AI Build Plan: What To Build, What To Skip, And What Each One Costs

Written Updated by
AI build plan · one department · three weeks · a written answer

Someone has told you to do something with AI. Nobody can tell you what, or whether it pays.

I spend three weeks inside one part of your business: interviewing the people who do the work, watching how it actually runs, and reading the systems behind it. You get a ranked, costed answer. What is worth automating, what architecture it needs, which models fit, what each one costs to run, and the ones that are not worth doing at all.

Fixed price, no hourly surprises. Three weeks, one department or workflow. Paid in four milestones, you release each one only when you approve it. No build attached, and no obligation to hire me for one.

The board asked what your AI strategy is. A vendor demoed something impressive last quarter. Someone built a ChatGPT thing last year and has since moved on, so nobody is quite sure what it does. Two consultancies quoted six figures for a transformation roadmap. Meanwhile nobody has sat with the people doing the work and asked where the hours actually go. Everyone has an opinion about AI. Nobody has looked at your business.

How it works

Look at the real work. Rank it by what it is worth. Write it down.

Three weeks, three milestones of work and one of writing, aimed at a single department or workflow: support, back office, sales, content, internal knowledge, reporting, whichever one is hurting. What I need from you is time with the people who do that work, read access to the systems behind it, and whatever volume numbers you already have.

01

I come and look at the actual work

Interviews with the people doing it, not only the people managing it. How a request arrives, what gets done to it by hand, where it waits, what gets rekeyed from one system into another. Plus the systems themselves: what data exists, what shape it is in, what is already written down.

02

I rank it, including what fails

Every candidate gets the same treatment: what it is worth in hours or money, what it would cost to build, what it costs per month to run, what breaks if it gets an answer wrong, and how hard it is to undo. Some candidates come out of that ranking as a clear no, and those are written up as carefully as the yeses.

03

You get it in writing, and we go through it

A document, not a deck: the ranked list, the architecture each surviving candidate needs, which models fit and why, build against buy, self-hosted against hosted, what to do first and what to leave until the first one is working. Then a session with whoever needs to act on it, to argue with me about it while I am still here.

WHAT EVERYONE SUGGESTS THREE QUESTIONS, ASKED OF EACH WHAT SURVIVESthe board, a vendor demo, a team wishlist · answer the tickets · summarise the calls · write the proposals · read the invoices · forecast the demand · score the leads plus the ones nobody says out loud 1 · Repetitive enough to pay? volume, frequency and hours, actually measured 2 · Is there material to use? documents, history, logs, and their real state 3 · Value over build plus run? including what a wrong answer costs you BUILD THESE, IN THIS ORDER 1. the one that proves the approach architecture, model, build + run cost 2. the one with the biggest number architecture, model, build + run cost 3. the one that needed those first architecture, model, build + run cost DO NOT, AND WHY NOT too little volume to ever pay back no usable data yet, and what to keep a cheaper fix that is not AI at all a wrong answer costs more than it saves each says what would flip the answer The rejects are the deliverable too Anyone can produce a list of things to build with AI. The useful half is the list of things you were about to fund that should not be funded yet. Costed in both directions Every candidate carries what it is worth, what it costs to build, what it costs per month to run, and what it costs on the days it is wrong.
Everything you could point a model at goes in. Three questions decide what comes out: is the work repetitive enough to be worth it, is there material to work from, and does the value beat what it costs to build and run. The rejects are part of the deliverable, with the reason attached.

What you get

A ranked list with real numbers, and the reasoning behind every one.

One fixed price for all of it. Everything is written down and handed over, because a recommendation you cannot re-read in six months was a conversation, not a deliverable.

Look

What the work actually is

  • A map of the workflow as it really runs, built from watching it and asking, not from the process document somebody wrote three years ago
  • Where the hours go, broken down by task, with the volumes attached: how often, how long, by how many people
  • What material already exists: documentation, ticket history, recordings, spreadsheets, logs, and what state it is in, because this is what decides whether half the candidates are possible at all
  • What is already running, including the AI nobody owns: which tools, who pays for them, what they are actually used for, what overlaps
  • The non-AI fixes you hit on the way, because some of what looks like an AI problem is a missing integration or a form with the wrong fields on it
Rank

Costed, both directions

  • Value per candidate, in the unit your business already uses: hours saved, queue cleared, cycle time, headcount not added
  • What it costs to build, as a range with the assumptions stated, priced as work anyone competent would do, not as a quote from me
  • What it costs to run, per month, at your real volume: model spend, infrastructure, and the human review it still needs
  • What it costs when it is wrong, which is the number nobody puts in these documents and the one that decides how much review a thing needs
  • The rejects, with reasons: too little volume, no usable data yet, a cheaper fix that is not AI, a risk your industry will not accept. Each one says what would have to change for the answer to flip
Decide

