How to compare AI design agency proposals
Compare AI design agency proposals by research, error handling, prototype scope, delivery roles and handover. Check what the price actually covers.
Two AI design agency proposals can promise the same outcome while buying very different work. One may cover interviews, interaction testing and an engineering handover. Another may offer a month of design capacity, with research participants and implementation decisions supplied by your team. A polished presentation can make those differences surprisingly difficult to see.
The comparison starts with the decision you need the agency to support. Are you testing whether people understand an AI feature, improving an existing workflow or preparing designs for a product your engineers will build? Each requires different evidence and deliverables. Comparing the proposals against the same brief makes their commercial differences easier to discuss.
This article provides an editorial review method for buyers of AI product design. The worked proposals are fictional and carry no implied market prices or quality ratings. Use the questions to clarify an actual offer, then confirm the proposed scope, people and terms directly with the provider.
Keep in mind
Key takeaways
- Give each provider the same task, users and intended design decision.
- Compare research, difficult states, implementation boundaries and handover as separate responsibilities.
- Resolve scope and commercial assumptions before turning offers into a purchasing decision.
Compare the research work behind the screens
A line saying discovery is included tells you little about how the agency will investigate the problem. Ask which questions the work will answer, which people need to participate and how the findings will change the design. The explanation should connect research activity to a product decision rather than simply list workshop formats.
Clarify recruitment responsibilities. Your organization may need to identify participants, obtain suitable access or arrange sessions. Ask what happens if the intended users are unavailable and whether the agency's proposed alternative can still address the research question. A workshop with stakeholders can be useful without establishing how end users understand a particular interaction.
Check what evidence you receive. Useful outputs can include the research questions, participant context, observations, design implications and unresolved issues. Agree the appropriate handling of research material with the people responsible for it. A buyer needs enough information to understand the decision without assuming that all participant recordings or personal details should be shared widely.
Look for a proportionate plan. An early concept may need a focused investigation before committing to a broader system. An established product with several user groups may require a different level of work. The useful comparison is whether the proposed research supports your next decision, not which proposal contains the largest number of activities.
Ask what happens when the AI result is incomplete
The successful answer is only one part of an AI interface. A proposal should explain how the agency will design for missing information, conflicting sources, slow responses and corrections. Those states often determine whether someone can use the result responsibly within the task.
Bring a representative example to the proposal discussion. Suppose an internal assistant drafts a support reply from approved product documents but lacks a customer detail needed to apply the policy. Ask how the team would explore that interaction. You are looking for a method of designing clarification and review, not a finished solution improvised during a sales call.
Microsoft's HAX guidelines consider interaction at first use, during ordinary use, when a system is wrong and over time. They provide a useful reference for discussing scope. Mentioning the guidelines does not establish that a particular agency follows them, and their presence in a proposal is not a certification of the eventual product.
Ask which states will be represented in the deliverables and evaluated with users. If only the ideal response is prototyped, the proposal may leave important product rules to your engineers. That can be acceptable when explicitly scoped, but it should be visible before you compare an apparently lower fee with a more complete interaction-design engagement.
Reference: Guidelines for Human-AI Interaction ↗
Related reading: Design AI answers people can inspect and correct
Establish what the prototype will actually demonstrate
An interactive prototype can support useful research without being connected to a real model. A coded prototype can explore integration without being a maintained product. Ask what the proposed prototype is intended to test and which parts will be simulated, connected or left outside scope.
For an answer-review workflow, the agency might use prepared outputs to test whether people find and correct an error. That is a different exercise from measuring the reliability of live retrieval. Both can be valuable. The proposal should explain which question the prototype addresses and what conclusions the team will be able to draw from the work.
Identify who provides realistic inputs and difficult examples. Without them, a prototype may look convincing while avoiding the task conditions that matter to your users. Ask how the agency will handle an example containing confidential information or a source the research participant cannot access. These are planning questions to resolve with the relevant owners.
Be precise about the handover status. If prototype code will be discarded, say so. If another team is expected to develop it further, specify what they will receive and what remains to be assessed. Treat the phrase production-ready as a claim requiring explanation of implementation, testing and ownership, rather than as a substitute for those details.
Inspect the proposed design and engineering handover
Ask what your engineers will know at the end of the engagement. Visual files alone may leave them to infer component behaviour, loading states, source relationships and approval rules. A useful handover describes the decisions that need to survive implementation, including the questions the design work has not resolved.
