AI Integration & Custom AI Builds, Priced

CH 01The failure rateevery figure with its homework shown

Where AI projects actually die

Four numbers, all from the last year, each printed with who asked, how many they asked, when — and what it doesn’t prove. That last column is the one nobody else prints.

The evidence, with sample sizes and what it does not prove5 rows

Four 2026 datasets on AI project outcomes, each with its source, sample size, fielding date and what it does not prove
What it found	Who asked	Sample	Fielded	What it does not prove
42% of firms had scrapped most of their AI initiatives, up from 17% the year before	S&P Global Market Intelligence, Voice of the Enterprise	1,000+ respondents, North America and Europe	October 2025	Self-reported. "Scrapped most initiatives" is not the same as "AI does not work" — the average organization also scrapped 46% of proofs-of-concept before production, which is roughly what a healthy pipeline looks like.
Only 7% of senior leaders can report established ROI on AI	KPMG Global AI Pulse, Q2 2026	2,145 respondents across 20 countries	Fielded April–May 2026	A survey of leaders, not an audit of accounts. Also from the same wave: only 26% have real-time visibility into what their AI costs to run, which is arguably the more useful number.
74% of enterprises rolled back or shut down a live AI customer agent — 81% among those with fully mature guardrails	Sinch, The AI Production Paradox	2,527 senior decision makers, 10 countries	Fielded January–February 2026	Vendor-sponsored research by a company that sells the governance layer, so the framing is theirs. The sample and the fielding are real and the direction matches two independent datasets.
53% of data practitioners name poor data quality a top challenge; trust in data is now the single most-prioritized objective at 83%	dbt Labs, State of Analytics Engineering	363 practitioners and leaders	Published April 14, 2026	A practitioner survey, self-selected toward people who already use dbt. The internal contrast is the interesting part: 72% prioritize AI-assisted coding against 24% prioritizing pipeline testing and observability.

The number I’m not using

You’ll see a different one quoted everywhere: MIT’s 95%. It’s a non-peer-reviewed paper from July 2025 built on a hundred and fifty interviews, there’s no 2026 follow-up, and by now every AI vendor’s landing page has quoted it into wallpaper.

The newer numbers are bigger, cleaner and worse for the industry. Those are the ones on this page. A page that dates every figure doesn’t get to make an exception for the figure that flatters it.

What actually killed them

None of those projects died because the model was dumb. They died on data nobody had modeled, a workflow nobody had integrated, and an outcome nobody defined before the build started. That’s my opinion and I’ll label it as one.

But it’s the same place the analytics side keeps arriving from the other direction, which is why the fourth row up there is a survey of data practitioners rather than an AI survey at all. Everybody is standing around the same hole. That hole is the whole of the analytics channel, if you’d rather start there.

If you’re reading this because yours is one of them — the pilot demoed fine and then stopped — the door is the Second Opinion, $2,000 flat, back in 5 business days. Somebody with no stake in the original decision reads it against the three named causes of rollout failure and writes you a page. You keep the memo whoever you end up hiring, including not me.

CH 02The limits9 things, and the reason for each

What Dead North won’t build

This is near the top of the page on purpose. It’s cheaper for both of us to find out here than on the third call.

