Celine Haaverstad, COO at SoftSync AI – Interview Series

Celine Haaverstad, COO at SoftSync AI, helps lead the company as it develops a new approach to customer relationship management built around AI agents and automation. Her work centers on making CRM systems less dependent on manual data entry and repetitive administration, while maintaining appropriate controls around permissions, customer data, and autonomous actions. Haaverstad advocates for an approach in which AI is integrated directly into everyday business workflows rather than added as a standalone chatbot, with humans retaining responsibility for sensitive decisions, customer commitments, and actions that require judgment. SoftSync’s own website identifies Haaverstad as the company’s COO.

SoftSync AI is a Norwegian-developed CRM platform designed to bring customer records, email, meetings, sales pipelines, and follow-up workflows into a single workspace. Its Autopilot system can run repeatable tasks in the background, while an MCP server allows external AI assistants such as ChatGPT, Claude, and Gemini to interact directly with CRM data and perform permitted actions using the user’s existing access rights. SoftSync also provides direct email and calendar connections and a REST API for conventional software integrations. The company says customer data is stored and processed in the EU on Microsoft Azure in Stockholm and is neither sold nor used to train AI models.

Can you take us back to the origins of SoftSync AI? What problems with traditional CRM systems convinced the team that there was an opportunity to build something fundamentally different around AI?

SoftSync started with a simple observation: the sales day was spread across too many disconnected tools. A lead might begin on LinkedIn, move into a spreadsheet, continue by email, become a calendar booking, generate meeting notes, and finally need to be entered into a CRM.

Traditional CRMs are often built around a fixed structure and a large amount of manual administration. Teams have to adapt their process to the system, wait for an admin to change fields, and repeatedly copy context between tools.

We wanted to build a sales workspace around the actual flow of work: capture the lead, keep the conversation connected, hold the meeting, record what happened, and make the next action clear. AI is valuable in that system because it can remove the administrative handoffs without becoming another isolated chatbot.

SoftSync includes capabilities such as Assistant, Autopilot, API and MCP within its paid plans, while capacity scales through factors such as seats, records, groups, mailbox connections and monthly activity. Why did you choose to make automation a core part of the product rather than reserving it for a higher-tier add-on?

The important distinction is between the agent and the amount of work it performs. Autopilot should be part of the product’s core workflow because automation is most useful when it understands the records, fields, permissions, conversations and processes already in the workspace.

We don’t want automation to be a luxury reserved for the largest customers. The current model is that the core CRM, Assistant, Autopilot, API and MCP capabilities are included in the paid plans. What scales with the customer is the room and usage: seats, records, groups, mailbox connections and monthly activity.

That gives customers predictable access to the product while still letting usage be metered fairly. Larger teams and higher-volume workflows move to plans with more capacity or custom limits, rather than having the most useful capabilities removed behind a higher permission tier.

Rather than limiting customers to an embedded chatbot, SoftSync runs an MCP server that allows assistants such as ChatGPT and Claude to work directly with the CRM. What does this architecture make possible that a conventional embedded AI assistant cannot?

An embedded chatbot is limited by the interface and model chosen by the vendor. MCP gives an assistant such as ChatGPT or Claude a standard way to work with the live SoftSync workspace.

That means the assistant can use the customer records, related activity, custom objects, fields, views, notes and permitted workflows already in the CRM. It can answer a question, prepare a meeting brief, propose a record update, or start a supported workflow without the user having to manually copy context into a separate chat.

The important point is not that an embedded assistant can never take action. It is that MCP gives customers more choice of assistant, a more open architecture, and a direct connection to the CRM system of record.

Your MCP implementation provides AI assistants with operational guidance on how SoftSync works before they begin using its tools. Why is that context important, and what have you learned about making agents more reliable when interacting with business systems?

Tool schemas explain what an operation accepts. They do not fully explain how the CRM should be used safely.

Before using the tools, SoftSync provides operational guidance such as discovering the live workspace schema, using internal field names, resolving related records, checking permissions, and following the correct order for actions like sending an email.

That context makes agents more reliable because they are less likely to guess a table, field, record ID, relationship or recipient. The broader lesson is that reliable agents need more than tool access: they need the business rules, boundaries and verification steps around those tools.

Giving an AI agent the ability to create records, move deals, send emails or modify customer data introduces important questions around control. How does SoftSync approach permissions, confirmations and other safeguards when agents are allowed to take real-world actions?

SoftSync does not give an agent a separate path around the CRM’s permissions. The connection uses the user’s account, OAuth, workspace membership, group access and role. A viewer cannot become an editor simply because an assistant is connected.

