On-site consulting
The Consult.
I come into your business for two to five days, find the hours you lose to work a machine should be doing, and build the things that give those hours back. Custom software, automations, and models running on your own hardware. Then I teach your people to run it, and I stay reachable for a month while real work proves it out.
Running a one-person shop, or three of you? That is a day, not a week, and it is priced like one: the Drop-In, $2,000 to $4,000.
Where I actually stand
I have real worries about AI. I also think you should be using it.
Both things are true and I am not going to pretend otherwise to sell you a week of my time. I think the long-term picture is genuinely unsettled: what this does to entry-level work, what happens when the models get cheap enough to run everywhere, who ends up owning the infrastructure. Anyone who tells you they know how that lands is guessing.
What I am confident about is the narrow part. For the repetitive, structured, joyless work that eats a small business alive, copying a lead into four systems, rebuilding the same report every Monday, chasing a status update across three inboxes, there is nothing better available right now. Not close.
So that is the only thing I sell. Not a strategy deck about transformation. The specific tasks in your specific week that a machine should be doing, automated properly, running on hardware you own, with a human still holding every decision that matters.
A worked example
One lead, typed in five times.
This is the most common thing I find, in almost every business, and it is usually invisible because everyone has stopped noticing they do it.
A lead comes in. Somebody types the same eleven fields into five different places, and then nobody is sure which one is right.
The form email arrives. Someone opens the CRM and creates the contact. Then the shared spreadsheet, because the CRM report does not have the column the owner likes. Then the calendar invite. Then a message to whoever handles that region. Then, three days later, somebody asks whether that lead was followed up, and the answer takes fifteen minutes to find because two of the five copies disagree.
Call it eight minutes of typing per lead and another few minutes of reconciliation. At twenty leads a week that is roughly four hours, every week, spent moving text between boxes. Nobody ever gets promoted for it and everybody hates it.
The parts, in plain English
Four words worth knowing before anyone sells you something.
You do not need to learn this to hire me. You do need it to tell the difference between a real system and an expensive subscription with a chat box on it.
01
The harness
A model on its own only writes text. The harness is everything around it: what it is allowed to see, which tools it can actually use, what it must do before it acts, and what happens when it gets something wrong.
This is where nearly all the value lives, and it is the part nobody sells you, because it has to be built for your business specifically. The same model with a good harness and a bad harness is two completely different employees.
02
RAG, or: how it knows anything about you
Retrieval-augmented generation is a long name for a simple trick. Instead of training a model on your business, which is expensive and goes stale the moment anything changes, you keep your documents where they are, and at the moment of a question you find the few passages that actually matter and hand them to the model along with the question.
The practical consequences are the ones worth caring about. It can cite the document it used, so you can check it. Updating its knowledge means updating a file, not retraining anything. And because only the relevant passages get sent, this is what keeps token costs sane instead of shipping your whole handbook on every question.
03
Agentic workflows, and the boring cousin that does most of the work
An agentic workflow is a model that can take steps: read this, decide, call that tool, check the result, try again. It is the right shape when the path is not knowable in advance, like reconciling two lists that disagree in a dozen small ways.
But a lot of what saves you real time is not glamorous at all. It is a scheduled job, a cron, running at 6am on your own machine, doing the same reliable thing every day and only bothering a person when something looks off. I will build the clever thing when the problem needs it, and the boring thing when the boring thing is better, which is more often than people expect.
04
On-premises inference, and why I default to it
On-premises just means the model runs on a machine you own, in your building or your rack, instead of somebody else's service. For a surprising amount of business work this is now genuinely practical: a model small enough to sit on one off-the-shelf GPU can classify an email, pull fields out of a document, draft a reply, or check one list against another, and it can do that all day at no marginal cost.
The reasons I reach for it first are ordinary ones. Client records and staff conversations never leave your network. There is no per-token bill that grows with your success. It keeps working when a vendor changes their pricing, deprecates a model, or has an outage on the morning you needed it. And you can leave it running around the clock, because the only thing it costs is electricity.
The rule of thumb for sizing is arithmetic, not magic: a model's memory need is roughly its parameter count times the bits per weight, divided by eight. A seven billion parameter model at eight-bit weights wants about eight gigabytes of video memory, which is a card you can buy today. For one or two people a simple runner is plenty; for a team hitting it constantly there are servers built for exactly that.
The kit
You pick the model. You keep the harness.