Clarify the design-system contribution. Does the proposal cover a small set of screens, reusable components, documentation or integration with an existing system? Identify who maintains the components afterward. Avoid assuming that a reference to a design system means every interaction in your product will be specified or that working frontend code is included.
Agree the accessibility target and evaluation responsibility. W3C's WCAG provides testable web accessibility criteria, but selecting a target is only part of the work. The proposal should state which design or implementation activities are covered and who will assess the delivered result. Do not treat a design file or a generic accessibility statement as proof that a finished application meets a stated standard.
Check the route for implementation questions. Your engineers may need clarification after a research decision is translated into code. Ask whether support is included, how it is requested and what happens if implementation reveals a constraint that changes the interaction. The answer affects the usefulness of the handover as well as the apparent cost of the initial design phase.
Reference: Web Content Accessibility Guidelines 2.2 ↗
Find out who will perform and review the work
A proposal can describe the capabilities of an agency without identifying the people available for your engagement. Ask for the delivery roles, relevant experience and planned involvement. Distinguish the senior person leading the sales conversation from the person who will run research, produce interaction work and maintain the day-to-day relationship.
Discuss review and continuity. A dedicated designer may be supported by a design lead, or a project team may include several specialists. Ask how work is reviewed internally and what happens when a named person becomes unavailable. The goal is a practical understanding of the service, not an assumption that one staffing model is always superior.
Identify your own required roles. A product owner may need to settle scope, an engineering lead may need to explain technical constraints and a subject specialist may need to check product facts. If your team cannot provide those inputs, discuss the consequences before accepting a timeline that depends on them.
Confirm working arrangements that matter to the project. These may include meeting overlap, decision turnaround and how feedback is consolidated. A listed office location does not by itself establish where the delivery team works or when it will be available. Request the arrangements for the proposed team rather than inferring them from the agency's company profile.
Compare the commercial unit and its assumptions
A fixed project fee, hourly rate and monthly design subscription describe different purchasing arrangements. Before comparing amounts, identify what each unit buys and what assumptions keep the estimate valid. A lower headline figure can cover a smaller scope or place more coordination work on the buyer.
For a fixed scope, ask which deliverables, research activities and revision decisions are included. For a capacity-based engagement, ask how available time is allocated, how competing requests are prioritized and what your team must prepare. A monthly service does not establish that a particular project will finish within one billing period.
Separate calendar duration from delivery effort. Waiting for participant access, internal approval or an engineering decision can affect the plan even when design work itself is ready to proceed. Ask how pauses, delayed feedback and new requests are treated within the proposed arrangement. Confirm the relevant terms with the provider rather than relying on a general service-page summary.
Keep a list of exclusions beside the price. Research incentives, specialist content, implementation and ongoing support may be handled separately. You do not need a universal budget benchmark to make this comparison useful. You need an estimate of the same work, with the assumptions and responsibilities visible enough to discuss.
Related reading: Scope an AI project before comparing budgets
Compare two fictional offers against one task
Consider a buyer redesigning an internal document assistant. The team needs to understand why users accept unsupported answers and prepare an interface for its own engineers to implement. Offer A proposes a defined research-and-design phase. Offer B offers a period of dedicated design capacity. Neither structure establishes which offer is better for this buyer.
The buyer can translate both offers into a shared comparison record. The following table is illustrative. It describes questions to settle, not a scorecard for any named agency.
Scroll to view all columns →
| Comparison point | Offer A needs to establish | Offer B needs to establish |
|---|---|---|
| User research | Participants, questions and findings included | Whether research fits the capacity and who arranges it |
| Difficult states | Which interactions are specified and tested | How those tasks enter the prioritized work |
| Engineering handover | Files, behaviour notes and support included | Expected handover outputs and time allocation |
| Changes to the brief | What triggers a revised estimate | What displaces other work within available capacity |
| Buyer dependencies | Required access and decisions | Prepared inputs and ongoing prioritization owner |
Suppose Offer A includes research planning but assumes the buyer will recruit users. Offer B includes a skilled designer but assumes the buyer supplies completed research. The difference is not merely how the work is billed. The offers rely on different inputs, and the buyer must determine which inputs it can actually provide.
Request clarification before forcing the proposals into a numerical ranking. If neither offer covers a necessary decision, revise the scope. If both cover it through different approaches, ask each team to explain the evidence and handover the buyer will receive. The comparison then supports a purchasing conversation rather than rewarding whichever document contains the most confident language.