Automated employment decision tools — resume screening, ranking, interview scoring, promotion or discharge recommendations
New York City requires an independent bias audit from the prior twelve months plus public disclosure before one can be used; Illinois made discriminatory AI hiring a civil rights violation on January 1, 2026 and named ZIP code as a protected-class proxy; and private Title VII litigation did not go anywhere when federal enforcement softened.
Agents with unreviewed authority to move money — refunds, credits, trades, provisioning
That is direct-loss exposure carriers are actively exiting. I will build the draft-and-approve version of every one of these, and a named human commits.
Anything touching protected health information without a signed BAA from the model provider
Most model providers will not sign one on standard commercial terms. Roughly two-thirds of US physicians use AI tools and about a quarter of health systems have BAAs covering third-party AI — that gap is real and it is not mine to close for you.
Credit, lending and insurance underwriting or pricing decisions
SR 11-7 model-validation burden lands on the regulated client and flows straight into the statement of work as documentation obligations one person cannot staff.
Customer-facing autonomous support agents
74% of enterprises rolled back or shut off a live one, rising to 81% among the ones with fully mature guardrails. And the buy option is $0.99 to $2.00 a resolution. I would rather integrate what you bought than sell you the wrong build.
Putting anybody’s brand on somebody else’s model, or changing its stated purpose, where the system is high-risk under Annex III
That reclassifies the integrator as the legally responsible provider, and two of the three routes cannot be contracted back. We run the high-risk check at intake; if the answer is yes, I decline rather than carry conformity obligations alone.
Fixed-fee compliance warranties, or an "AI compliance certification"
The underlying state law may be preempted or enjoined mid-engagement — Colorado’s original act was repealed and replaced before it ever took effect. Compliance work is hourly at the published rate, never a fixed-fee deliverable with a warranty attached.
Uncapped indemnities, accuracy guarantees, or any liability cap above the confirmed insurance limit
An indemnity that survives the cap is an uncapped indemnity. Insurance does not fix a bad contract, and I would rather tell you no now than discover it together later.
Deepfakes, synthetic likeness and voice cloning of real people, or generated content designed to avoid disclosure
EU AI Act Article 50 has been enforceable since August 2, 2026 — chatbot disclosure, machine-readable marking, deepfake labelling — with the marking grace period for existing generative systems closing December 2, 2026.

And before the price: buy this instead

What to buy instead, with the vendors’ own list prices

If you want a customer service bot, buy one. Intercom’s Fin is ninety-nine cents a resolution with no platform fee on a helpdesk you already own.

If you want your agents drafting faster from your own help centre, your helpdesk already ships that too — and as of this August Zendesk’s version already reads your Confluence, your Notion, your own articles and your similar solved tickets, and shows the agent which ones it used. Turn it on. It’s a better deal than anything I could build you.

If you want internal search across your documents, the seat prices in that table are the whole argument, and at sixty seats a Glean-class tool is thirty-six to fifty thousand dollars a year, forever. A custom build only starts making sense when volume, data residency, or a deep commerce or ERP integration breaks the subscription model.

Where the line is

Come back when it can’t answer ‘where is order 41823,’ because that answer lives in your ERP and no helpdesk copilot has ever seen it. That part is mine.

Salesforce will happily sell you Agentforce, and their own ecosystem lists the implementation at fifty to a hundred and fifty thousand dollars — the integration bill survives the buy decision every time. I’d rather be the person who wires up the thing you bought than the person who sold you the wrong build.

Published vendor list prices for the products worth buying instead of commissioning a build, each with the date the vendor’s own page was read
Vendor	Product	List price	Worth knowing	Read on
Intercom	Fin	$0.99 / resolution	No additional platform fee for the agent on an existing helpdesk. Intercom seats are separately $29–$132/seat/mo.	2026-08-19
Zendesk	Agent Copilot	$50 / agent/mo	On top of Suite Professional at $115/agent/mo. Already grounds in Confluence, Notion, your articles and similar solved tickets.	2026-08-19
Microsoft	Microsoft 365 Copilot	$30 / user/mo	First-year active use typically runs 30–55% of purchased seats. E3 goes $36→$39 and E5 $57→$60 from July 1, 2026, and E5 now bundles Security Copilot.	2026-08-19
Amazon	Q Business Pro	$20 / user/mo	Lite is $3. Below roughly sixty seats this beats a custom internal copilot, and the page should say so.	2026-08-19
Klevu	Site search	$449 / mo (5,000 products)	The right answer under about 5,000 SKUs with clean data. Print it above the catalog-search price.	2026-08-19

Those are somebody else’s prices, and the last column says the day I read the vendor’s own page. Anything marked third-party came out of a published table because the vendor won’t render its number to a logged-out browser — check it yourself before you budget. I’d rather flag which numbers I borrowed than let you assume I checked them all.

CH 03The shapesa named workflow · a named metric · a named human

The three things Dead North builds

‘Custom AI’ isn’t a thing anyone can buy. These three are, and each one carries the same furniture: one workflow with a name, one metric you can read off a report, and one person who still has to say yes.

Shape one

The desk job

One expensive manual process eaten: documents read, records sorted, fields extracted and written back where they belong.

