Which AI product design agency can add AI to existing software without a bolted-on feel?

An AI assistant can answer a question and still leave the user with the same work. The answer sits in a chat panel; the record they need to change sits somewhere else. So they copy, paste, and check it by hand. That gap between the answer and the finished task is what people mean when they say an AI feature feels bolted on.

Eight agencies publish named AI work that speaks to this problem in different ways. Eleken, Humbleteam, and Neuron show AI added inside established B2B workflows. Punchcut shows assistant and machine-learning work for large platforms. Lazarev.agency, MetaLab, reloadux, and ustwo show products where AI is the core of the experience, which matters if you are building an AI-native product rather than extending one.

Humbleteam publishes this comparison and is one of the agencies included. Entries are alphabetical, without a ranking. We re-checked every source on October 20, 2026. The fit notes are editorial judgments about published work, not a guarantee of results.

What does a bolted-on AI feature look like?

You can usually spot it in five places:

• The AI lives in a side panel or a new tab, behind a sparkle icon, while the task happens on the main screen.
• The answer is a dead end: it cannot go into the record, form, or report it was meant for.
• The prompt says “Ask anything” and gives no hint of what the feature can do with the data on screen.
• A slow answer, a refusal, and a wrong answer all get the same red banner.
• Nobody can tell which change the AI proposed and which one a person approved. Worse, the AI may show data the user is not allowed to see.

An integrated feature turns each of these around. It starts where the work already happens and writes its result back into the user’s record. Failure and review get their own states, and the AI respects the roles the product already has.

How we chose the agencies

Every agency below meets the same three conditions: a named AI or machine-learning product case on its own site, a described design contribution in that case, and a public location. Humbleteam is held to the same standard, using its own case pages. We then read each case for the four integration signals above. The table records what each source describes; “not described” means the source is silent, not that the agency lacks the skill.

Which agencies belong on your shortlist?

Eleken: new AI features in an existing design system

Eleken’s Avid case describes a designer working with the client’s developers on an AI fundraising product: a Playbook module that presents AI recommendations as cards, with a table for sorting and filtering campaign data. The Siena case goes further into integration. The team designed new AI features inside Siena’s existing design system, put rating and feedback inside each automation test so users do not switch context, and placed citations beside responses.

Consider Eleken when your developers will build and you need a designer embedded in the product. Siena’s features were still being tested with customer teams when the case was published, so ask what shipped. Eleken lists Kyiv, Ukraine, and a US entity in Newark, Delaware.

Humbleteam: AI that stays inside the task

Humbleteam’s Cluely work redesigned a meeting assistant around attention: suggestions sit at the edge of the screen, the prompt to act is separated from supporting detail, and summaries keep the meeting’s context next to the recap. In the Afternoon case, an assistant prepares client meetings, turns recordings into notes, and drafts reviews. The adviser sees every step and signs off.

The AI-in-SaaS service describes drawing slow, refused, empty, and wrong-answer states before the happy path. Your engineering team implements the screens; Humbleteam does not deliver production code. Its public pages show design scope, not retrieval or permission architecture, so settle those owners early. Humbleteam lists Prague and Dubai.

Lazarev.agency: from conversation to a research result

The Accern Rhea case shows a financial research tool that mixes prompts with charts, references, and widgets. When a query is unclear, Rhea asks clarifying questions. Users drag text and widgets from the conversation into a split-screen report, and the prompt field also works as a command line for files, alerts, and scheduling.

This is strong evidence for a research or analysis product whose output must become an editable document. Rhea was a new product in Accern’s ecosystem, so ask how the team would fit the same patterns into screens you already have. Lazarev lists Mountain View, California.

MetaLab: AI experiments next to a trusted product

MetaLab’s Atlantic Labs case covers strategy, design, and engineering for an opt-in testing ground where readers try AI experiments. Its first experiment, Companion, turns a prompt into a ranked list of articles from The Atlantic’s archive. The model never reads the archive directly: an intermediary converts each prompt into a vector-database search, which gave the team control over misuse. For AI-native work, the Midjourney case covers Rooms, a shared creative space now part of Chat.

