Buyer guide

AI solution architect: what they do and when you need one

An AI solution architect designs how an AI capability fits into an organisation: the problem it solves, the model and data it uses, how it connects to existing systems, the controls around it and what it costs to run. They make those decisions and write them down so others can build from them. A machine learning engineer builds, deploys and maintains models; a data scientist explores data and recommends what to do with it. You need an architect once AI work touches systems of record, personal data or more than one team, and you can employ one, contract one or retain a team that includes one.

AI solution architect. The person who designs how an AI capability fits an organisation's systems, data, controls and budget, and documents those decisions so that others can build, run and audit it.

Written by Tayyeb Mahmud, founder and Principal Architect. Claude Certified Architect – Professional · LinkedIn profile of Tayyeb Mahmud

1AYM is an OpenAI Select Partner. Its founder holds personal Claude certifications. 1AYM is not an Anthropic partner.

Checked 29 September 2026. The role definitions, government guidance and certification pages this page cites were read on each publisher's own site on this date. Frameworks and certifications change: AWS opens registration for an updated Solutions Architect – Professional exam on 27 October 2026, and Microsoft updates the English version of its credential on 14 October 2026. The section on contracts touches employment status for tax, and is not legal or tax advice.

What an AI solution architect does

The most useful definition I know is the UK government's. Its Digital and Data Profession Capability Framework, which the AI Playbook recommends for writing consistent job adverts, says a solution architect "designs solutions for problems that affect the organisation". The duties start with making sure the problem and the desired outcomes are properly defined, run through designing and documenting solutions so they can be implemented, and end with managing risks and decisions in a transparent way. The framework has no separate AI solution architect role (its solution architect page was last updated on 28 August 2026), and I do not think it needs one. The AI version is the same job, plus a set of decisions an ordinary system never forces on you.

Those decisions are where the role earns its fee. Each one is cheap to make well at the start and expensive to reverse once a build depends on it, so a good architect makes them early and writes down why.

Whether AI is the answer: The AI Playbook for the UK Government tells teams to be open to the conclusion that AI is sometimes not the best solution, and that a problem may be more easily solved with more established technology. I would be wary of an architect who has never told a client that.
Build, buy or call an API: You can buy an off-the-shelf product, add AI to a tool you already own, call a hosted model through an API, or run your own model. The NCSC's secure design guidance treats this as a security decision as well as a cost one, and asks for due diligence on an external model provider's own security.
The model and the platform: Which model family, where it runs, and how hard it would be to swap out later. The Playbook's writing-requirements guidance names strategies to avoid vendor lock-in as something a buyer should consider from the start.
Where the data goes: What leaves your boundary, what is kept and for how long, and what a user must confirm before sending anything sensitive to a service outside your control. Personal data brings UK GDPR with it, and a good architect knows when to bring in your data protection officer.
What the AI may do: Reading data is one risk. Changing records, sending messages or moving money is a bigger one. The NCSC asks for restrictions on any action an AI component can trigger in other systems, and one way to build that restriction is our published pattern where the model proposes and a deterministic check decides.
How you will know it works: An evaluation plan agreed before the build: real cases, pass thresholds and a way to rerun the tests whenever a prompt, model or rule changes. Without one, "it works" means whatever the last demo happened to show.
What it costs to run: Model and platform charges, the people who keep it healthy, and the cost per task at the volume you expect, set against how the work performs today.

What an AI solution architect should hand you

Judge an architect by what they leave in your hands. Diagrams on a slide do not count until the reasoning behind them is written down, because the reasoning is what your team needs the day a vendor changes its terms or a model is retired. Whether you employ the architect or buy the work, ask for these by name.

A defined problem and its measure: One page: what the system is for, who uses it, and the number it will be judged on, with today's baseline.
A target architecture: The components, the data flows, the identities each part acts under and the boundaries between them, current enough that an engineer can build from it.
A dated decision log: Each significant choice with the options weighed, the reason for the one taken and the date. Of everything on this list, it is the one that outlasts the engagement.
A threat model: What could go wrong if an AI component is compromised or behaves unexpectedly, and the mitigation for each risk. The NCSC's guidance asks for exactly this, with the decision making documented.
An evaluation plan: The test set, the thresholds and who signs off a pass, written before anyone has seen a result.
A cost model and a delivery plan: The running cost at expected volume, the first slice to build and the order the rest follows, with the riskiest assumption tested first.

