Nobody is coming to implement it for you

There is a version of the stalled AI project that almost everyone runs into once. The tool works in the chat window. You have watched it do the thing. Six weeks later it is still not part of how the work actually gets done, and the instinct is to go looking for a better model, a different subscription, one more comparison video.

That instinct is aimed at the wrong thing, and the money in the industry has already moved on. This guide is about the part that comes after you pick a tool: how to tell an implementation problem from a model problem, how to write the job down so someone else could do it, and how to decide whether to hire that someone or accept that it is you. It assumes you are running a business of one to about ten people and already pay for at least one AI tool. It is not for a team with engineers on staff, and it is not for someone who has not used these tools enough to have an opinion about them.

The capital moved to the part after the model

In May 2026 Anthropic launched Ode with Anthropic, a $1.5 billion AI implementation company, as a joint venture with Blackstone, Hellman & Friedman and Goldman Sachs [2]. It was built on the acquisition of a startup called Fractional AI, which ended an 11-month partnership with OpenAI when it was bought, and it runs around 100 engineers. Anthropic and OpenAI have both spun up separate businesses dedicated to deploying AI engineers to their customers’ offices [2]. Ode describes itself on its own site as an end-to-end partner that helps companies turn frontier models into measurable business results, working from roadmap through deployment [1]. It competes with OpenAI’s own services arm, The Deployment Company, and with the forward-deployed engineering teams that Deloitte and Accenture have built [2].

None of that is a service you will buy. Ode’s chief executive, Chris Taylor, describes the ideal customer as one where the AI work is among the top one or two priorities for the chief executive of the company [2]. The signal matters anyway, because of what it says about where the difficulty sits. Ode’s chief technologist, Eddie Siegel, put it this way: “I think model selection matters, but it’s not where the majority of calories are spent” [2]. Serious institutional money agreed with him strongly enough to build a company on the sentence.

Look at the other half of that trade and it is obvious why. Model capability is a published price list. Anthropic sells Claude Haiku 4.5 at $1 per million input tokens and $5 per million output, Claude Sonnet 5 at $2 and $10, Claude Fable 5.1 at $10 and $50 [3]. OpenAI’s list runs the same shape, gpt-5.6-luna at $0.10 and $0.60, gpt-5.6-terra at $1 and $6, gpt-5.6-sol at $2 and $10 [4]. Anyone can buy either page this afternoon with a card. The scarce thing is not on those pages. It is the labour of connecting one of them to a specific business, and that labour is expensive enough to build a billion-dollar company around.

Test which problem you actually have, in an afternoon

Before you spend anything, find out whether the model is failing or the setup is missing. The test is manual and it takes about two hours.

Collect ten real cases from the last month. Real ones, with the messy names and the missing fields and the one that arrived as a photo of a receipt. Open a chat window and do the task yourself, ten times, pasting in everything you would have to know to do it by hand. Judge each result the way you would judge a new hire’s first week.

If the model gets eight or more of the ten right while you are feeding it, you do not have a model problem. You have an assembly problem: the inputs it needed were in your head, your inbox and two other systems, and nobody has built the path that brings them together without you. If it fails most of the ten even with you sitting there supplying context, then either the task is genuinely beyond current models, or the task is not one task. Split it and run the test again on the smaller piece.

This distinction decides everything downstream. An assembly problem is buyable work with a defined end. A capability problem is not, and no contractor can sell you a way around it, though several will try.

Implementation is five written artifacts

The reason implementation feels vague is that it is usually described as a feeling. It is not. At your size it comes down to five things written in plain sentences, and until they exist there is nothing for anyone to build.

The trigger: what starts the job, in observable terms. A message arriving in one specific inbox. A row appearing in one specific sheet. Not “when a client asks for something”.

The inputs: what has to be in front of the model, and where each piece lives right now. This is where most projects quietly die, because “the client’s history” turns out to mean three systems that spell the client’s name three ways.

The rules and the exceptions: the sentences that exist only in your head. Which two customers must never receive an automated reply. Which product line you refuse to quote by email. What “urgent” means when your largest account says it, versus everyone else.

The destination: where the output goes and who sees it before a customer does. A draft in your inbox is a different project from a message that sends itself, and it costs a different amount.

