How Is AI Implemented?

In nine steps, and five of them happen before anybody writes a prompt. I read the five most cited published frameworks this morning. Between them they name 28 stages, 3 of the 5 open with the business problem rather than the technology, and not one of the 28 uses the word integration.

I am Hayat Amin. I spent twenty years as a technology chief financial officer and sold three companies in that seat, and I now build this work inside other people's companies rather than writing a report about it, which means I have signed off the invoices and then had to make the thing run. On 18 September 2026 I read Microsoft's Cloud Adoption Framework for AI, Google Cloud's AI Adoption Framework whitepaper, the AWS Well-Architected Machine Learning Lens, the NIST AI Risk Management Framework and CRISP-DM, which is thirty years old and still the one most people are quietly using. Below is the order they agree on, the step only one of them names, and the step none of them names at all.

How I counted

I read five documents end to end and wrote down every stage each one names, in its own words. Microsoft names 6 phases of AI adoption: Strategy, Plan, Ready, Govern, Secure, Manage. Google Cloud names 6 themes, Learn, Lead, Access, Scale, Automate and Secure, sitting across four areas it calls people, process, technology and data, plus 3 maturity phases it calls Tactical, Strategic and Transformational. The AWS Machine Learning Lens, published on 19 November 2025, names 6 lifecycle phases: business goal identification, ML problem framing, data processing, model development, model deployment, model monitoring. The NIST AI Risk Management Framework, released on 26 January 2023, names 4 functions: Govern, Map, Measure, Manage. CRISP-DM, first published in 1996, names 6: business understanding, data understanding, data preparation, modeling, evaluation, deployment.

That is 28 stage names. Lining them up leaves 9 steps that actually happen in a company, in the order they happen. I have put the tally against each one, so you can see which steps the industry agrees on and which it has quietly left to you. The order below is chronological, not a ranking of importance, and the two steps that decide whether the project pays for itself are 4 and 8.

1. Name the business problem before you name the technology

Named by 3 of the 5. Microsoft opens its whole framework with use case identification, and its first instruction is to start with business problems: look for where the organisation needs better results before you consider AI at all, and it suggests two plain questions, where do results miss expectations and where do people spend time on repetitive tasks. AWS goes further and calls business goal identification the most important phase of the machine learning lifecycle. CRISP-DM has said the same since 1996 and calls it business understanding.

Google and NIST do not open here, and their reasons are fair. Google opens with Learn, which is about skills, and NIST opens with Govern, because it is a risk framework rather than a build guide. What this step is good for is killing bad projects in the first hour, for free. Where it goes wrong is when someone arrives with a tool and works backwards to a problem, which is most of the meetings I get invited to.

2. Write down the number you will be judged on

Named by 3 of the 5. AWS is the most explicit: you must be able to measure business value against specific business objectives and success criteria, and you should evaluate your organisation's ability to move toward that target before you start. Microsoft asks you to turn each business problem into a short statement naming the activity and the expected result, and gives the example of helping support agents answer from internal documents so resolution time drops. CRISP-DM puts success criteria inside business understanding.

One number, written before the build, owned by a named person. Invoices posted per day. Hours from enquiry to quote. Percentage of orders that need a human to retype them. If you cannot say what the number is today, step 2 is not finished, and everything after it is decoration. This is the cheapest step on the list and the one skipped most often.

3. Decide whether the answer has to be the same every time

Named by 1 of the 5, and it is the most useful thing Microsoft publishes. Its guidance splits the choice into generative AI, which it calls non-deterministic and which produces outputs that can vary even for the same input, and nongenerative AI, which it calls deterministic and which produces consistent repeatable outputs from structured inputs. It tells you to lean generative when you do not know the exact steps in advance and some variation is acceptable, and deterministic for tasks that depend on accuracy, such as prediction or anomaly detection.

I would put this first if it were my framework. A VAT calculation, a credit limit and a payroll run must be identical every time, so a language model does not belong in the middle of them. Reading a supplier email and working out which of your 40 product codes it means is the opposite. Most failed projects I have looked at put the two the wrong way round, then blamed the model.

4. Find out whether the data exists and whether you can reach it

Named by 4 of the 5, and CRISP-DM spends 2 of its 6 phases here, data understanding and data preparation. AWS calls it data processing. Microsoft makes data strategy the fourth of its five strategy decisions and lists data needed as one of four factors you weigh on every solution, alongside capabilities, skills required and cost. Google calls its theme Access, which is the honest name, because the question is rarely whether the data exists.

In a company of 20 to 500 people the data almost always exists and is almost always in four places that do not talk to each other. Reaching it means credentials, an API that your version of the software actually has, and someone with the authority to say yes. That is a diary problem, not a technical one, and it is why I ask for credentials day in week one. Budget real weeks for this step. Every plan I have seen that came in late came in late here.