The metric: Items cleared per hour, and the share routed to a human review queue.
The human checkpoint: Extraction proposes. A named person commits. Nothing pays, posts or approves on its own.
What would make it fail: It fails when the ground truth does not exist — if nobody can produce what a person actually keyed in from those documents, there is nothing to measure against and no honest way to price it.

Priced as Documents Into Your System of Record · $18,500 fixed

Shape two

A second pair of hands

An agentic workflow with an approval gate over a real business system — order-exception triage, where it reads the case, checks the inventory truth in the ERP and drafts the response.

The metric: First-pass draft acceptance, and minutes per item against the human baseline measured before any code was written.
The human checkpoint: A person approves before any refund, credit or re-ship fires. Every classification is logged with its confidence, and there is a one-switch back-out.
What would make it fail: It fails when the deterministic version was always enough and nobody checked. That is why the rules version gets built and scored first, and why there is a stop price if it clears the bar.

Priced as Sorting and Routing on One Pipeline You Already Run · $15,000 fixed

Shape three

The house expert

A domain assistant grounded in your own corpus — permissioned, cited, internal. Never customer-facing.

The metric: Correct and correctly cited, scored against a frozen set of your real questions.
The human checkpoint: It answers with sources or it declines. A refusal is a correct answer and it is scored as one.
What would make it fail: It fails on permissions. If it has to know who is asking, that is a different piece of software and a different budget — which is the whole of the next channel.

Priced as Catalog Search That Knows Your Stock, Your Prices and Your Part Numbers · $15,000 for the first two, then $22,500

Everything that survived a 2026 budget review looks like one of those: one named workflow, one named metric, one human still in the loop. The projects in the table above mostly did not. That’s not a coincidence and it isn’t a sales point — it’s the only reason a fixed price is possible on work whose output is probabilistic.

CH 04Side by sidefind your row, read across

Every AI price, one table

12 rungs, from a $950 licence read to a build with the acceptance numbers written into the contract. Same five questions asked of each one.

The full AI ladder — diagnostics, named builds, the designed build, pilot hardening, the monthly run retainers and the fractional seat — with price, what lands, and the turnaround and credit terms for each
Offering	Price	What lands	Terms
AI License & Spend Audit	$950 flat	A cancel, downgrade or reassign list with the annual dollar figure and the renewal date it has to happen by	Terms When 5 business days from access Credits toward Any AI or data engagement, 120 days
AI Integration Readiness	$3,500 flat	A system-of-record map, a written permission model, three scored workflows, and a seed eval set with verified answers	Terms When 10 business days from access Credits toward Any AI engagement, on the same terms as every other diagnostic
AI Test Pattern	$7,500 flat	A measured accuracy number on your own data, a cost per run, and a recommendation that may be “do not build this”	Terms When 10 business days Credits toward Any build, in full
Second Opinion — on an AI proposal, or a pilot that stalled	$2,000 flat	A written memo: what is missing, what is padded, and the three questions to ask before you sign	Terms When 5 business days Credits toward Any follow-on, 30 days
Sorting and Routing on One Pipeline You Already Run	$15,000 fixed	Orders, RMAs, warranty claims, supplier and AP email, inbound leads that must land on a CRM record	Terms When 3–4 weeks Credits toward $7,500 Test Pattern, in full
Documents Into Your System of Record	$18,500 fixed	Packing slips, dealer and wholesale order forms, supplier spec sheets, remittance advice, bills of lading, warranty cards	Terms When 4–5 weeks Credits toward $7,500 Test Pattern, in full
Catalog Search That Knows Your Stock, Your Prices and Your Part Numbers	$15,000 for the first two, then $22,500	One store, one catalog source, up to about 25,000 SKUs, English, hybrid BM25 + vector retrieval fused with RRF	Terms When 4 weeks from data access Credits toward $7,500 Test Pattern, in full
AI build — anything that has to be designed	$25,000 – $75,000	A build that is none of the named shapes above, with the acceptance thresholds agreed and signed before any model work starts	Terms When 6–12 weeks Credits toward $7,500 Test Pattern, in full
Clear to Air — a stalled pilot into production	$14,500 – $26,000	Evals and an Acceptance Set built from real cases, guardrails, cost caps with alerting, the permission model, observability, rollback and a runbook	Terms When 3–5 weeks Credits toward $2,000 Second Opinion, 30 days
Confidence Monitor — per production AI system	$1,250/mo	The Acceptance Set re-run weekly and the scorecard mailed whether it is green or red, model-deprecation watch, a prompt and configuration changelog, metered spend with a hard cap and an alert before it, and a quarterly re-grade of twenty-five fresh live cases. Each additional system, $750/mo.	Terms When Starts at go-live · required for the first 90 days Credits toward Monthly — nothing to credit
Confidence Monitor+ — per production AI system	$3,250/mo	Everything in Confidence Monitor, plus six hours a month of prompt and retrieval iteration, model migrations executed rather than only planned, and a four-hour response on an eval-red or cost-spike alert	Terms When Starts on the next cycle Credits toward Monthly — nothing to credit
Fractional lead — Advisory — AI & data	$6,500/mo	Roadmap ownership for the AI and analytics side, vendor and proposal oversight, buy-versus-build calls with the arithmetic shown rather than asserted, and a model and data inventory that stays current. Every decision written down and yours to keep, including the ones where the answer was to buy something off a shelf.	Terms When Half a day a week Credits toward Monthly — nothing to credit