The technical answer, not a direction

  • An architecture per surviving candidate: what runs where, what it reads, what it writes, where a human sits in the loop, and how it fails safely
  • Model picks with the reasoning, against your actual constraints on latency, privacy, cost and quality, including where a small cheap model is the right answer and where it is not
  • Build against buy, named: where an off-the-shelf product wins, which one, and roughly what it costs. Sometimes the answer is a tool you can buy this afternoon
  • Self-hosted against hosted, with the break-even worked out at your volume rather than asserted
  • A sequence, so the first thing you do is the one that proves the approach cheaply and leaves the hard one until you have learned something

What this will not do.

I do not build anything inside this engagement. Three weeks is spent looking, thinking and writing, and that is the whole product. I am not a change management practice and I do not run training programmes. What comes out is a technical answer about your own workflow rather than a strategy narrative. The findings are only as good as the access: if I cannot talk to the people doing the work, I am guessing from an org chart like everyone else. And I will tell you not to do things. If the honest answer for your department is one small automation and a better form, that is what the document says, and you will have paid six thousand dollars to not spend sixty.
$6,000one engagement · three weeks · one department or workflow

One price for the whole thing: the interviews, the system review, the ranking with its numbers, the architecture and model recommendations, the written report and the session where we go through it. Scope is one department or one workflow, plus the systems behind it. A second department is quoted up front as its own engagement, never as surprise scope.

  1. 1

    Interviews and mapping $2,000

    Time with the people doing the work, a read of the systems behind it, and the workflow written up as it actually runs, with volumes attached.

  2. 2

    Candidates, ranked and costed $2,000

    Every candidate with its value, build cost, monthly run cost, failure cost and verdict, including the ones that come out as a no and why.

  3. 3

    Architecture and model picks $1,000

    Per surviving candidate: what the system looks like, which models fit your constraints, build against buy, self-hosted against hosted, and where the human stays in the loop.

  4. 4

    Report and walkthrough $1,000

    The whole thing written up and handed over, then a working session with whoever has to act on it, so the decisions get argued out while I am still in the room.

Each milestone is agreed before it starts and paid only when you approve its deliverable, so you never pay up front for work you have not seen. If what you see at milestone two tells you the answer already, you can stop there and keep it.

Why this is not the usual deck

Written by the person who would have to build it.

The costs are real because I pay them

A monthly run cost is easy to state and hard to get right. Mine comes from having shipped this work and paid the bills: what a retrieval call actually costs at volume, where prompt caching helps and where it does nothing, when a GPU you own beats an API and when that maths is a trap. I build and run a self-hosted AI gateway, so the self-hosting answer is operational experience rather than a slide.

The no is the valuable half

A consultancy whose next invoice depends on you adopting AI will always find a reason to adopt AI. I am paid once, for an answer, and the answer is allowed to be no. Half the value of three weeks is the list of things you were about to fund that should not be funded yet, and exactly what would have to change for that to flip.

It is three weeks, not six months

The long engagement is not more rigorous, it is more people billing. One department, three weeks, one person who has done the work reading your systems directly. If the scope genuinely needs more than that, I say so at the scoping call and quote it as a second engagement, rather than discovering it in month four.

I also sell builds. Here is how I keep that honest.

You should assume a conflict of interest, because there is one: three of my other services are exactly the kind of thing this review might recommend. So the report is written to survive me walking away. Build costs are quoted as what the work is worth with any competent engineer, not as a price from me. Where an off-the-shelf product beats a build, it is named, with what it costs. Where the answer is do nothing yet, it says that. You owe me nothing after milestone four, the document is yours, and taking it to your own team or another shop is a completely normal outcome I have priced for.

The answer is yours to keep.

🔓

Everything written down and handed over

The workflow map, the ranked candidates with their numbers, the architectures, the model picks and the rejects with their reasons. A document your team can act on without me, and that a new hire can read in a year to find out why you decided what you decided. My open-source work is on my GitHub if you want to see how I actually build this stuff before you hire me to have an opinion about it.

🏠

Nothing leaves your building

I work on your systems in place, through access you create and can revoke. No data gets uploaded anywhere for analysis, no documents get run through a model to summarise them, and I sign your NDA before we talk specifics if that is your policy.

🛡

Paid in four phases, never up front

Each milestone is agreed before it starts and paid only when you approve its deliverable. Direct B2B, on your own contract and invoicing, with no platform in the middle.

∞

No retainer, and no obligation

It ends at milestone four. There is nothing to renew, nothing to cancel, and no requirement to hire me to build any of it. If you want me for one of the builds afterwards, that is a separate engagement quoted then.