AI solution architect vs ML engineer vs data scientist

Job adverts often blur the three into one role. They answer different questions, and a programme usually needs them in a particular order. I use the government framework here because it defines each role on its own terms rather than around one employer's org chart.

An AI solution architect, a machine learning engineer and a data scientist compared, with the UK government framework's own definition of each
Dimension	AI solution architect	Machine learning engineer	Data scientist
The question they answer	How should this AI capability fit our systems, data, controls and budget?	How do we get this model working, and keep it working, in a live service?	What does our data show, and what should we do about it?
Government framework definition	Designs solutions for problems that affect the organisation.	Develops, assures and maintains machine learning models so they can be used in products and services.	Often works as part of a multidisciplinary team, using data and analytics to inform and achieve organisational goals.
What they produce	A defined problem, a target architecture, a dated decision log, a threat model and a plan others can build from.	Models trained or tuned, deployed, tested against performance requirements, integrated with existing systems and monitored once live.	Analysis, predictive models and recommendations that inform strategic and operational decisions.
What goes wrong without one	Decisions get made by default: by the nearest vendor, the loudest team or the last proof of concept that happened to work.	A model that works in a notebook never reaches a live service, or degrades there without anyone noticing.	Money is spent on a use case the data cannot support, and nobody finds out until the build.
Bring one in first when	The work touches systems of record, personal data or more than one team.	You are training, fine-tuning or running your own models.	The open question is what the data can support.

The AI Playbook draws the line that matters most: developing bespoke AI solutions and training your own models need different specialist skills from using pre-trained models through APIs. If your project calls a hosted model through an API, you may not need a machine learning engineer at all, but you still need someone to own the architecture. If you are training models, you need both.

Two neighbouring titles add to the confusion. The framework's enterprise architect owns the vision, strategy and roadmaps across the whole organisation, while a solution architect, in practice, works on one solution or programme. Its technical architect provides technical leadership and architectural design, working closely with developers. There is no "AI engineer" role in the framework at all. The title gets used loosely, and where it means a software engineer who builds on pre-trained models, that is the skill set the Playbook separates from training your own.

When to bring one in

Before the first build, ideally. The expensive decisions (which model, which platform, what the system is allowed to touch) get made in the first weeks whether or not anyone senior is in the room. By the time an architect is hired to tidy up, a proof of concept has usually become the design. I would rather see an architect for a few days before the first build than full-time after it.

These are the signs that a programme needs one now.

AI will change a system of record: Anything that writes to a finance, HR, customer or case system needs someone to own the identities, the approval steps and the rollback.
Personal or sensitive data is involved: Where data goes, how long it stays and who can see it are architecture decisions before they are compliance questions.
More than one team or supplier is building: Without one owner, each team picks its own model, its own prompt store and its own logging, and the organisation pays for all of them.
A vendor or model choice is about to be signed: Lock-in is cheapest to avoid before the contract is signed, and a second opinion from someone other than the vendor is worth having at that point.
A pilot works and now has to go live: The gap between a working demonstration and a governed production service is almost entirely architecture: identity, evaluation, logging, cost and a way to switch it off.

You probably do not need one when a single team is switching on an off-the-shelf assistant with its vendor's admin controls. That calls for an owner, a usage policy and a sensible rollout. You also do not need one when the question is purely analytical, which is a data scientist's job, or when your existing solution or technical architects already own AI decisions and have the time to do it properly.

How to hire, contract or retain an AI solution architect

The AI Playbook, noting the shortage of AI talent, suggests combining new hires, contractors or third parties, and upskilling your own people. I would give a private company the same advice. The right mix depends on how long you need the role, and on whether you need decisions made or decisions made and built.