The failure behaviour: what happens on the case the system does not recognise. Silence is the wrong answer. Something has to land somewhere a human looks.

That list is not a bureaucratic exercise. NIST published its AI Risk Management Framework as version 1.0 in January 2023, intended to be voluntary, and organised the whole approach around four functions, Govern, Map, Measure and Manage. Map is the one that establishes the context to frame risks related to an AI system, and it asks that intended purposes and the prospective settings in which the system will be deployed are understood and documented [7]. That is the same work, at a different scale. You can write yours in an hour, on one page, because you already know all of it.

Write the spec before you ask anyone for a price

With those five artifacts on a page you can do something you could not do before: get comparable quotes. Send the same page to three people and the replies become legible, because they are all pricing the same job rather than describing their own capabilities.

Scope the first engagement to one process and one milestone with a fixed price. Not a retainer, not an audit, not a strategy phase. One process that runs end to end on your real data, delivered by a date, for a number agreed in advance. If the person cannot price one narrow process at a fixed number, that is information about how well they understand it.

Insist that the deliverable includes the boring half: the written spec as built, the credentials list, and instructions for changing the thing without them. Automation platforms make this reasonable to ask for. Zapier’s free tier runs 100 tasks a month and its Professional plan starts at $19.99 a month [5]; n8n’s Starter plan is €20 a month billed annually for 2,500 workflow executions [6]. The tooling is not where your money goes. The labour is, which means an engagement that leaves you unable to touch the result without rehiring the same person has sold you the cheap half and kept the expensive half.

What to check before you hand anyone your accounts

Ask what they have already put into production, specifically, and for whom. Not which models they know. A demo proves the model works, which was never in question.

Ask how they choose between models and what happens to that choice if the recommendation stops suiting them. Ode is open about being close to Anthropic’s applied AI team [1], which is a legitimate position honestly stated. The problem is not preference, it is undisclosed preference. Someone whose income depends on one vendor should say so before you ask.

Keep the accounts in your own name and give access rather than ownership. This is a five-minute administrative act with a large consequence. Anthropic’s Admin API lets an organisation manage its members, workspaces, invites and API keys programmatically, including renaming and deactivating keys, with spend limits available on Claude Enterprise [8]. The practical version for a small business: you create the account and the payment method, you create a separate workspace for the contractor, and when the engagement ends you deactivate that key rather than negotiating for access to something built on someone else’s login.

Agree what “done” means in the same sentence as the price. Done is the process running on your real cases for two weeks with the failure route working, not a walkthrough on a call.

Do the arithmetic before you sign anything

Implementation help is priced against enterprise budgets because that is who buys it. Whether it makes sense for you is a question of hours, not of enthusiasm.

Count the hours the process currently takes each week. Multiply by what an hour of your time is worth, honestly, which for most solo operators is closer to their billable rate than to their salary equivalent. Then work out how many months of that saving it takes to cover the quote. If the answer is more than about twelve months, the sensible move is usually to do the narrow version yourself first, badly, and learn what the exceptions are before paying anyone to encode them.

The number this produces is a floor, not a forecast. It ignores the hours you will spend answering the implementer’s questions, which are considerable and unavoidable, because the answers exist only in your head. Budget a day of your own time for every week of theirs and the arithmetic gets more honest.

What still goes wrong

The most common outcome is not a bad contractor. It is a good one who finishes, hands over something that works, and watches it decay. Processes drift, a supplier changes their invoice format, a model version is retired, someone renames a folder. Whatever you buy needs an owner afterwards, and at your size that owner is you. If nobody in the business can open the automation and read what it does, you have bought a dependency rather than a capability.

The second failure is scoping to the interesting process instead of the frequent one. The task you most want automated is often the one with the richest judgement in it, which makes it the most expensive to specify and the most likely to produce a result nobody trusts. The dull, high-volume, low-stakes process is where a fixed-price engagement actually lands.

The third is more uncomfortable. Some of what looks like an implementation problem is an operations problem wearing a costume. If your client records are inconsistent, your intake arrives on four channels, and nothing is written down, an implementer’s first month will be spent fixing that at their day rate rather than yours. That work is worth doing. It is just not AI work, and paying AI prices for it is a choice you should make on purpose rather than by accident.