Designing AI experiences

Sage builds software to help small and medium businesses thrive, from managing cash flow to paying employees and staying compliant. Our AI agents must reflect this mission: empowering users to make better decisions, save time, and focus on what matters most.

To ensure that our customers feel supported in adapting to new AI features, we must:

1. Meet user needs
2. Help users feel in control
3. Build trust with users over time

The following principles guide how we aim to design, deploy, and evolve AI agents within Sage products.  

1. User needs

Empowerment 

Our agents should act as professional partners, not robotic assistants. They support the customer’s expertise and help them work more effectively, without taking control. 

Security

Sage is a connected platform. Our agents must integrate seamlessly with Sage products. This includes utilising secure APIs, managing permissions effectively, logging actions, and adhering to compliance standards. The experience should feel native and secure. 

Credibility

Multi-user environments, compliance needs, and traceability are everyday requirements. The agent must work within that context where customers are involved. 

2. Control

Guided workflows

Our customers use Sage to achieve specific goals: submit returns, pay staff, and reconcile bank feeds. AI agents should guide them through these tasks with clear, structured steps. Intelligence should make the process quicker, not more confusing, with clear checkpoints, decisions, and options for humans to take control when needed. 

Predictable, reliable, and auditable 

Sage customers rely on us for accuracy in their books, their payroll, and their tax returns. Our AI agents must behave consistently and transparently, applying intelligence at well-defined moments. We avoid “black box” behaviour and instead build auditable workflows where similar inputs produce reliable, trusted outcomes. 

Contextual understanding 

Running a business is complex - our AI agents must help manage that complexity. This involves remembering ongoing tasks (such as VAT returns), understanding financial history, interpreting business rules (like payroll deadlines), and maintaining context across multiple Sage products (e.g. from Accounting to Payroll). We design agents that feel like they “get” the user’s business, not just the last thing they completed. 

3. Trust

Ethical guardrails 

Working with payroll, tax, and business advice means Sage must hold itself to a higher standard. Bias, manipulation, or opaque reasoning can have real-world consequences. To support this, we apply the Sage AI Trust Label - a public pledge to ensure safe, transparent, and responsible AI. This offers clear criteria to assist our teams in developing trustworthy agents and helps customers understand how our AI systems function and are governed. 

Validation and error handling

Business data can be messy, and mistakes happen. Our agents are built with this in mind. They verify inputs, manage errors smoothly, and escalate to humans when needed. The goal is to complete tasks efficiently, not just quickly. 

Built-in feedback loops 

AI agents are not static. Continuous improvement based on real usage ensures they stay aligned with customer needs and product evolution.

Currently, there are 3 established models to consider for human-AI interaction:

1. Supportive
2. Focused
3. Integrated

Each has a different emphasis and persistence of interaction relative to the main context (product or suite interface). They can be used separately or together, letting users navigate between them based on their needs.

1. Supportive

Provides users with on-demand support through a contextual conversational interface that:

• Sits alongside the main context.
• Can be triggered from content within the main context.
• Provides conversational interaction specifically related to the main context.

Potential features

The AI could create responses by inferring what the user wants based on their context and history, relying less on highly specific prompts and commands.

Past AI suggestions, user queries, and system responses could be grouped and stored for easy access, ensuring continuity and easy reference. 

2. Focused

Provides users with a maximized version of the supportive interface where they can:

• Browse previous conversations.
• Create new conversations unrelated to the main context.
• Navigate back to the main context if needed.

Potential features

Within this environment, components and their data could be generated directly in the workspace, ensuring seamless integration and real-time updates. Humans could interact with the generated UI, modifying and refining it to meet their specific requirements.

Interaction could be multi-modal, combining text input with buttons, sliders, and interactive cards for a highly flexible and dynamic experience designed to accomplish goals, not tasks.

3. Integrated

Provides users with AI features that are embedded into the main context to help them complete small, discrete tasks. It can offer:

• Enhanced components that are added directly to the UI.
• Contextual support without disrupting the user experience.
• Relevant recommendations through in-context UI elements.

Potential features

Embedded elements could highlight critical insights, recommended actions, or task summaries, enabling users to quickly assess and act on key information.

To maintain accuracy and user control, validation elements could ensure that data is only submitted after a human confirms or modifies AI-generated suggestions.

Agents act on behalf of users by automating routine and complex tasks. They also enable users to make proactive decisions by surfacing meaningful insights and empowering them to take the right action.

