cost structure instead of replacing any part of it. Margin is where this work should show up first, and frankly, a point of margin on revenue you already have beats the same dollars won at the old cost structure.
If the line item keeps growing, usage is up across the team, and nobody can point to a margin or a cycle time that moved, that is the whole problem in one sentence. Often the highest-return decision is to not run the initiative at all, which returns more than accelerating it would.
Faster
Table stakes. Everyone claims it and most deliver it, because AI makes work faster almost regardless of how carelessly it is deployed. Nobody should be paid a premium for speed alone.
Lower overhead
Claimed nearly as widely and delivered far less often, because most vendors are not running efficiently themselves. A deployment that spends model calls on work that should be deterministic is adding overhead while selling a reduction in it.
More consistent
Almost nobody delivers this, and it is the only one of the three that verification produces. The result has to hold on the hundredth run, not just in the demo.
Those three are what buyers wanted before any of this and will want after it. We anchor to them rather than to whatever a model can do this quarter, because a claim resting on a current capability has a shelf life measured in months.
Roughly 95% of generative AI pilots produce no measurable P&L impact. That is an overhead and consistency failure wearing other clothes, rather than a speed failure. (That figure is the NANDA study, which is preliminary rather than peer-reviewed, measures P&L impact inside six months, and skews toward sales and marketing pilots. We would rather state that than have you find it.)
Anyone can spend tokens. The claim is the same result on a fraction of them.
Our own underwriting run is the receipt. We underwrite land deals for our own account at Serious Land Capital, our capital at title, and that weekly review is the process we automated first. The first fully automated version took 3.5 hours and 10 to 15% of a week's model allowance.
Two things changed it. The procedure moved into code, so the deterministic parts stopped being model calls at all. Then each call that remained got matched to the model the task actually needs, with cheaper models carrying extraction, formatting and the mechanical passes, and the expensive model reserved for judgment and the written read. The same run now takes under 30 minutes unattended at roughly 1% of weekly allowance, with the top model still doing the work that actually needs it.
Model allocation is the part most deployments skip. Sending every call to the best model available is the AI equivalent of overnighting everything, and it sets a cost floor several times higher than the work requires.
You already own leverage that is paid for, meaning customers, reputation, distribution, a team that knows the work, years of accumulated judgment. AI is one more lever on top of that, not a replacement for it. We are not here to sell you more AI. We are here for the discipline that makes it pay, and the verification that proves it did.
Proof of effort
We ran it on ourselves first, including the part that hurt.
We automated SLC's underwriting, the single thing we're best at, on our own book, with our own capital exposed to the result. We call it the Deal Engine. It's running now, and it reviews our own deals every week.
The data extraction layer gave us the most trouble, and it failed intermittently rather than outright, which is the worse of the two. We could have kept patching it. We rebuilt that layer instead.
Before any of it, we put roughly $100K into Land Pricer, an AI-enabled system for pricing land to SLC's own standards, two years before the tooling was ready for it. We set it down to focus on the land business itself, and by the time we came back the comparable build cost about $160. We count it as tuition that paid, because what it forced into codified form is what the underwriting inside the Deal Engine now runs on. We would not claim the method itself as a moat. Anything written down can be copied. Two-plus years of building, breaking, and rebuilding has iterated into a design we keep getting better at, and the smaller wins compound the same way (our newsletter workflow cut editing time from roughly three hours to roughly one, with the voice intact).
How we work
Three steps, in this order.
Engagements run as a fractional Head of AI or fractional CTO seat, or as a scoped build, depending on how much of the operation is in play. Either way the sequence below does not change.
Constraint diagnosis.
We find the decision that's actually gating the business. The most expensive mistake in this work is automating something that was never the bottleneck.
Architecture and build.
Designed before it's written, with the verification sitting where a human still carries the liability.
Operation.
We run it with you and keep tuning it. Systems nobody owns are systems that quietly stop being true.
Who you're working with
The operator, start to finish.
Chris Duff, M.D.
Fractional Head of AI, fractional CTO, AI workflow engineer
I spent 10+ years directing engineering teams without writing the code. These systems need exactly that skill, specifying the work, breaking it down, and correcting it when it comes back wrong. It is why these systems went from new to operational here quickly, and why they keep working long after the build. You can check the record rather than take my word for it.
I trained as a physician and never practiced, but I never stopped practicing differential diagnosis. Rule out what can kill before you treat what might heal. That is the same discipline as underwriting, and it is the one I bring to every system I build.
Then I spent years applying both to a business where my own capital sat on every decision, and where being wrong cost me money. That is the part most people selling AI implementation cannot say.
- M.D., University of Illinois College of Medicine
- Managing Partner, Serious Land Capital. Over $6.5M funded and realized at 41% operating margins, own capital at title. No asset held longer than 15 months, most under nine months. Now automated down to a few hours a week
- 10+ years directing engineering teams, since 2015, much of it managing developers without doing the development
- Success Advisor, The Uncommon Business, since October 2026, after serving as Senior AI Advisor earlier in 2026. Working with roughly 2000 students on their most technically demanding builds
Start with the constraint.
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.
Tell us what's breaking