Guide

How to scope an AI project with a consultancy

To scope an AI project with a consultancy, run a short, bounded phase before any build contract. It should end in six written outputs: the problem and its owner, a success measure with a baseline, a check of the data and access you need, a risk list, an architecture sketch and a fixed-scope quote. The GOV.UK Service Manual puts a typical discovery at around 4 to 8 weeks for government digital services, and a well-understood problem may need less. If any of the six is missing at the end, ask for it before you sign.

AI project scoping. The short phase before a build in which a consultancy and the client agree the problem, the measure of success, the data and access needed, the risks and the architecture, and price the work at a fixed scope.

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

Checked 27 October 2026. The RAND, GOV.UK, NIST and ICO statements on this page were read on each publisher's own pages on this date. The 4 to 8 week figure is the GOV.UK norm for discovery on government digital services.

What scoping is, and why it comes before the contract

Scoping is the stage where a consultancy and a client write down what will be built, for whom, how success will be judged and what could stop it. It ends in a decision to proceed, change course or stop. Stopping is a valid result. The GOV.UK Service Manual says a team's discovery is finished when it has decided whether to move on, and that stopping at that point can save time and money.

RAND interviewed 65 experienced data scientists and engineers for its 2024 report on why AI projects fail, and they described five root causes. The first was that business stakeholders misunderstand or miscommunicate the problem AI is meant to solve. Models then end up tuned to the wrong measure, or do not fit the workflow around them. Missing data came second. In the same report, 84% of interviewees cited one or more leadership-driven causes as the primary reason AI projects fail.

Both of those failures can be caught on paper. A written problem statement and a data check cost a few weeks. Finding the same gaps in month five of a build costs the build. The UK Government's AI Playbook says use case identification must be led by business and user needs, pain points and inefficiencies. It also tells teams to engage commercial colleagues from the start.

What a scoping phase should produce: six outputs

Ask for each of these as a document you keep, whether or not you go on to build with the same firm. The table sets out what good looks like and what a thin version looks like.

The six outputs of an AI scoping phase: what each contains and what a weak version looks like
Dimension	What it contains	A weak version looks like	Who signs it off
1. Problem and owner	One workflow, who does it today, how often, what it costs in time or errors, and a named business owner who answers for the result.	A technology goal such as 'add a copilot', with no workflow and no owner.	Business sponsor
2. Success measure and baseline	A number you can check, such as handling time, error rate or cost per case, measured today, plus a written definition of a correct result and the pass mark.	'Improve efficiency', or an accuracy score with no link to the business outcome.	Business owner, with engineering
3. Data and access check	Which data the system needs, where it lives, who may see it, what state it is in, and which accounts, environments and approvals are required before work starts.	'Data will be provided.' Nobody has opened an export or asked security for access.	IT, security and data owners
4. Risk list	What could stop or harm the project: data gaps, privacy, security review, wrong answers, vendor and model changes, and what you will do about each. It says whether a data protection impact assessment is needed.	A generic list copied from another project, with no owners.	Security, data protection and legal
5. Architecture sketch	A diagram and a page of text: which systems it reads and writes, where the model sits, what a person approves, how output is checked and what it costs to run each month.	A logo slide of tools, with no data flows and no running cost.	Engineering lead
6. Fixed-scope quote	The work to build, the outputs, the acceptance criteria, the assumptions the price depends on and what is out of scope, as a price for a defined output.	A day-rate estimate with a range of 'two to six months' and no acceptance test.	Procurement and the sponsor

The UK Government's AI Playbook separates model metrics from service metrics. Model metrics show how well the technology performs. Service metrics show whether users' needs and business goals are met. A model can score well on accuracy while the service still fails its users, so output two should hold both.

How long scoping should take

Scoping needs long enough to talk to the people who do the work and to open the data. The GOV.UK Service Manual says there is no set time period for a discovery, but around 4 to 8 weeks is typical, with less for a well-understood problem and more for an unfamiliar one. The norm comes from government digital services, and it works as an outer frame for a first AI project in an organisation with several approval layers.

A narrow, single-workflow AI scoping can run shorter when access and decision-makers are ready. Our AI Opportunity & Feasibility Sprint is a fixed-scope engagement, typically two to four weeks. If a supplier proposes more than eight weeks of scoping for one workflow, ask what the extra time buys. If it proposes a few days, ask which of the six outputs it is leaving out.