However, it's essential that users maintain control. They must have the ability to enable or disable agent capabilities and choose where they want to validate, check, or approve actions before execution.

This configurable approach builds trust by ensuring users remain in control of critical business decisions.

The following patterns might inform how we design for these interactions.

Maintain contextual awareness

• Context summary. Shows current goals, active agents, and relevant data.
• Conversation history. Scrollable log of interactions, decisions, and user inputs.
• Goal tracker. Visual progress bar or checklist showing how close the user is to achieving their goal.

Reveal agent activity

• Live activity feed. A real-time stream of agent actions, similar to a system log but human-readable.
• Task timeline. Chronological view of tasks initiated, completed, or paused.
• Agent profiles. Each agent has a visual tile showing:
• Current task
• Status (“In progress”, “Awaiting confirmation”, “Completed”)
• Recent actions or decisions

Request feedback and validation

• Clarification requests. Ask users to clarify ambiguous requests.
• Confirmation prompts. Before executing critical actions, the system asks for user confirmation.
• Editable suggestions. Agents propose actions or content that users can review and edit before approval.
• Inline feedback buttons. Allow agents to iterate and learn from user feedback.
• 👍 / 👎 for agent responses
• “Was this helpful?” prompts
• “Improve this” or “Try again” options

Enable adjustment and correction

• Undo/Redo buttons. Make actions reversible.
• Adjustable parameters: Users can tweak inputs to better match their goals.
• Agent traffic lights: Allow users to start, stop, and pause agent activity.

Allow multi-modal interaction

• Text input. Natural language or guided prompts.
• Voice commands. For hands-free interaction.
• Quick actions. Buttons or dropdowns for common tasks.
• Draggable elements. For organizing agents, tasks, or data.

Explain failure and recovery

• Failure messages. “The report could not be generated because the data source is missing.”
• Recovery suggestions. “Would you like to upload the missing file or request a different report?”
• Fallback options. Manual workflows or human support escalation.

Agentic AI delivers the most value when it operates independently.

In order to fully embrace this technology and truly harness its power, we will eventually need to shift from control-based workflows to trust-based interactions.

That shift will bring its own set of design obstacles that go far beyond traditional UX concerns.

Here are some of the challenges we might face.

Trust is fragile

Agentic systems make decisions independently, often without user visibility. If users don’t understand the “why” behind actions, they lose confidence, even if the system is technically accurate.

• Users distrust opaque logic and overconfident AI behavior.
• Minor mismatches between user expectations and system output feel like significant errors.
• A single unexplained decision can break long-term trust.

Solution

Design for visible, contextual reasoning that aligns with user expectations, even when the intent is implicit.

Recovery must be effortless

When an agent acts independently, users need quick ways to reverse or redirect without having to reset everything. The following scenarios pose a challenge to the workflow of AI agents.

• Errors without undo options increase user frustration.
• Backtracking often breaks the interaction flow.
• Users feel powerless when AI overrides go unresolved.

Solution

Make every AI action reversible with simple user inputs and clear fallback paths.

Intent is often unclear

Users express goals, not tasks, and often without enough detail. But to be truly effective in saving time and effort, agents can’t pause every time for clarification.

• Inputs are vague, but outcomes are expected to be precise.
• AI must infer intent from partial signals and past behavior.
• Asking too many questions can be frustrating for the user.

Solution

Build conversational models that can clarify instructions intelligently and enrich user intent with context.

Voice UI: People do not always say what they mean or speak in a structured way that is easily understood by an LLM. Do not use voice to replace visual interfaces entirely, but to complement them, especially for quick, low stakes interactions where precision is not critical. The moment complexity increases, people want to see options, not just hear them.

The interface feels unfamiliar

Agentic UIs don’t always follow traditional software patterns or use components with recognizable affordances.

• Users often don’t know how or where to begin.
• Lack of structure creates hesitation.
• Success favors informed users, not casual ones.

Solution

Lower cognitive load with innovative prompt suggestions, starter templates, and contextual cues.

Ethical design demands long-term thinking

Autonomy introduces deeper responsibilities around fairness, privacy, and inclusion.