Straight answers

What does the price cover?

Three weeks in four milestones. Interviews with the people doing the work and a review of the systems behind it; the workflow mapped as it actually runs with volumes attached; every candidate ranked with its value, build cost, monthly run cost and failure cost; an architecture and a model recommendation for each one that survives; the rejects written up with their reasons; and the whole thing delivered as a document plus a working session with whoever has to act on it. $6,000 total, one department or workflow. A second department is quoted up front as its own engagement.

How does payment work?

In four phases: interviews and mapping ($2,000), candidates ranked and costed ($2,000), architecture and model picks ($1,000), report and walkthrough ($1,000). A phase is agreed before it starts and paid when you approve that phase’s deliverable, so you are never paying up front for work you have not seen. If the ranking at phase two already tells you what you needed to know, you can stop there and keep everything produced so far.

What if the answer is that we should not do any of it?

Then that is what the document says, with the reasoning and the numbers that got it there, and with what would have to change for each rejected candidate to become worth doing. That is a real outcome and it happens. Six thousand dollars to find out that your department needs one small automation and a fixed intake form is a good trade against funding a six-figure programme that was never going to pay back. I would rather write that than invent work.

Are you just going to tell us to use ChatGPT?

Sometimes a seat licence and two hours of training genuinely is the answer for a team, and if so you will get that in writing instead of an architecture diagram. More often it is not, because the thing that makes an AI feature valuable is the plumbing around it: what it can see, what it is allowed to do, how it fails, who checks it. That plumbing is what this review is actually about, and no chat subscription supplies it.

How do you choose which models to recommend?

From your constraints, not from a leaderboard. What latency the workflow can tolerate, whether the data can leave your network, what an error costs, and what the thing will cost per month at your real volume. That usually produces a mix rather than one answer: a small or self-hosted model for the high-volume mechanical work, something stronger for the calls that need it, and a hard rule about which is which. The recommendation names what you would have to re-check if you swapped a model later, because you will.

Do you build it afterwards?

Only if you want me to, and it is a separate engagement quoted then. Several of the things this review recommends are services I offer, which is exactly why the report prices builds as what the work is worth with anyone rather than as a quote from me. Handing the document to your own engineers or to another shop is a normal outcome. You are buying an answer, not a funnel.

What do you need from my team?

One scoping call to pick the department and agree what is in scope. Then, during the three weeks: a few hours total with the people who actually do the work, ideally at their desks rather than in a meeting room; whoever owns the systems involved, for read access and a walkthrough; any volume or cost numbers you already track; and one person who can answer the small questions as they come up. Nobody needs to prepare a presentation for me.

Is our data safe during the assessment?

I work in place, through access you create and can revoke when it is over. Nothing gets exported for analysis, and no document, transcript or ticket of yours goes through a model as part of this work. Where I need to see a sample to judge whether something is feasible, it is whatever sample you choose to show me, on your systems. NDA signed before specifics if that is your policy.

We already have some AI running and it is a mess. Is that in scope?

Only what touches the department we are looking at, and only as context for what to build next. Taking an inventory of everything running across the company, costing it and deciding what to switch off is a different job, and it is the AI Stack Audit: two weeks, from $4,000, whole company, a keep, kill or rebuild verdict on every line. If that is the state you are in, do that one first. It is two weeks instead of three, and its output is the list this one would otherwise spend its first week assembling.

What does the deliverable actually look like?

A written document, in plain language, long enough to be useful and short enough to be read: the workflow map, the ranked candidates with a page each, the architectures, the model picks, the rejects with their reasons, and a recommended sequence. Plus a live session going through it with your people, which is where most of the real decisions get made. A document rather than a slide deck, because the detail is the point.

Who is this not for?

Companies that have already decided what they want built and need someone to build it, which is a different conversation and probably a cheaper one. Teams that cannot give me access to the people doing the work, because then I am guessing. And anyone who needs the document to support a decision that is already settled, because it might not.

How do we actually engage?

Direct B2B, on your own contract and invoicing, through whatever procurement process you already have. There is no platform in the middle and no marketplace fee. Reach me on LinkedIn, tell me which part of the business you are thinking about, and we sort the terms directly, including an NDA first if your policy wants one.

Tell me which part of the business hurts.

This is for companies under pressure to do something with AI and short of anyone who can say what, or whether it pays. Tell me which department you have in mind, what the work looks like, and what you already tried. We get on a call, I scope it, and three weeks later you have a ranked and costed answer, including the parts where the answer is no.

Get in touch →

Direct B2B, on your own contract and invoicing. Reach me on LinkedIn and we sort the terms. Fixed price, four milestones, no platform in the middle.