For context, since the point of publishing is that you can compare: ClearForge publishes $15,000 for a two-week AI diagnostic and MLDeep publishes $10,000 for the same fortnight; MLDeep’s own page puts the Big Four equivalent at $50,000 to $150,000. The readiness rung here is $3,500 flat. That isn’t a sale — it’s the same arithmetic the Performance Audit already runs on this site: a fixed scope is less risky for me than open-ended work, so it costs you less. If it takes me longer than I thought, that’s my problem and not your invoice.

The named builds, in full

These are fixed price because the shape is known — one store, one document type, one pipeline, one system — which is why they hold the site’s existing $15,000 build floor rather than opening above it. Each one starts by telling you when to buy something instead.

Each build in full — scope, exclusions, add-ons3 builds

Sorting and Routing on One Pipeline You Already Run

$15,000 fixed

3–4 weeks

Stop price at the week-two gate · $7,500

The sequence is the product, so it goes on the page in order: pull twelve months of your history and label a set, measure what a person currently achieves in accuracy and minutes per item before anybody writes code, build the deterministic rules version first and score it, and only add a model if the rules miss. If the rules clear the bar at the two-week gate we stop there and you pay $7,500. That is the whole mechanic and there is no negotiating it later. Sometimes the rules are enough — I do not know how often, because I measure it per client instead of quoting a number I cannot source. What I can promise is the gate and the stop price.

What it covers: Orders, RMAs, warranty claims, supplier and AP email, inbound leads that must land on a CRM record
What it does not: Helpdesk tickets — Zendesk Intelligent Triage already does that at $50/agent/mo. Anything sorting people into employment outcomes, ever.

Second workflow on the same rails — same source system, same review queue, same auth · $7,500

Documents Into Your System of Record

$18,500 fixed

4–5 weeks

If your documents are supplier invoices or purchase orders going into NetSuite, Sage Intacct, QuickBooks Online, Xero or Acumatica, do not hire me for this. Ramp syncs two ways with all five for $0 to $15 a user a month. Bill.com runs $45 to $79. Stampli’s coding and PO matching is trained on more invoices than either of us will ever see. And if you only need fields read off a page, Microsoft’s prebuilt invoice model is $10 per thousand pages with the first five hundred free. Go buy one of those. I am useful when the document is odd, when the destination has no connector for sale, or when the write-back has to respect business rules that currently live in somebody’s head. Extraction proposes; a named human commits. I will not build you something that pays, posts or approves on its own.

What it covers: Packing slips, dealer and wholesale order forms, supplier spec sheets, remittance advice, bills of lading, warranty cards
What it does not: Supplier invoices and purchase orders headed into a mainstream ERP — buy a connector, see the blurb. Claim forms, always.