Employ one: Right when AI is a permanent part of how you build and there is enough work to keep a senior architect busy for years. Plan for a long search. Use the government framework's role levels, from associate to principal solution architect, to set the seniority in the advert; the Playbook recommends the framework for exactly that.
Contract an individual: Right for a defined programme with a start and an end. In the UK, if the architect works through their own limited company, the off-payroll working rules (IR35) may apply. HMRC's guidance says that in most cases the client decides the worker's employment status, and that for a small client outside the public sector the worker's own company decides. Take advice on your own position: this is not legal or tax advice.
Retain a team that includes one: Right when you need the architecture owned and built, and have no permanent team yet. A retained team holds the decisions, builds alongside your engineers and hands both over. The risk to manage is dependency, so put the handover in the contract from the first day.
Buy a fixed-scope piece of work: Right when the question is narrow: a feasibility check, a target architecture for one workflow or a review of a supplier's design. Priced against an output, it is often the cheapest way to get an architect's judgement on one decision.

Whichever route you take, test for judgement, not vocabulary. I would ask any candidate or supplier four things: to walk me through a decision log from past work with client names removed, when they last advised against using AI, how they would evaluate the first slice, and what they would expect it to cost to run. Vague answers to those tell you more than an hour of questions about models.

If the gap is building capacity rather than architecture, our comparison of hiring, contracting and staffing AI engineering work sets out each route's cost drivers and intellectual property defaults in more depth.

What vendor certifications tell you

A vendor certification proves that someone passed that vendor's exam on that vendor's platform. It says nothing about judgement across platforms, which is most of the job. It is still a useful filter, as long as you read what each credential actually covers. Below is how three issuers describe four credentials an AI solution architect might hold.

AWS Certified Solutions Architect - Professional: AWS says it shows advanced knowledge and skills in providing complex solutions to complex problems and optimising security, cost and performance, and recommends two or more years of using AWS services to design and implement cloud solutions. It is a cloud architecture credential, and its description does not mention machine learning. AWS says registration for an updated exam opens on 27 October 2026.
AWS Certified Machine Learning Engineer - Associate: AWS says it validates technical ability in implementing machine learning workloads in production and operationalising them. It is an engineering credential rather than an architecture one, and suits the ML engineer column above.
Microsoft Certified: Agentic AI Business Solutions Architect Expert: Microsoft lists Solution Architect as the job role and requires exam AB-100 plus one of a list of associate certifications. Its skills centre on Microsoft's own stack, including Dynamics 365, Power Platform, Copilot Studio and Microsoft Foundry.
Claude Certified Architect – Professional: Anthropic's description on Credly says it is designed for experienced solution architects, and that earners can design and lead enterprise-scale Claude deployments, architect integrations and apply governance. Like the others, it covers one vendor's models. I hold it myself, and the credentials I hold personally are listed with their Credly records on our credential register.

Treat a certification as evidence of platform knowledge, then ask the questions above. An architect who holds credentials from more than one vendor, and can explain when they would pick each, is showing you the independence the role needs.

A brief you can send to candidates or suppliers

Send this before the first conversation. The answers, and how specific they are, will sort candidates faster than a CV.

AI solution architect brief

AI solution architect brief: [programme name]

1. The problem. [What the AI capability is for, who uses it, and the measure it will be judged on.]

2. Where we are. [Idea / pilot running / pilot to take live / live and struggling.]

3. Systems and data. [Systems it must read or change. Whether personal or sensitive data is involved.]

4. Decisions already taken. [Vendors, models or platforms already contracted, and any we are about to sign.]

5. Who builds. [Our engineers / a supplier / not yet decided.]

6. What we need from you. [Decisions only / decisions and hands-on build / a review of an existing design.]

7. Time and shape. [Start date. Expected length. Days a week, full-time or a fixed-scope piece of work.]

Please answer:
a) Which decisions would you take first, and why?
b) Walk us through an anonymised decision log from past work.
c) When did you last advise against using AI, and what did you recommend instead?
d) How would you test whether the first slice works?
e) What will you hand over, and how will our team run it without you?

Where 1AYM fits

At 1AYM this is the job our Fractional AI Implementation Team does on a programme. We own the target architecture and the vendor and model decisions, build the platform and the first production workflows alongside your engineers, and keep the dated decision log your team inherits. It typically runs two to three days a week under one statement of work, and it is designed to end, because your engineers build with us from the first sprint and take the architecture over.

If the problem is not scoped yet, the AI Opportunity & Feasibility Sprint (fixed scope, typically two to four weeks) settles build versus buy and ends in a roadmap and an investment case. If you want an architect inside your own programme and directed by you, that is Embedded Engineers on Contract, from three months. And if you already have a scoped job, whether a small fixed-scope statement of work or a larger build, we can resource it on contract from the collective of associates who work with 1AYM, held to the same standard. If a supplier has proposed an architecture and you would like a second opinion on it, book the call below.