5. Choose build or buy, from four real options

Named by 1 of the 5. Microsoft is alone in setting out the build or buy ladder, and it names four adoption models that trade customisation for simplicity: ready to use Copilots, low code SaaS development, managed platform development, and Azure infrastructure. Its own summary is that as you move from the first to the last you gain control and give up speed. The named products at each rung are Microsoft 365 Copilot and the in-product and role based Copilots, then Copilot Studio, then Microsoft Foundry Agent Service, Azure Machine Learning and Microsoft Fabric, then Azure Virtual Machines, Azure Kubernetes Service and Azure Container Apps.

Read that ladder as a warning rather than a menu. Most companies under 500 people belong on the first two rungs and get sold the third. Rung one is wrong for you when the work crosses systems, because a Copilot inside one application cannot fix a handoff between four. Rung four is wrong for almost everybody, and Microsoft says so plainly: it takes the longest to build and carries the most operational ownership. If you are not Microsoft's customer, the same four rungs exist under other names.

6. Build the narrow version against the real system

Named by 3 of the 5. CRISP-DM calls it modeling, AWS calls it model development, and Microsoft carries create an AI proof of concept as one of the 6 items on its enterprise plan checklist, alongside assess AI skills, acquire AI skills, access AI resources, prioritise AI use cases and implement responsible AI. All three mean the same thing. Build the smallest version that answers the one question from step 1.

The word I would add is against. Build it against the real system, with the client's credentials, in their environment, on their messy live data. A pilot built on an exported spreadsheet proves nothing, because the spreadsheet is the part a human already cleaned. That is the whole argument for a forward deployed engineer instead of a vendor demo, and it is why the demo that dazzled you in the sales meeting has a failure rate.

7. Test it against the number from step two

Named by 3 of the 5. CRISP-DM gives evaluation a whole phase, second to last. AWS folds evaluation into model development and adds model monitoring at the end. NIST makes Measure one of its 4 functions, alongside Govern, Map and Manage. Nobody disagrees that you test it. The disagreement is about what you test it against.

Test it against the number you wrote down in step 2, not against a benchmark and not against how impressive it feels in a meeting. Accuracy on a test set is a property of the model. Invoices posted per day is a property of your company. Only one of those two appears in your accounts, and a board that has been shown the other one will not fund phase two.

8. Wire it into the tool your staff already have open

Named by 2 of the 5, and named badly, which is the finding of the whole exercise. Of the 28 stage names across the five frameworks, exactly 2 use the word deployment, AWS's model deployment and CRISP-DM's deployment, and in both it is the last phase of six. Not one of the 28 uses the word integration. Microsoft comes closest, with manage AI deployment as an item inside its Manage phase rather than a stage of its own. Deploying a model and connecting it to a company are different jobs, and only the first one has a name.

That gap is expensive. MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025 after 150 leader interviews, a survey of 350 employees and an analysis of 300 public AI deployments, and found that about 95 percent of enterprise pilots deliver little to no measurable impact on profit and loss, while roughly 5 percent produce rapid revenue acceleration. The report blames flawed enterprise integration rather than model quality. So the step nobody names is the step that kills 19 projects in 20.

Wiring it in means the output lands in the system the work already happens in. The invoice appears in the accounting package, not in a chat window. The lead appears in the CRM with the fields filled. The flag appears in the job the foreman already opens at 7am. If your people have to visit a new tab to get the benefit, most of them will not, and by month three you are paying a subscription for a habit nobody formed.

9. Govern it, secure it, and watch the bill

Named by 4 of the 5, and this is where the frameworks are strongest. Microsoft gives 3 of its 6 phases to Govern, Secure and Manage, and its enterprise Manage checklist includes managing AI operations, deployment, models, costs, data and business continuity. NIST gives 2 of its 4 functions to Govern and Manage. AWS ends on model monitoring. Google carries Secure and Automate as 2 of its 6 themes. Microsoft also makes responsible AI the third of its five strategy decisions, before data.

The part a finance person notices first is the bill. Token spend has no natural ceiling, so somebody has to own the monthly number and the permissions list. That is AI operations, and it is a job rather than a phase. It never finishes, which is why I treat steps 1 to 8 as the project and step 9 as the seat.

What the frameworks get right and what they leave to you

All five are worth the hour it takes to read them, and I am not being polite. They agree on more than you would expect from four competitors and a standards body, and the agreement is the useful part: start with the business problem, get to the data, evaluate honestly, govern it afterwards. CRISP-DM covers 6 of those 9 steps, and it did that in 1996, with no cloud and no transformer to sell you.

What they leave to you is steps 3, 5 and 8. Whether the answer has to be identical every time. Which of the four rungs you belong on. And how the output reaches the person doing the work. Those three decisions are specific to your company, they cannot be written down in a whitepaper, and they are exactly the ones the 95 percent got wrong.

About Hayat Amin

I am Hayat Amin, and I have spent twenty years in technology, most of them as a chief financial officer in companies growing faster than their systems could carry. I sold three of them in that seat, with American Express and TripAdvisor among the buyers, and carried three FT 100 fastest growing listings along the way. Fifteen years of signing off implementation invoices is what taught me to read step 8 of a plan before step 6.