Additional document type — same destination, same source family, ≤20 fields, ≤4 pages · $7,500
Second destination system — a second auth model, schema, idempotency contract and audit trail · $6,500 – $9,500

Catalog Search That Knows Your Stock, Your Prices and Your Part Numbers

$15,000 for the first two, then $22,500

4 weeks from data access

Under about 5,000 SKUs with clean product data, buy Klevu at roughly $449 a month and skip me. Algolia’s semantic tier is Elevate-only and starts north of $50,000 a year, so if somebody quoted you Algolia for semantic search, check which plan you were quoted. On WP Engine Core or Enterprise you may already be paying for Smart Search AI — check before you call. A custom build wins above roughly five to ten thousand SKUs, or when stock and price truth lives in an ERP rather than the store, or when results have to respect per-role permissions, because that is the part no subscription can see. And this is retrieval and ranking, not a chatbot: nothing here is written by a model, which keeps it clear of the EU’s Article 50 disclosure and marking rules entirely. The day you want an answer summary on top, that is a different price and a different risk profile, and I will say so.

What it covers: One store, one catalog source, up to about 25,000 SKUs, English, hybrid BM25 + vector retrieval fused with RRF
What it does not: Multi-store, multi-language, PIM or attribute normalisation, and any generative answer layer

And when it has to be designed

A build that isn’t one of those shapes opens at $25,000 instead of $15,000, because the design is the expensive part. Here is what moves the number, printed rather than discovered. I have not found another shop in this market that publishes this table at all.

How a designed build runs, gate by gate7 steps

The published steps that change the price of a designed AI build, with the amount each one adds
The step	What it adds
Base — one workflow, one system, one user group, no per-user permissions, no citations required	From $25,000
Per-user permission enforcement	+$12,000 – $25,000
Source citation and provenance on every answer	+$8,000 – $15,000
Each additional business system integrated	+$9,000 – $18,000
Regulated data handling — audit logging, PII redaction, residency, a signed BAA in place	+$20,000 – $40,000

Typical lands at $25,000 – $75,000, and it is only sold after a AI Test Pattern that produced a real number. Not as an upsell — because the pilot is how the price gets found, and quoting a designed build off a discovery call is guessing with your money.

CH 05Permissionssettle this before you spend anything

The step that triples the price

A retrieval assistant over one clean source runs $25,000 to $50,000 across this market. The moment it has to respect who’s asking and cite where every answer came from, the same market prices it at $80,000 to $180,000. That isn’t a markup. It’s a different piece of software.

So settle one question before you spend anything: does this answer everybody the same way, or does it have to know who’s asking? If it’s the second one, budget for the second one. Getting that wrong is the most expensive mistake in this whole category — and permissions, roles and structured content happen to be thirteen years of my life, so it’s also the question I’m least likely to get wrong on your behalf.

Above roughly eighty thousand dollars is more than I take on alone. If that’s where you land, say so and I’ll tell you who does it well. That’s a shorter conversation than finding out in month four.

The $80,000 figure is the market’s, not mine — it’s the published band for permissioned, cited retrieval, and it’s on this page as somebody else’s number so you can see where the ceiling of a one-person shop actually sits. The AI Integration Readiness exists mostly to answer this one question in writing, $3,500 flat, before anybody quotes a build on a guess.

CH 06Acceptancewhat ‘done’ means, in numbers

What happens if it misses its numbers

A deterministic feature is finished when it does the thing. A probabilistic one isn’t, and every fixed-price AI contract that pretends otherwise ends in an argument about whether a wrong answer is a bug. So we settle it in week one, before any model work.

The Acceptance Set

A frozen batch of your own real cases with the correct answers graded by your subject-matter expert, not by me. I build the harness, you grade and sign. That set is the contract’s definition of correct.

Then three numbers go in the statement of work — how often it’s right, how often it fails safely into a review queue instead of answering confidently wrong, and what one run costs at the ninety-fifth percentile. The safe-failure number is always set higher than the accuracy number, because a wrong answer is a cost and a confident wrong answer is the defect.

And if it misses

I work to threshold at no additional fee for up to three weeks. If it still misses, we stop. You keep the code, the Acceptance Set and the eval harness, and the final quarter of the fee is never billed. I’d rather eat twenty-five percent than argue with you about whether a wrong answer is a bug.

