AI Consultant or Development Agency: Who Should You Hire First?
TL;DR
Choose the engagement by the uncertainty you need to resolve. Compare an assessment, a proof of concept and production delivery on the same fictional task.
Updated . Replaced unsupported claims with a worked supplier brief, engagement comparison and selection worksheet.
Hire AI advice when the main uncertainty is what to do, whether it is feasible or whether it is worth doing. Hire a delivery team when you have a sufficiently clear problem and need a working system, integration and operating handover. The same firm can provide both, so compare the contracted work and evidence rather than relying on the label “consultancy” or “agency”.
You may also need neither. A missing reporting rule, an inconsistent process or a standard feature already available in your software can be a better first fix than an AI project. A useful supplier should be able to explain that possibility.
This guide uses a fictional support workflow to compare three engagements. The deliverables below are a proposed buying brief, not a report of a Pixelfield client engagement or an independent test of suppliers.
Identify the decision you cannot make yet
“We need AI” is not a brief. A more useful starting point is: staff spend time finding the right policy before replying to delivery queries, and we do not know whether retrieval, a better help centre or a simpler system integration would help most.
That problem contains several uncertainties. Are the policies consistent? Can staff access current order records? How often is the right answer unavailable? Is the time spent searching actually the largest part of handling the request? Those questions need evidence before a full build estimate becomes useful.
The GOV.UK discovery guidance recommends understanding the problem and user needs before committing to a solution. A commercial discovery engagement should similarly leave you with a decision and its supporting evidence, rather than a document that merely restates the original ambition.
If the uncertainty is already narrow, you may not need an organisation-wide strategy exercise. Ask for the smallest investigation that can resolve the decision.

Compare the work you would actually buy
| Engagement | Question it answers | Useful output | What it does not automatically include |
|---|---|---|---|
| Advisory assessment | Which problem should we address and why? | Prioritised options, assumptions, evidence and a recommendation | A deployed application |
| Proof of concept | Can this approach meet a specific test? | Runnable experiment, test cases, results and limitations | Production support or complete security assurance |
| Delivery engagement | How do we build and operate the agreed workflow? | Integrated software, acceptance evidence and handover | An unlimited scope or guaranteed business return |
The boundaries can be combined in one contract, but name them. A demonstration should not quietly become the production system just because people begin using it. Equally, a team hired to deliver a defined feature should not be blamed for failing to answer a strategy question outside the brief.
Ask each supplier to show a sample output from the type of engagement you are considering. Permissioned, anonymised examples are useful; a client logo alone does not tell you which part of the work the supplier performed.
A worked brief for a support assistant
Imagine a company with current policy documents, an order system and a support inbox. It wants to reduce searching time while keeping a person responsible for sending customer replies.
For an advisory assessment, ask the supplier to map the existing task, inspect a permissioned sample, identify missing data and compare a retrieval assistant with simpler alternatives. The decision might be to improve policy ownership before testing a model. That can be a useful outcome even though no assistant ships.
For a proof of concept, narrow the scope to a defined class of delivery-status questions. Specify expected answers, unsupported cases and the evidence a reviewer needs. Ask for results on a held-out test set and an explanation of failures. The AI proof-of-concept example shows how to make that decision testable.
For delivery, add integration permissions, access controls, review states, monitoring, operational ownership and a fallback route. The acceptance brief should say what happens when the order system is unavailable or the policy cannot answer a question. A polished chat window is not enough.
The three engagements now have distinct completion conditions. You can compare proposals without pretending they are interchangeable products.
Ask who will do the work
Find out who investigates the workflow, who writes the integration and who handles problems after release. A senior person in the sales meeting may not be the person delivering the project. That is not inherently a problem, provided responsibilities and access to expertise are clear.
For advice, look for the ability to challenge assumptions and communicate a decision. For implementation, ask about relevant systems, testing and operational handover. For both, ask what evidence would make the supplier recommend against the proposed approach.
A firm can be strong at both advisory and delivery work. It can also be strong in one industry and unfamiliar with the systems your project depends on. Ask for comparable constraints rather than a superficially similar company name.
Our UK AI consulting shortlist and AI development company shortlist support those separate searches. They should follow the decision about the work you need, not replace it.
Compare fees after aligning scope
There is no single useful price for “an AI consultant”. A short feasibility assessment and an enterprise change programme include different work. A day rate, a fixed fee and a recurring support charge also represent different commitments.
Ask proposals to separate discovery, data preparation, the experiment, production work, third-party charges and ongoing support. Identify what your own team must supply. A lower supplier fee may assume substantial internal work that another proposal includes.
Check what happens when an assumption fails. Can the engagement stop after discovery? Is the next phase optional? What evidence triggers a change request? A staged commitment can be more informative than comparing two large totals built on unresolved requirements.
This article does not provide market-average pricing. Use a scoped proposal and the AI consulting fees guide for the relevant cost discussion, subject to its stated assumptions.
Define ownership and handover before starting
List the outputs you need to keep: research notes you are entitled to receive, data mappings, test cases, source code where agreed, configuration, deployment instructions and an explanation of third-party dependencies. Ownership and licence terms should be explicit in the agreement.
For a proof of concept, ask what can be reused and what would need rebuilding. Experimental shortcuts may be reasonable, but they should not be hidden. For delivery, establish access to repositories, accounts and operational documentation appropriate to the agreed engagement.
A useful handover includes someone showing that the receiving team can perform a real task, such as deploying a change or investigating a failed integration. Receiving a folder of documents is not the same as being able to operate the system.
Use a small set of evidence-based selection questions
Download the supplier selection worksheet. It asks each candidate the same questions: what decision or outcome will this phase deliver, what evidence demonstrates completion, which assumptions could change the scope, who does the work, what remains outside the contract and what can we operate afterwards?
Record the answer and the supporting example separately. “We follow best practice” is a claim; a redacted acceptance test or handover record is something you can inspect. Do not invent a numerical ranking if you have not defined comparable criteria.
If you are still choosing the problem, start an AI consulting discussion. If the problem and acceptance conditions are already clear, discuss AI development. The next step should match the uncertainty you actually need to resolve.
1 Comment