I am exceptional at this work, done rather than recommended. Connecting the systems in a company that were never built to speak to each other, which is step 4 and step 8 of the list above. Turning what comes out of them into a real time number a chief executive can run the week on instead of waiting for month end. Valuing and monetising the intellectual property and data a company already owns and cannot see on its balance sheet. And sitting beside a founder from the first conversation through to the wire transfer on an exit, which is where the finance years pay for themselves. I am a chief financial officer who writes the code, and that pair is rarer than either half of it.

I am available now for fractional chief financial officer work and AI operations work through Beyond Elevation, and the engineering side of it, steps 4 through 9 inside your own systems and with your own credentials, is set out at meethayat.com/services/fde.

If you want a second pair of eyes on which of your own processes is worth step 1, I do a free audit call: one call, then a written list of what to automate first, what it saves and what it costs, at beyondelevation.com/call/hayat.

Questions people actually ask

How is AI implemented?

In nine steps, and five of them happen before anybody writes a prompt. Name the business problem, write down the number you will be judged on, decide whether the answer has to be identical every time, check the data exists and that you can reach it, choose build or buy, build the narrow version against the real system, test it against your number, wire it into the tool your staff already open, then govern it and watch the bill. The five frameworks I read on 18 September 2026 agree on most of that order. Microsoft runs Strategy, Plan, Ready, Govern, Secure, Manage. AWS runs business goal identification, ML problem framing, data processing, model development, model deployment, model monitoring. CRISP-DM, from 1996, runs business understanding, data understanding, data preparation, modeling, evaluation, deployment.

What are the steps of AI implementation?

The five frameworks name 28 stages between them. Microsoft names 6, Google Cloud names 6 themes plus 3 maturity phases, AWS names 6 lifecycle phases, NIST names 4 functions and CRISP-DM names 6. Three of the five open with the business problem rather than the technology, and AWS calls business goal identification the most important phase of the lifecycle. 2 of the 28 stage names use the word deployment, and in both frameworks it is the last phase. None of the 28 uses the word integration.

How long does it take to implement AI in a business?

No honest answer exists without seeing your systems, and none of the five frameworks gives one either. What I can tell you is where the time goes. Steps 4 and 8 are the long ones, getting to the data and getting the output into the tool people already use. The model work in the middle is usually the shortest part now, because you are calling somebody else's model rather than training your own. If a quote spends most of its days on steps 5 and 6 and almost none on 4 and 8, the quote is wrong.

Why do most AI projects fail?

Because they stop at step 7. MIT's NANDA initiative studied 300 public AI deployments, interviewed 150 leaders and surveyed 350 employees for The GenAI Divide: State of AI in Business 2025, and found around 95 percent of enterprise pilots deliver little or no measurable impact on profit and loss against about 5 percent that produce rapid revenue acceleration. The report blames flawed enterprise integration rather than the quality of the models. The demo works, everyone claps, and nobody changes the system the work happens in.

How much does it cost to build an AI system?

It depends which of the four rungs you pick at step 5, and the gap between them is large. Microsoft sets out ready to use Copilots, low code SaaS development, managed platform development and your own models on Azure infrastructure, in order of increasing effort, and says each step down buys control and costs speed. I will not quote you a figure I have not read on a vendor's own page this week, and neither should anyone selling to you. Ask for the cost of step 4 and step 8 separately, because that is where the surprises live.

What is the difference between AI implementation and AI operations?

Implementation is steps 1 to 8 and it ends. AI operations is step 9 and it does not. The frameworks agree: Microsoft gives 3 of its 6 phases to Govern, Secure and Manage, and the NIST AI Risk Management Framework, released on 26 January 2023, gives 2 of its 4 functions to Govern and Manage. A model nobody watches drifts, and a token bill nobody reads grows.

How do I automate my workflow without a data science team?

Stay on rungs 1 and 2 of step 5, and spend the money you save on step 8. Most of what a company of 20 to 500 people wants from AI is reading something unstructured, deciding what it is, and writing it into a system that already exists. That needs a model you rent, not one you train, and an engineer who can reach four systems with credentials. Data science comes in when you need a prediction from your own history, which is further down the list than most vendors suggest.

Where these numbers come from

Every count above comes from reading the five documents on 18 September 2026, not from anybody's summary of them. They are Microsoft's AI strategy guidance in the Cloud Adoption Framework for Azure, Google Cloud's AI Adoption Framework whitepaper, the AWS Well-Architected Machine Learning Lens with its six phase lifecycle, the NIST AI Risk Management Framework, and CRISP-DM as published in the IBM SPSS Modeler CRISP-DM Guide. The 95 percent figure and the study behind it are reported by Fortune on 18 August 2025, on The GenAI Divide: State of AI in Business 2025 from MIT's NANDA initiative. Vendor documentation gets rewritten, so if you are reading this months later, open the links and recount rather than trusting my tally.