We also encourage a read-first workflow: identify the exact record, inspect the current state, propose the change, and review the target, values, recipients and side effects before approving it.

For higher-risk work, the system can require human review or confirmation. That matters especially for sending customer-facing email, changing deal stages, modifying sensitive records, using personal mailboxes, or starting connected external operations. Autopilot is well suited to repeatable checks and reminders, but people should remain responsible for ambiguous, irreversible or relationship-sensitive decisions.

Autopilot is designed to handle routine work in the background. What types of CRM workflows are particularly well suited to autonomous execution, and where do you believe humans should remain firmly in the loop?

Work with a clear trigger, clear inputs and a repeatable outcome. For example:

• reminding an owner when a deal has gone quiet
• alerting a seller when a high-value opportunity is created
• classifying and routing inbound email
• creating a follow-up after a proposal is sent
• checking records on a schedule
• notifying a person when a known condition is met

Humans should remain firmly involved when the task involves judgement, an uncertain identity match, a customer-facing commitment, a sensitive stage change, deletion, or an external action that cannot easily be reversed. The goal is not to remove responsibility; it is to remove the repetitive work around responsible decisions.

SoftSync describes the AI assistant as an integration layer rather than trying to route every task through hundreds of individual connectors. How far can that model go, and where do traditional APIs or dedicated integrations remain the better approach?

MCP can cover a broad class of assistant-led work because it exposes the workspace as structured, permissioned tools rather than requiring a separate custom interface for every assistant.

It is particularly useful for questions, record work, meeting preparation, workflow execution, and actions that benefit from natural-language context. It can also work with connected operations where those accounts and permissions are available.

Traditional APIs and dedicated integrations remain better for deterministic, high-volume, low-latency processes; continuous synchronisation; strict service-account governance; webhooks; bulk data movement; or systems where an AI assistant should not be involved at all. Our position is not “everything must go through AI.” It is to use the simplest reliable connection for each job.

SoftSync is designed to work with multiple AI assistants rather than tying customers to a single model provider. As models from OpenAI, Anthropic, Google and others continue to evolve, how important is model independence to the long-term architecture of an agentic business application?

A business application should not force customers to redesign their data and workflows every time the preferred AI model changes. SoftSync keeps the durable layer, including records, schema, permissions, workflows and audit expectations, inside the workspace, while allowing different assistants to work through a common interface.

That gives customers more choice and lets us benefit from progress across the model ecosystem. It also requires discipline: model providers differ in tool use, approvals, context limits, reliability and cost. Model independence is not about pretending all models behave identically. It is about keeping the customer’s business system independent of any one provider.

Customer data becomes particularly sensitive when AI agents can actively operate on it. How is SoftSync approaching privacy, data residency and governance, particularly for European customers, while still giving AI enough access to perform useful work?

The starting point is least-privilege access. Each person connects their own account, and the assistant inherits the workspace and group permissions that account already has. Private groups remain private, viewers remain read-only, and connected mailboxes or calendars are subject to their own access controls.

SoftSync AI is hosted in the EU on Microsoft Azure in Stockholm. We never sell customer data and never train AI models on customer data. Every sub-processor is also listed along with the region in which it operates.

We also separate access from action. An assistant may be able to read a record without being allowed to change it, and a connected workflow may require additional inputs or human confirmation before it runs. For sensitive work, teams can begin with read-only requests and verify the result against the source record.

The broader product principles are data minimisation, explicit access, user-level authentication, controlled visibility, reviewable actions and clear handling of connected accounts. Recording meetings also requires appropriate participant notice and compliance with applicable consent and privacy laws.

Looking several years ahead, do you expect the traditional idea of maintaining a CRM, manually entering data, updating deal stages and creating follow-up tasks, to largely disappear? What does the CRM become when AI agents are responsible for much of that operational work?

The manual CRM does not disappear completely; its role changes. People should spend less time typing routine updates and more time reviewing the context, exceptions and decisions that matter.

The CRM becomes a shared context and action layer: a place where customer history, communication, meetings, ownership, permissions, workflows and next steps remain connected. Agents can capture information, keep records current, identify work that needs attention, and run repeatable processes in the background.

Humans still own the relationship, judgement, commitments, pricing and important decisions. The future CRM is less a database people constantly maintain and more a trusted operational system that people supervise.

Thank you for the great interview. Readers who wish to learn more can visit SoftSync AI. Unite.AI readers using this link can access a 30-day trial and receive 30% off.