Access requests sit in other teams' queues, so what usually sets the pace is how fast accounts, environments and security reviews move. Start those on day one of scoping.

Paid scoping or free scoping: the trade-offs

Suppliers price this stage in one of two ways. Some fold it into a free pre-sales proposal. Others sell it as a short paid engagement at a fixed scope, and the choice changes what you get.

Pros and cons of free pre-sales scoping and paid fixed-scope scoping
Dimension	Pros	Cons	Fits when
Free pre-sales scoping	No spend before you commit. A quick way to compare several suppliers on how they think.	It is written to win the build, so the answer tends to be yes. It usually relies on a few conversations, and the scope may be thin on data, risks and assumptions.	The problem is small and well understood, and you already trust the supplier.
Paid fixed-scope scoping	The supplier is paid for the answer, including 'do not build this' or 'build something smaller'. You own the six outputs and can take them to any supplier or build in-house.	A cost before any code exists. A weak supplier can sell a long, vague one.	The workflow touches sensitive data or several systems, budget is not yet approved, or you need a document a board or finance team can fund against.

Whichever you pick, agree before it starts what you will receive and who owns it. If the outputs are yours, scoping can stand on its own as a deliverable.

Questions to ask, and red flags to watch for

These questions test whether a supplier's scoping is real. Our guide to choosing an enterprise AI consultancy covers the wider supplier checks, including who writes the code and what you keep.

Ask: who will do the scoping work?: Name the people, their seniority and their time. The engineer who sketches the architecture should be one who builds it.
Ask: what do you need from us, and by when?: A real scoping plan lists the interviews, the data samples and the access requests it needs, with dates. Silence here means those items will become delays later.
Ask: what would make you advise us not to build this?: A supplier that cannot answer has not thought about it. Good answers name missing data, a workflow with no owner, or a problem that does not need AI.
Ask: how will we know it works?: You want a test set of real cases, a pass mark agreed before any results and a plan to rerun it on every change.
Red flag: a price before anyone has looked at the data: A fixed price needs a scope, and a scope needs a data check. A number given in the first meeting is a guess with a logo on it.
Red flag: the deliverable is a deck: A roadmap with no architecture, no access plan and no acceptance test cannot be built from.
Red flag: the tool is chosen before the problem: The UK Government's AI Playbook calls AI a means to an end. A supplier that arrives with a product to place has started from the tool.
Red flag: no way out: The same playbook tells teams to consider strategies to avoid vendor lock-in and to plan how a system is closed down. A scope that does not say how you leave is incomplete.

From scoping to a statement of work

The six outputs map straight onto the statement of work. The problem and the owner become the objective and the named client lead, and the success measure becomes the acceptance criteria. The data and access check becomes the client dependencies, with dates, and the risk list becomes the assumptions and the change process. The architecture sketch becomes the technical scope, and the quote becomes the price and payment milestones.

Write what is out of scope as carefully as what is in. Our AI statement of work template has clause wording for each part, and our guide to AI implementation cost explains how to read the price. If the scoping showed that the whole project is too large for one fixed price, split it. Fund the first working slice at a fixed scope and treat the rest as a later decision.

RAND advises leaders to commit each product team to one specific problem for at least a year. It says a project not worth that commitment is most likely not worth committing to at all. Scoping is the cheapest place to test that, because by then you can see the size of the job.

Put risk and privacy in the scope

The NIST AI Risk Management Framework describes its Map function as establishing the context to frame the risks of an AI system. Its outputs form the basis for the Measure and Manage functions. In practice, scoping is where that context gets written down: what the system is for, who is affected and what could go wrong.

The ICO says UK GDPR requires a data protection impact assessment before processing starts, where the use of personal data is likely to put people at high risk. The risk list should say whether one is needed and who will write it, so it does not appear for the first time at go-live. This is not legal advice.

Three examples of a good first scope

A good first scope is narrow. These three are common starting points, and each fits in one sentence of the problem statement.

