Dovetail Sun’s Out: Product deep dive

A Digital Twin is an Agent that answers as one of your customers. It’s not purely synthetic, and it’s not a general model guessing what a customer might say. It draws on the calls, tickets, interviews and feedback already sitting in your workspace, and it’ll cite the quote behind any answer you push it on. You choose how tightly to scope it, and can provide specific instructions on how you want it to act. You could scope it to anything relevant to your business, including things like particular market segments, a user persona, someone from your team, or a spend band leveraging enriched data from your CRM. It doesn’t have to be a customer either—some we use at Dovetail include an Enterprise Twin, an ‘Ask leadership’ Twin, and an ‘Ask Design’ Twin our design team built so people could settle low-stakes design-system questions without waiting on a designer. We even experimented with a Digital Twin of our CEO built on the full history of his strategy documents, Slack activity and customer calls, so people from the team could get a quick opinion on something, grounded in how he thinks, if he was unavailable.More in Digital twins

A traditional persona is a document. Someone researches it, writes it up, and it holds until the market moves—then it needs doing again, and new archetypes tend to show up faster than anyone can refresh a persona document or slide deck. A Twin is the same idea without that maintenance problem, because it reads live data and reflects the full suite of data in your workspace. That means your Twins can update in real time, based on new customer data entering your workspace that reflects how those profiles actually think in the market. Plenty of teams just use Twins as their personas. You can run both, too: twin a named account, then attach a broader archetype as a reusable Skill. That’s how we handle segments internally. Customer data comes in from several sources, Fields map it to each archetype, and there’s a Twin sitting behind each one.

Chat starts cold every time. You’ll get somewhere good with it, but it usually takes a few passes to steer it toward the right data and the right framing. A Twin already has its scope, Instructions, metadata and Skills set before you type anything, so the first answer is normally the useful one, and it stays in character for the whole conversation. The bigger difference is that it’s reusable. One person builds it, and anyone with the right permissions can turn it on in Chat and ask. Nobody else has to learn how it was put together.More in Chat

The diversity of use cases we have seen for Digital Twins in such a short space of time is really interesting. We have seen customers establish Digital Twins for their most important personas, particular market segments, particular user groups within their products, and more. They are then using those Digital Twins for everything from getting feedback on a design concept, to customer success managers role playing upcoming calls with a Twin of a specific customer, to sales reps role playing pricing negotiations with an Enterprise Digital Twin built on a recording of every pricing negotiation that exists in the workspace, and marketing teams checking whether a message lands. We also see many customers using a Digital Twin of their customer base to get live responses from their customers during internal planning or strategy meetings. It is the evolution of the ‘empty chair in a meeting to represent your customer’ concept.More in Digital Twin use cases

No. And it’d be a mistake to treat one that way. A Twin is a first pass. Our designers still run structured concept testing with real customers, and the difference is they walk in already knowing which assumptions are shaky, so the sessions and the recruitment budget go further. Where a Twin earns its place is when recruitment is the bottleneck. Niche roles, low response rates, users nobody can reach quickly. Especially when you’ve already run months of interviews with those exact people and that work is sitting there unused. You can ask now instead of in six weeks.

It comes down to the Instructions. You can filter by Field and by metadata, or point the Twin at particular projects and data sources. You can also scope it to a person—our research coach Twin is pointed at our own researcher, using her calls as the benchmark for what good looks like. And if your Fields aren’t tidy, you don’t have to pretend otherwise. Describe the characteristics you care about and let the Agent search the workspace semantically for data that matches. That route tends to work better for the groupings nobody ever got round to formalising.More in Guide to writing instructions

Three parts. Instructions cover what it does and the context it can see. Skills cover how it works, and they’re reusable across Agents: a report format, a tone-of-voice guide, a renewal-prep routine, an archetype. Tools cover what it can reach, from Dovetail’s own tools to web search and external services over MCP, each one switchable. Then you decide when it runs. On a schedule, say every Friday synthesising the week’s feedback into Slack. On something happening in Dovetail, like data landing in a project or a tag going on. On something happening elsewhere, like a CRM record changing. Or on demand, when someone asks.More in Agents overview

Every answer cites what it came from, so you can ask for the source and check it instead of trusting the tone. A Twin only speaks to the data you’ve attached, and the guardrails are built so it tells you when no customer has covered something rather than filling the silence. It won’t forecast behaviour either. It reflects what people have actually said. In practice, most underwhelming answers turn out to be a scoping problem rather than a model problem, and hitting ‘Run now’ is the fastest way to catch that before anyone leans on it.

You can ask Digital Twins and Agents to leverage both, and there are a number of different options. Where quant and call data already sit side by side in your workspace, describe the cut you want in plain language in the Instructions—feedback from enterprise accounts in a particular region, say—and slice across both. An example we discussed covered someone pulling in PDF reports of quantitative analytics for a particular market segment by uploading those files in to Dovetail and asking an Agent to consider them alongside the qualitative data that lives in Dovetail. You can also leverage quant data that lives elsewhere. At Dovetail our Digital Twins and Agents are often querying our data warehouse directly, or reporting that lives in our BI tool, Hex.