Atlantic Labs deliberately sits apart from the main product to protect the editorial brand. That is a useful pattern for testing AI with real readers, but it is not evidence of in-workflow integration. MetaLab says it has worked remotely since 2006.

Neuron: AI flags inside a manager’s workflow

Neuron’s Workforce case redesigned a B2B product for understanding hourly staff. AI flags survey comments that mention safety, harassment, or discrimination. Managers triage each flag in the same product: they reply, assign it to a colleague, and record how it was resolved. The case also covers one to three permission tiers and what to show before a new customer has enough data.

This is the closest public example on this page of an AI result written back into a record. The AI is a flagging layer, not a conversational assistant, and the case is undated. Neuron lists San Francisco, California.

Punchcut: assistants and machine-learning platforms

Punchcut’s Google AI case says the team helped integrate Gemini into a range of Google experiences and helped design, test, and build a premium assistant service. Its AWS SageMaker Studio case describes research, a no-code interface, and a shared design system, worked on in parallel with AWS engineering.

Consider Punchcut for a large platform where the assistant’s behavior is itself a design question. The Google page names few specific screens, so ask which work shipped and which was vision. Punchcut lists San Francisco, California.

reloadux: agents with review gates

The eminnt case describes an AI content platform built and owned by tkxel. reloadux designed a workflow where agents, not forms, move the work, with visible handoffs, states, and review gates, and an expert approves every piece before it publishes. Published results flow back into the system, and the assistant reports where each piece stands.

Consider reloadux when several roles must inspect AI work before it proceeds. Because the product is owned by tkxel, ask for an external client reference if that matters to you. reloadux lists Reston, Virginia.

ustwo: conversation as the product

ustwo’s Inflection AI case describes a year-long embedded team on Pi covering strategy, research, UX, and frontend engineering. It helped bring the first iOS app to market, co-created Discover to start conversations, and helped launch Call Pi for voice. Weekly interviews and usability tests fed each release.

Pi is a consumer assistant, and its case does not cover enterprise permissions or actions on customer records. ustwo’s studios are in London, Malmö, New York, and Lisbon.

AI-native product or AI feature in existing software?

The two jobs reward different evidence. For an AI-native product, the interaction model is open, and the useful cases show how a team handled conversation, voice, or creative tools from the first release: ustwo on Pi, MetaLab on Midjourney, Lazarev.agency on Rhea, reloadux on eminnt, and Humbleteam on Afternoon. For a feature in existing software, look for a case where an established product kept its screens, roles, and design system while AI arrived: Eleken on Siena, Humbleteam on the Cluely redesign, Neuron on Workforce, and Punchcut on Google. MetaLab’s Atlantic Labs shows a third route: test AI with real users beside a trusted product before it enters the main one. Several agencies cover more than one of these, so judge the closest case, not the agency’s label.

How should you choose for an existing product?

Give each candidate the same current screen, a sample of the data the AI can use, and the place where the finished work must end up. Then ask them to show the closest case against the symptom you are trying to fix. The scenarios below are illustrative.

Take one illustrative support product. The assistant drafts a correct reply but attaches it to the wrong account. A proposal that covers only the chat interface leaves that unresolved. Ask the candidate to show where the account is visible, when it can change, and how the user undoes the mistake.

For a deeper brief, read the guides to AI search in an existing product, agent approval and recovery, and permissions in enterprise AI. If the feature already exists but people avoid it, an AI copilot UX audit defines the problem first, and our comparison of copilot adoption audit agencies covers who runs one.

End the selection with a written split of work: the agency’s deliverables, your team’s technical responsibilities, and the open questions the first engagement will test. You should be able to trace one task from the existing screen, through the AI result, and back into the record. If a candidate can walk you through that path on your own product, the feature is less likely to feel bolted on.