• Privacy risk. Autonomous agents may access sensitive data, such as financial information, emails, or health records, and may unintentionally expose or misuse it without clear boundaries.
• Bias risk. AI decisions, such as hiring, loan approvals, or support prioritization can reflect biases in training data, reinforcing systemic inequalities based on location or demographics.
• Inclusivity risk. Interfaces designed for English-speaking, tech-savvy users may exclude people with disabilities, neurodivergent traits, or non-native language backgrounds from fully engaging.

Solution

Embed ethical checks into design workflows to ensure privacy, fairness, and inclusivity are prioritized from the start.

User interfaces aren’t disappearing—they’re evolving.

When mobile became ascendent, new UI components were created to take advantage of what it could do. AI is no different. Great UI is more important than ever so that we can build better mental models of what this technology can and cannot do, and help our users to effectively harness its power.

We can neither understand nor communicate the capabilities and limitations of AI with just a text box.

Multi-modal beats mono-modal

Voice assistants are adopting screens

Humans consume visual information faster than spoken information.

We can read around 250 words per minute but speak or listen to only around 150. Why would you listen to an AI assistant read a list aloud when it’s so much faster to glance at a screen showing the list? Scanning items on a screen takes a fraction of the time it would take to hear them.

Combining AI with visual cues can give a better experience than voice alone. A UI should be as visible as it needs to be.

Prompting isn't for everyone

Not everyone can (or wants to) type elaborate prompts all day.

Many people are not fast typists. And crafting the “perfect prompt” to elicit an effective response from AI is a new kind of literacy that not everyone has—or will ever have. Relying completely on chat UI will exclude and frustrate a lot of users.

Sometimes tapping a button or swiping a screen is just easier and more universally accessible.

Users need UI

Affordance

Users need interfaces designed to harness AI’s power without dumping the burden on them to figure it out.

Invoking AI functionality through a well-designed UI affordance means less cognitive load on the user. The interface remains in the foreground, guiding the AI and the user. It’s a partnership: the AI does the heavy lifting, but the UI gives the user understandable control and feedback.

People need tools, not just text boxes.

Hyper-personalization

AI can tailor layout, content, and functionality to each user’s needs in real time.

Every interaction can be an individualized experience, with interfaces adapting on the fly to a user’s context and preferences. One user's app might rearrange its layout or surface different options than another user's does, because the AI behind it knows what each of them likely wants next.

This kind of dynamic, adaptive, context-aware UI can satisfy users more than a one-size-fits-all chat prompt can.

Agents are becoming users

AI accessibility

We’re moving towards AI systems that can operate apps and other software.

Researchers and companies are building AI agents that can click, scroll, and type in apps on your behalf, essentially using the UI just like a human would. As a result, designers need to create even more clear and consistent UIs because now both humans and AI bots need to understand them.

Designing for agent users could improve accessibility for all users.

Human oversight

Many agents will of course bypass the UI and use APIs to perform actions.

But even if an agent performs most of a multi-step task behind the scenes, the user will still want visibility and control over critical decisions and actions. They'll need a UI that shows what the AI is doing, and that allows them to confirm, adjust, or reverse important steps.

Users aren’t going to accept invisible AI systems. They’ll demand dashboards, notifications, and other controls to monitor and manage what AI assistants do.

Voice

Agents speak in plain language that's clear, concise, and well-organized.

They do not use jargon. When an agent needs to use a specialist term, it gives a clear explanation to the user.

Tone

Agents speak with a conversational tone that respects professional boundaries.

Responses should have a consistent tone. This allows agents to build user trust and establish their reliability.

However, there may be situations where the tone needs to be adapted. For example:

• If an agent has failed to answer a user request, the tone could be apologetic.
• If a user has achieved a goal with the help of an agent, the tone could be celebratory.
• When tax filing is due, agents should communicate in a more straightforward tone so that users know exactly what the next step of the conversation means for them.

Reflect our values

Human

Agents should be conversational and helpful. They should support users in their tasks and avoid verbose responses. Conversations should be friendly, but concise.

Simplify

Agents should speak the user's language. They should respond in plain language with answers that anyone can understand. They should guide users through necessary next steps with well-structured responses. For example, by using bullet points or headings.

Trust

Agents should be transparent about responses and whether or not a user can rely on the integrity of a response. They should always cite references when offering information that the user may rely on to make decisions.

Bold

Agents should help users to maximize the potential of their business. They should make bold, empowering calls-to-action for users to complete tasks faster and more accurately.

For example, prompting the user to automate a task after successfully completing it the first time.