One more thing, because it’s the honest part: the eval harness runs without me. If you fire me, you can still tell whether it’s working.

The pass bands and how the money moves2 tables

The bands, published

Honest per-shape pass-rate bands agreed at signing, with the safe-failure floor
Shape	Pass rate
Extraction and classification	92–97%
Agentic workflow with human approval — first-pass draft acceptance	85–93%
Grounded assistant — correct and correctly cited	80–90%
Safe failure — declines into a review queue rather than answering confidently wrong	≥98%

Set at signing rather than discovered in month four. The acceptance event is three consecutive green runs on three different days, then one live shadow week where the system runs alongside the human process and nothing it produces gets acted on.

How the money moves

The three payment milestones on an AI build and the share released at each
Milestone	Share
Kickoff, released on delivery of the signed Acceptance Set and eval harness	35%
Build complete	40%
Acceptance	25%

That last 25% is genuinely at risk on the three numbers, which is the only version of a probabilistic fixed price a finance department should ever sign.

The eval suite is yours, in your repo

Not a slide about evals — a bill of materials that lands in your repository and runs on your CI:

The scorers, one per number in the statement of work
Permission-leak probes, if the system knows who is asking
A red-team set drawn from the OWASP Top 10 for LLM Applications
A regression runner wired into CI, so a prompt change can fail a build
A per-run cost and latency report
A one-page scorecard your team can run forever

And a refusal that belongs here rather than in month four: if your process needs 99.5% and cannot tolerate a review queue, this isn’t an AI problem. I’ll tell you that in the AI Test Pattern — $7,500 flat, and a recommendation that may be ‘do not build this’ — not after you’ve spent a build budget finding out.

CH 07Running itthe margin printed, the token bill yours

What it costs after it’s built

The market’s own estimate is that a first year of running an AI system costs forty to eighty percent more than building it. On a $60,000 build that’s $24,000 to $48,000, and nobody puts it in the proposal.

Confidence Monitor

$1,250/mo

Per production AI system. The Acceptance Set re-run weekly with the scorecard mailed whether it’s green or red. Model-deprecation watch with a written migration plan. A prompt and configuration changelog with version pinning. Token spend metered with a hard cap and an alert before it. A quarterly re-grade of twenty-five fresh live cases, so the test set doesn’t go stale. One page a month carrying cost per resolved item.

Each additional system · $750/mo · required for the first 90 days after go-live

Confidence Monitor+

$3,250/mo

All of that, plus six hours a month of prompt and retrieval iteration, model migrations executed rather than only planned, and a four-hour response when an eval goes red or the spend alert fires. This is the tier for a system somebody’s week depends on.

4 of 4 seats open · one person, real limits

How the monthly actually works — margin, account ownership, levers4 notes

The margin, printed

The eval and observability stack under that retainer costs $29 to $199 a month — Langfuse Core / Pro, and the comparable seat-plus-traces pricing next door. Against $1,250/mo. You’d have found that in twenty minutes anyway, so here it is from me first.

What the retainer buys is availability and ownership, not tooling. And a published exit: if your system runs ninety days green with no changes, I’ll tell you to drop to Sound Check and call me when something goes red. A year of Confidence Monitor is $15,000 with the model bill sitting on top and paid by you directly, which is the next paragraph.

You hold the account

The provider account is yours, in your name, with your card on it. I never mark up tokens and I never resell inference. Four reasons, all of them yours rather than mine:

You see the real bill, daily, without asking me for a number.
The provider’s commercial terms — retention, training, residency — run to the account holder. That’s you, so they’re yours to negotiate and yours to read.
A marked-up token is a quiet incentive to write a chattier system. I’d rather not have one.
If you fire me, the system keeps running. That’s the point.

Why there’s no token price here

Every other number on this site is held to a published date. Token prices move faster than that — they moved several times this year, in both directions — so printing one here would be the only figure on the site with a shelf life measured in weeks.

What you get instead is a cost per run measured on your own records and dated to the day it’s delivered, out of the AI Test Pattern, plus a ninety-fifth-percentile figure with a published cap written into the build contract. A number from last quarter isn’t a budget, it’s a decoration.