Nothing here is proprietary to me, and that is deliberate. Every piece is either open, or something you can hold an account for directly. If we part ways, none of it stops working.
What the days look like
One to five days, in your office.
Two days is the minimum, because one day is only ever an audit and I would rather leave you with something running. A full week gets a system your team is already using by Friday afternoon. Either way, the month after the visit is part of the fee.
- Day one Watch, and write it down I sit with the people doing the work and time it. Not a survey: the actual clicks. By the end of the day we have a ranked list of every repeated task, with hours attached and an honest note on which ones are worth automating.
- Day two Wire the plumbing Access, credentials, a server or a machine to run on, the model chosen, and the first job running end to end on real data in draft mode. Usually this is the lead flow, because it is usually the worst one.
- Day three Build the specific things The custom pieces: the reconciliation nobody wants to do, the report that gets rebuilt every Monday, the intake that gets retyped. Written as code you own, on your hardware, with the guardrails agreed out loud.
- Day four Teach it to your people Hands on their keyboards, on their real work, not a slideshow. How to change a rule, how to read the log, how to tell when something is wrong, and how to turn it off. If they cannot run it without me, I have not finished.
- Day five Harden it and hand it over Failure modes, backups, what happens when a service is down, and the written record of how everything works. Then a list of what I would do next, whether or not that involves me.
- The month after Refine it against real work Included, not an upsell. Nothing survives first contact with a live week untouched: a rule that was slightly too strict, an exception nobody thought of, a field that turns out to matter. For thirty days after the visit I am reachable and I tune it, because a system that gets abandoned in week two was never worth building.
On a two-day visit these get compressed rather than skipped: the audit and the plumbing on day one, the highest value automation and the training on day two, and more of the building happens in the weeks afterwards.
What it is worth
Do the arithmetic yourself.
I am not going to quote you a case study with a made-up percentage on it. Here is the only calculation that matters, and you can run it on your own numbers in about a minute.
Take one task your team repeats. Multiply the minutes by how often it happens by how many people do it. Multiply that by a loaded hourly cost. That is the annual number sitting in one habit, and most businesses I walk into have four or five of them running at once.
One habit usually pays for the week. The rest is upside, and it keeps compounding, because the tools underneath get faster and cheaper every quarter while your harness stays the same.
Where this ends up
Some of these stop being tools and start being staff.
A script saves you an afternoon. A well-built system with a defined role, a schedule, and permission to act saves you a position, and it keeps saving it every year without a raise, a holiday, or a handover.
The honest framing: not a replacement for your people, a replacement for the part of their week they resent.
Once a system has been right for a few weeks, you stop reviewing every action and start reviewing the exceptions. That is the moment it changes category. It has a job description now: watch this inbox, triage what arrives, draft the standard replies, file everything where it belongs, and put the three things that need a human in front of one.
It works at two in the morning and on the Friday before a holiday. It does not get bored on the four hundredth record and start making mistakes, which is exactly what happens to a person doing data entry, and it never leaves and take the process knowledge with it.
Stacked over years the arithmetic gets hard to argue with. Several roles' worth of repetitive work, running continuously, on hardware you already bought, built once and owned outright. I am not going to put a headline number on it, because yours depends on your week. But this is the part of the engagement whose value people consistently underestimate at the start and overestimate having ever doubted a year later.
If you are small
The Drop-In: One day, for one person or a few
The Consult is priced for a team of about ten or more, because that is how many processes there are to untangle. If you are one person, or three, you do not need a week of me and I am not going to sell you one. A day gets the time audit, the worst job in your week automated properly, and you trained to run it. Same approach, same ownership, sized to what you actually have.
$2,000 to $4,000
One day on site or remote · 2 weeks of refinement · one-time fee
- A day on site, or split across two remote sessions if you prefer
- A time audit: the jobs eating your week, ranked, with hours attached
- One or two automations built and working before the day ends
- Honest advice on what to pay for and what to stop paying for
- Two weeks of refinement afterwards, once real work has run through it
- Everything built is yours: the code, the accounts, the notes
- If a week would genuinely serve you better, I will say so instead
The The Consult is priced for a team of about ten or more, because that is roughly where the number of processes, handoffs and systems makes a week worth it. I would rather size the engagement honestly than sell you the bigger one.
On site · The Consult
$8,000 to $20,000
Two to five days on site · 30 days of refinement · for teams of about ten or more
One fee, once. Not a retainer, not a seat licence, nothing to cancel later. Everything built during the week is yours outright: the code, the accounts, the documentation.