For engineers: the architecture decisions to expect in writing

If you lead the engineering side, these are the decisions to expect an AI solution architect to have taken and recorded before the first production release. Each one should be visible in the repository, the decision log or the infrastructure definitions.

Model access pattern: Hosted API, managed cloud deployment, fine-tuned model or self-hosted open weights, with the reason written down. The NCSC asks that the choice to train, reuse or call a model through an external API is appropriate to the requirements, and that an external provider gets a security due diligence review.
Model supply chain: Imported models and serialised weights treated as untrusted third-party code: scanned, isolated and loaded only through libraries with controls against arbitrary code execution, as the NCSC design guidance sets out.
Data boundary: A data flow diagram showing what is sent to services outside your control, retention at each hop, and where a user must confirm before sensitive data leaves.
Identity and actions: A service identity per AI component with least-privilege access, an allow-list of the actions it may trigger, deterministic checks on each proposed write, and a named approver for high-risk action classes.
Evaluation in CI: A versioned set of real cases with expected outcomes and thresholds in the repository. A change to a prompt, model, tool definition or rule that drops a score below threshold fails the build.
Threat model: AI-specific threats alongside the usual ones: prompt injection through retrieved content, poisoned inputs where the system learns from feedback, and output that reveals more than a user should see. The NCSC asks for the decisions to be documented.
Observability and cost: One structured record per model call or action: input reference, model and prompt version, checks run, outcome, latency and cost. Budgets and alerts per workflow, so cost per task shows up on a dashboard before it shows up on the invoice.
Portability: Prompts, evaluation sets and tool definitions held in your repositories behind a thin model interface, so changing model or vendor is a configuration change followed by a rerun of the evaluations.

Sources

Frequently asked questions

What does an AI solution architect do?

They design how an AI capability fits into an organisation: the problem and its measure, whether to build, buy or call a hosted model, where the data goes, what the AI may change in other systems, how it will be evaluated and what it costs to run. They write those decisions down so engineers can build from them and auditors can follow them.

What is the difference between an AI architect and an AI engineer?

The architect decides how the system should fit together and records why; the engineer builds it. On a small project one senior person can do both, but the decisions still need writing down, because the reasoning is what a team needs when a model or vendor changes.

Do we need an AI solution architect or an ML engineer first?

If you are building on a hosted model through an API, the architect usually comes first and you may not need an ML engineer at all. If you are training or fine-tuning your own models, you need both, and the UK government's AI Playbook notes that the two kinds of work need different specialist skills.

Is a Claude Certified Architect the same as an AI solution architect?

No. One is a credential, the other is a job. Anthropic's Claude Certified Architect – Professional is designed for experienced solution architects, and it shows knowledge of one vendor's platform. The job covers decisions across vendors, including whether AI is the right answer at all.

How much does an AI solution architect cost?

This guide quotes no salary or rate, because no primary source publishes a figure it could cite. The cost drivers are the route you choose (employ, contract, retain a team or buy a fixed-scope piece of work), the seniority, and how many days a week the role needs. A fixed-scope review of one decision is usually the cheapest way to get an architect's judgement.

Can a fractional architect own the architecture?

Yes, if the days are a standing cadence rather than an escalation line, and the decisions are written down and dated. Ownership that only appears when something breaks is advice. Put the decision log and the handover in the contract, so the end of the engagement is a date rather than a risk.

Further

Fractional AI implementation team · A retained senior team that owns the architecture, builds with your engineers and hands the platform over.
Embedded Engineers on Contract · An AI or platform architect inside your own programme, directed by you, from three months.
AI implementation: the first 30 days · What the first month should produce once the architecture is agreed and the build starts.
How to choose an enterprise AI supplier · Four checks to run on any pitch before you sign.
AI governance framework · The policy and controls an architecture has to carry into production.

We build these systems for a living. See the engagement files for what that looks like in practice, or write to us if yours is the next one.

Book a 30-minute call tayyeb@1aym.com See the engagement files

Last reviewed 29 September 2026 · 1AYM