Themes are Ideas now, data points are Evidence, and both tabs stay where they were. The real change is ranking. Channels reads what comes in continuously and ranks Ideas by commercial signal instead of volume, so the loudest request stops winning by default. With your CRM connected, every piece of feedback ties back to a contact and a company. That means an Idea with less Evidence behind it than its neighbours can still sit near the top, because of the revenue attached to it. Channels 2.0 is rolling out in beta—you can join the waitlist at dovetail.com/channels.More in Channels

It classifies everything that arrives, then ranks on business value instead of frequency. Because feedback associates to contacts and companies through your CRM, you can sort Ideas by the revenue sitting behind them, and see which sources a signal is showing up in: support tickets, sales calls, app store reviews. Something appearing in every channel at once reads very differently to something stuck in one. You can also see how the Evidence behind an Idea is trending, so a small signal that’s accelerating hard doesn’t stay invisible until it’s a problem.More in Build a Channels strategy

Yes, and that’s deliberate. Open any Idea and you get the Evidence underneath it: the full verbatim, the whole support thread rather than a clipped quote, and whatever your CRM knows about who said it. Not every piece of feedback has to belong to an Idea either. You can go at the raw feedback directly, move Evidence between Ideas when the grouping is off, and build filtered views by field, sentiment or priority so different squads read the same channel their own way. With AI in the loop, being able to get back to the source matters more, not less.

This is the bit teams tend to underrate. Once an Idea has been built and released, you can ask Chat to draft a note to everyone whose feedback fed into it, explain what actually shipped, and pull the associated contacts out as a list. What comes back reads like a personal message rather than a release announcement, so people hear that the specific thing they raised is done. We built it to close that loop, on the bet that customers who hear back are more willing to tell you the next thing.

Yes. From an Idea you can open a Chat to dig further, drop it into a doc, or push it out to where the work happens—Jira or Linear for tickets, and other MCP destinations including Claude, Claude Code, Cursor and Figma Make. Attaching an Idea to an epic means the insight travels with the work instead of being retyped into a ticket and losing its Evidence along the way. Teams can comment and react on Ideas too, which keeps the conversation that usually happens in Slack sitting next to the Evidence it’s about. Ours go into Linear.

Yes, both directions. This launch added ten integrations, including Gong, Zendesk, Salesforce Service Cloud and Pendo, plus our first data warehouse connector for Snowflake. That takes the total past thirty. The warehouse one solves a specific problem: you often can’t get credentials to a source system yourself, but the data has already been centralised somewhere, so you or your data team can import a table directly. Past bulk import, Chat and Agents can reach connected tools at the moment you ask, which means numbers that will never live in Dovetail can still shape an answer. Auth is OAuth as you, so nothing runs on shared credentials.More in Integrations

Yes. Chat and Agents reach a set of external services over MCP, so analysis doesn’t have to stop where Dovetail stops. A realistic run: ask what customers are saying about part of your product, then ask for a proposal drafted straight into your team’s doc tool. The draft picks up the tone-of-voice rules, vocabulary and acronyms you’ve set in your workspace, so it turns up sounding like your organisation instead of generic. An Agent can also pull quant context from a BI tool mid-task. That’s how we combine numbers from ours with the qual already in Dovetail.

Yes. Dovetail Intelligence is a first-party connector in Claude, ChatGPT, Copilot and Teams, so you can question your customer data from whichever assistant you already work in. It authenticates with your Dovetail login and acts as you, with the same permissions you hold in the product. Our MCP server opened the year with a handful of read-only tools and a couple of hundred sessions a week. It now exposes more than forty tools, and has handled a couple of hundred thousand external sessions. You can push analysis back into Dovetail as a doc as well, so the thinking doesn’t get stranded outside.More in MCP server

Nobody, by default. No standing access for anyone—not support, not engineering, not the ops team. When support needs to see something to help you, you grant it explicitly, and that access is scoped, logged, time-bound, and expires on its own. The only way in is if you open the door. That’s how the storage architecture was built rather than a policy bolted on afterwards, which is why we can say it plainly instead of hedging.More in Security information

No. Generative AI in Dovetail runs through AWS Bedrock on pre-trained models. Your data goes out, an answer comes back, the model learns nothing from it. That was an architectural call made early, specifically so there’d be no ambiguity to explain later. We also hold to zero operator access and zero data retention, and we require the same of our cloud providers and data processors, which rules out models that keep your prompts around for a safeguard window.More in Dovetail AI

SOC 2 Type 2, ISO 27001 and ISO 42001. The 27001 means our information security runs as an externally audited management system, which we’ve held for years with a clean audit history. ISO 42001 says the same thing about our AI: how risk gets assessed before an AI feature ships, and how bias and accuracy are handled, audited by the same kind of third party. That second one is quietly becoming the useful one, because plenty of procurement questionnaires were written before AI and don’t yet ask the right questions.More in Trust Center

We’ve spent years serving customers in heavily regulated environments—banks, airlines, large cloud providers—and the bar those customers set is the bar the whole platform is built to, not a separate enterprise tier bolted on the side. Thousands of customers use Dovetail every day on some of the most sensitive material a company holds: what their own customers actually said to them. Guardrails aren’t a feature in that context. They’re a precondition.