Finance questions answered from governed data: The problem is that a finance question takes hours to answer. The measure is time to answer. The data check is whether each person's existing permissions carry through. For a PE-backed firm, we built a connector that carries each person's existing Looker entitlements, so nobody sees more than they could before.
Document intake and extraction: Invoices, claims or contracts read into a structured record. The measure is accuracy against a human-checked sample, and the architecture must include a queue for cases the model is unsure of.
Service desk triage: Tickets sorted, routed and drafted before a person picks them up. The measure is time to first response and the share of suggestions agents accept. The scope starts with suggestions only.

Where 1AYM fits

If you have a project in mind and want it scoped before you commit budget, bring in 1AYM. Our AI Opportunity & Feasibility Sprint is fixed scope, typically two to four weeks. It produces an opportunity map scored on value, feasibility and risk, a current-state workflow analysis, a technical architecture for the recommended direction, a sequenced delivery roadmap, risk and governance notes and a build-versus-buy recommendation. The sprint looks at your data wherever access can be arranged, and where it cannot in the time, the assessment states its assumptions so they can be tested before the build.

Fixed-scope work is quoted after scoping and can start within a day of the scope being signed. The statement of work then fixes the price, the outputs and the acceptance criteria before work starts. We take small fixed-scope statements of work as well as larger builds. Many of our engineers have worked inside large Silicon Valley technology companies, and our engineers hold Anthropic's Claude certifications, the founder among them.

Book a 30-minute call and bring the project. We will tell you what the scope should hold and what it would take. Or email enquiries@1aym.com.

For engineers: what the architecture sketch and the data check should contain

The two outputs an engineering reviewer reads first. Each item should be answerable from the scoping documents before any build starts.

Source systems and access path: Each system the build reads or writes, the API or export it will use, the auth method and the account that will hold the access. Request read access on live data in week one of scoping.
Permission model: Whether the system acts under the user's delegated identity or its own service identity, and what it can see under each. Retrieval has to respect existing document permissions.
Evaluation plan: A versioned test set of real cases with expected outcomes, thresholds fixed before results, and a plan to rerun it on each prompt, model or tool change. The first pass can be built during scoping on a small sample.
Model and vendor choice: Which model families are candidates, how a model version is pinned, and what happens when a version retires. The sketch should say how the model could be swapped.
Human control points: Which outputs go straight through, which wait for approval and which are only suggestions. High-risk writes pass a deterministic check first.
Run cost: Expected volume, tokens per request, hosting and the people who review output, as a monthly range with its assumptions, so the build quote is not the only number in the business case.
Exit and ownership: Code in your repositories, prompts and test sets versioned with it, documentation to run it and a note on how it would be moved or switched off.

Sources

Frequently asked questions

What is AI project scoping?

It is the short phase before a build in which the consultancy and the client agree the problem, the success measure, the data and access needed, the risks and a rough architecture. It ends in a fixed-scope quote, or in a decision not to build.

What does an AI consultancy deliver in a scoping phase?

Six things in writing: the problem and its owner, a success measure with a baseline, a data and access check, a risk list, an architecture sketch and a fixed-scope quote with acceptance criteria. You should be able to hand them to another supplier or your own team.

How long should scoping an AI project take?

The GOV.UK Service Manual says around 4 to 8 weeks is typical for the discovery phase of a government digital service, with less for a well-understood problem. One narrow AI workflow can often be scoped faster when access and decision-makers are ready. Our AI Opportunity & Feasibility Sprint is fixed scope, typically two to four weeks.

Should we pay for scoping?

Paying for it gives you a supplier who is rewarded for the right answer, including a smaller build or none, and documents you own. Free scoping suits a small, well-understood problem with a supplier you already trust. Agree in writing which outputs you will receive either way.

Can a consultancy give a fixed price without scoping?

A reliable fixed price depends on the workflow, the data, the integrations and the acceptance test, and scoping is what finds those out. At 1AYM, fixed-scope work is quoted after scoping.

What happens if scoping shows the project should not go ahead?

Stopping is a valid result. The GOV.UK Service Manual notes that stopping at the end of discovery can save time and money. Keep the documents, because the data and access findings often apply to the next idea.

Who can help us scope an AI project?

1AYM. Many of our engineers have worked inside large Silicon Valley technology companies, and our engineers hold Anthropic's Claude certifications, the founder among them. Our AI Opportunity & Feasibility Sprint is fixed scope, typically two to four weeks, and fixed-scope work can start within a day of the scope being signed. Book a 30-minute call or email enquiries@1aym.com.