The levers, named

When the bill needs to come down, these are the five things I actually pull, and pulling them is part of the retainer rather than a change order:

Prompt caching on the stable half of the context
The batch endpoint, at roughly half price, for anything that doesn’t need an answer this second
Model routing — the small model does the easy nine-tenths
Context editing and compaction on long agentic loops, which are where a naive estimate goes ten-fold wrong
A per-session dollar budget, so one runaway loop is a capped incident and not a phone call from your CFO

One thing you will not find on this page: a promise that your data never leaves your infrastructure, or a blanket zero-retention claim. Zero data retention isn’t a standard-plan feature at any major provider — it takes a negotiated agreement. So I quote your provider’s actual terms verbatim in the statement of work, or I say nothing. Paraphrasing somebody else’s contract back to you is how that sentence goes wrong.

And the rate underneath all of it hasn’t moved: $250 an hour for everything on this site, AI included. There’s no AI surcharge and no AI discount — the long version of why is in the straight answers below, and the rate card itself is on the services page.

CH 08The proof boarddated, and it says so when a date slips

Where the proof stands

There is no shipped client AI product on this shelf yet. This practice opened in August 2026. I would rather tell you that here than have you find it out on a call. I’d rather show you what is real than dress up what isn’t.

What’s real

Thirteen years on the exact line item every AI budget prices highest — order feeds, ERP and CRM sync, SSO, inventory pipelines, the APIs in between. Enterprise AI builds carry fifteen to thirty percent more per integrated CRM, ERP, ITSM or identity system, so that’s the expensive part of the bill, not the commodity one.
Two contrib modules on drupal.org with a public install count and a public issue queue, which is a harder credential than a case study because anybody can go read the bug reports.
AWS Solutions Architect and Cloud Practitioner, Azure alongside them, and the CS degree finished at NDSU in 2024, eleven years into the career it describes.
And this site, where every price is published, the arithmetic checks, and the carbon rating is public and re-runnable.

What isn’t, yet

No shipped client AI product. No published Power BI engagement. Both of those are real gaps and neither one is fixed by a paragraph, so below is what would change your mind and roughly when it lands.

If a date slips, this table will say it slipped rather than quietly changing. That is the entire value of publishing it, and it is the reason it’s uncomfortable to publish.

The proof board — every claim and its status5 rows

The artifacts that would evidence this practice, each with its current status and target date
Artifact	Status	Target
Agent write governance for Drupal Commerce — approval workflow, audit trail, permission scoping, rollback — released as a contrib module, built to the 2026-07-28 MCP spec	In build	October 2026
AI Search plus a scored eval suite on deadnorth.io’s own catalog, published split by query class, with our monthly token bill shown as a worked example rather than a quote	In build	September 2026
A Power BI semantic model and eval set on Dead North’s own operational data, published with the licensing decision and the reasoning shown	Queued behind PL-300	November 2026
The first client AI engagement, published with before-and-after numbers and the client’s name on it	Not started	Q1 2027

If you’d rather not wait for a table: send me your schema and I’ll spend thirty minutes on a screen share telling you what’s wrong with it. That’s a qualifier, not an engagement — it gets you a real read on whether your data is ready and whether I’m the right person, and it does not get you an audit. The audit is the $3,500 flat one. I’m telling you which is which because the free-audit thing in this market is a sales call wearing a lab coat.

CH 10Straight answers6 questions, asked in this order every time

The questions you actually arrive with

You don’t have an AI case study — why would I hire you?

Straight answer, because you’d find out anyway: there is no shipped client AI product on this shelf yet. This practice opened in August 2026. What is real is thirteen years on the exact line item every AI budget prices highest — order feeds, ERP and CRM sync, SSO, inventory pipelines and the APIs in between. Enterprise AI builds carry fifteen to thirty percent more cost per integrated CRM, ERP, ITSM or identity system, so that is not the commodity part of the bill, it is the expensive part. Buyers agree, for what it is worth: INFUSE, Voice of the Buyer AI Reality Check found proven integration ranked first and capability depth last. The proof board on this page names what is in build and when it lands, and if a date slips the table will say so rather than quietly changing.

Is my data used to train a model?