Agree how the design work will be accepted
Acceptance should relate to the work commissioned. An agency can deliver research, a prototype and design documentation without establishing the reliability of the underlying AI system. Separate the design outcomes you will assess from engineering or operational outcomes that belong to another scope.
Write observable expectations for the key deliverables. For example, the prototype covers the agreed review task and its selected failure states; research findings identify the tested users and conditions; handover material explains the proposed behaviour and unresolved questions. Tailor these expectations to the project instead of copying a long checklist that neither side intends to use.
Define how feedback becomes a decision. Ask who can approve a direction, how conflicting stakeholder comments are resolved and when a design choice is considered settled for the current phase. Without that process, revisions can continue while each participant assumes a different authority to reopen the work.
Keep changes visible. A new user group, another source system or a request for frontend implementation can alter the engagement. Agree how the provider will explain the impact before proceeding. This helps both sides distinguish a reasonable correction within the agreed work from a new piece of work requiring another decision.
Check the experience that matters to your proposal
Use published case studies to identify a relevant contribution, then ask for detail about the work being offered now. A project involving an AI product may demonstrate interface design without establishing model development, and a broad agency service list may cover capabilities outside the proposed team.
Ask the provider to connect its example to your task. What users were studied, what decisions did the designers own and which work was performed by the client or another supplier? Request the boundaries as well as the attractive screens. An honest explanation of a narrower contribution can be more useful than an unsupported claim of complete ownership.
Check the stage and timing. Research on an early concept can be relevant to discovery, while an implemented design can reveal different handover experience. Neither automatically establishes that the same team is currently available or that your product's constraints are equivalent. Ask for a current reference when the match depends on those details.
Do not convert directory badges into a quality score. Point of AI's profiles describe the available documentation and its limits, with dated source links. Use them to prepare a more specific conversation. The final assessment should consider the proposed team, work and responsibilities for your brief, not a general impression created by a recognizable client name.
Related reading: Inspect AI product design agencies and their published work
Use the clarification meeting to resolve differences
Send the important questions before the meeting so the provider can bring the right people. Group them by the decision they affect: research access, prototype scope, engineering handover, staffing or the commercial basis. This gives the conversation a clearer purpose than asking the agency to present the same document again.
Bring your own accountable roles. If engineering must explain an existing interface constraint, have someone able to settle that point. If a product owner must choose which user group comes first, identify who will make the decision. The agency cannot produce a dependable estimate while several incompatible buyer assumptions remain open.
Record the answer and its effect on the offer. A clarification may confirm that work is included, identify a missing activity or require a revised proposal. Ask for the material points in the updated document rather than relying on a reassuring comment that does not change the stated scope. Keep the comparison brief consistent when another provider needs to answer the same issue.
Leave uncertainty visible when it cannot yet be resolved. A short discovery phase may be a sensible next purchase when the problem needs investigation before a full design estimate is credible. The useful outcome is an agreed route to the next decision, with clear responsibilities, rather than forcing a precise-looking total from information neither side has.
Finish with a short decision record
Summarize the comparison in a document someone else can understand. Record the problem, the phase being purchased, the chosen approach, the required buyer inputs and the questions resolved during clarification. Keep the important exclusions and remaining assumptions beside the commercial decision.
Explain the trade-off you are accepting. You may choose a bounded research phase because the interaction is still uncertain, or ongoing design capacity because your team already has a prioritized backlog and the people needed to support it. The record should make that reasoning visible without pretending that every other purchasing model is unsuitable.
Identify the next checkpoint. It might be a research review, an agreed prototype task or an implementation handover discussion. Give it an owner and a decision to support. That keeps the engagement connected to your product rather than allowing deliverables to accumulate without a clear use.
A strong proposal comparison leaves both parties with fewer hidden assumptions. The buyer understands what it is purchasing and what its own team must contribute. The agency understands the decisions and evidence expected from the work. That shared understanding is a more useful starting point than choosing a document because its screens look the most finished.
Related reading: Prepare for an AI project handover
Read the original thinking
Sources & further reading
This is original Point of AI analysis. The sources support the statements attributed to them. Hypothetical examples and suggested questions are editorial guidance.
- Guidelines for Human-AI Interaction ↗Microsoft · Checked November 13, 2026
Provides interaction-design guidance across first use, ordinary use, errors and change. The proposal comparison is original editorial guidance, not a Microsoft certification.
- Web Content Accessibility Guidelines 2.2 ↗W3C · Checked November 13, 2026
Provides web accessibility success criteria. Buyers must agree the applicable evaluation scope; a design proposal alone does not establish conformance or legal compliance.