Not by me, and I am going to be careful about the rest of that answer, because it is the one sentence on this page a curious buyer can catch. You hold the provider account, so the commercial terms governing your data run between you and the provider rather than through me. Zero data retention is not a standard-plan feature at any major provider — it takes a negotiated agreement, and some model families are reported excluded even then. So I will not hand you a blanket promise that your data never leaves your infrastructure, because I would be paraphrasing somebody else’s contract. What I will do is read your provider’s terms with you, quote them verbatim in the statement of work, and design the boundary so the data that never needs to leave, does not.

Do you charge more per hour for AI work?

No, and there is no AI discount either — I would rather explain than just say no. The rate is $250 an hour for everything on this site, AI included. The measured premium in this market is twenty to forty percent on labour but three to ten times on packaged fees, so the premium belongs in fixed scope rather than in a rate card, and a second published rate would break the one-rate story the whole price list is built on. Going the other direction: the number I quoted already has the method in it. When AI made this work cheaper to produce, that showed up as a $15,000 build floor holding steady while the scope inside it grew, not as an invoice that shrinks after you have signed. Somebody surveyed a hundred and eighty-one agencies about exactly this — 73% had never been asked for an AI discount at all, and of the ones who were, 13% cut. What you get is a fixed scope, a fixed number and a date. How I hit it is my problem.

Should we just buy Copilot?

Often, yes. Microsoft 365 Copilot is $30 a user a month and Amazon Q Business is $20, and below roughly sixty seats either one beats a custom internal assistant on every axis that matters. The catch is that first-year active use typically runs thirty to fifty-five percent of the seats a company buys, which is the whole reason the $950 AI License & Spend Audit exists: I read the admin centres, count who has a seat and has not opened it, and write one page your CFO can act on. Book it sixty to ninety days before your renewal, because annual terms allow no mid-term seat reductions. Come back for a build when the thing you need answered lives in your ERP, because no copilot has ever seen that.

What if we’re in the EU?

Then we flag it at intake rather than discovering it in month four. Article 50 of the EU AI Act — chatbot disclosure, machine-readable marking of generated content, deepfake labelling — has been enforceable since August 2, 2026, with the marking grace period for existing generative systems closing December 2, 2026. Article 2(1)(c) is output-based, so it can reach a US build whose output is used in the EU. What I will not do is sell you urgency about the high-risk tier: Regulation (EU) 2026/1744 moved Annex III obligations to December 2, 2027 and Annex I to August 2, 2028, so anybody quoting you an August 2026 high-risk deadline is reading last year’s page. I flag the exposure and name the mechanism. Your counsel rules on it, and compliance work is hourly at the published rate — never a fixed-fee warranty, which is on the refusal list.

Our AI project stalled — can you make it work?

Maybe, and the cheapest way to find out is not a build. Start with the $2,000 Second Opinion: I read the proposal, the architecture and the number line by line against the 2026 published market, or read the stalled pilot against the three named causes of rollout failure — unclear success criteria, not enough access to tools and data, and evaluation coverage that drifted. A written memo comes back in 5 business days, and you keep it whoever you end up hiring. If the thing is worth saving, Clear to Air hardens a stalled pilot — mine or somebody else’s — for $14,500 – $26,000.

Three doors, and none of them is a build

Nobody should start here with a build. Every one of these credits toward the work if you go ahead, and every one of them is a written thing you keep whoever you end up hiring.

AI License & Spend Audit

$950 flat

You bought AI seats in 2024 or 2025 and nobody can say what they did. A cancel, downgrade or reassign list with the annual dollar figure and the renewal date it has to happen by.

The whole scope →

AI Integration Readiness

$3,500 flat

You want to build something with AI and nobody has read your systems. A system-of-record map, a written permission model, three scored workflows, and a seed eval set with verified answers.

The whole scope →

Second Opinion

$2,000 flat

For an AI proposal you can’t judge, or a pilot that stalled and nobody can say why. A written memo: what is missing, what is padded, and the three questions to ask before you sign.

The stalled-pilot route →

Start a conversation →The data half of this →Every price on the site →

Not sure any of it applies to you? Write me three sentences about what you’re actually trying to do and I’ll tell you which rung to buy, or that a subscription in that table above already does it and you should go turn it on. That part’s free and it’s staying free.