AI for Learning in 2026: What’s New, What’s Next, and What It Means for Educators

Insights from a recent Digital Learning Institute (DLI) webinar with Joe Houghton.

Generative AI is moving fast, but the real shift is not “new chatbots”. It is how AI tools are starting to connect to your work, your files, and each other.

In our February webinar, Joe Houghton joined the Digital Learning Institute (DLI) to share the most important developments educators and L&D teams should pay attention to right now. From AI-generated Word documents and slide decks to connectors, agentic workflows, and assessment redesign, this session was a practical tour of what is already possible and what is coming next.

Below is a recap of the most useful takeaways, with clear examples you can apply to teaching, training, content design, and learning operations.

• AI is shifting from “chat” to connected workflows across tools like Notion, Google Drive, Gmail, and slide and document builders.

• Claude demonstrated a leap in productivity, including creating a polished Word document from a single prompt and generating decks through Gamma.

• AI “connectors” are becoming a major capability layer, enabling agentic workflows where AI can carry out multi-step tasks.

• In education, the biggest implication is assessment: AI detection is unreliable, so the focus needs to move to competence, performance, and experiential assessment.

• Research workflows are improving, including more powerful “deep research” capabilities and better ways to organize and query your own curated resources.

For many teams, the early AI phase looked like this:

1. Ask a chatbot for an outline

2. Copy and paste it into a document

3. Edit it manually

4. Repeat across tools

The webinar’s central message was that we are starting to move beyond that. AI is increasingly able to connect to other platforms and help with multi-step work. This reduces friction and creates opportunities for faster content development, better knowledge management, and more responsive learning support.

Joe described this as a move toward agentic workflows, where AI agents can complete tasks across systems rather than producing isolated answers.

One of the standout moments from the session was Joe’s reaction to Claude 4.6, describing it as the biggest single jump he had seen in some time.

What impressed him most

• Claude produced a fully polished Word document from a single prompt, including professional formatting elements such as a title page, table of contents, headings, and consistent layout.

• This signals a shift from “drafting text” to generating finished outputs that are closer to publish-ready.

For learning teams, this is relevant because many workflows rely on repeatable document formats:

• learning briefs and design documents

• facilitator guides

• stakeholder reports

• policy and governance templates

• program proposals and evaluation write-ups

When AI can produce a stronger first draft in the right structure, designers can spend more time on instructional decisions, quality, and outcomes.

A major part of the webinar focused on connectors, which Joe described as a way to amplify what AI can do by allowing it to work across platforms.

Examples mentioned in the session included connecting AI to:

• Google Drive

• Gmail

• Notion

• slide tools such as Gamma

• workflow tools and automation platforms

Why connectors matter for learning and L&D

Learning work is rarely contained in one place. Content, briefs, stakeholder notes, and assets often live across drives, knowledge bases, and email threads. Connectors point toward a future where you can:

• search and summarize documents across folders

• pull information from past decisions and stakeholder emails

• assemble a brief using content from multiple sources

• generate outputs aligned to your existing templates and style

A key caveat raised during the webinar was privacy and control. The value is real, but granular permissioning is still evolving, and teams need to treat access thoughtfully.

The session repeatedly returned to one theme: AI workflows are becoming joined-up.

Instead of doing tasks in separate tools and stitching them together manually, the goal is for AI to:

• retrieve the right material

• complete steps in sequence

• produce an output that is usable

Joe described this as the early stages of agentic workflows, where multiple agents can operate autonomously, share findings, and assemble more complex outputs.

Practical learning examples of agentic workflows

• “Create a microlearning pack from our policy PDF and last month’s stakeholder notes, then draft a 5-question knowledge check.”

• “Summarize our last 10 learner feedback emails, cluster themes, and draft improvements for the next cohort.”

• “Generate a slide deck and facilitator guide for a workshop using our Notion knowledge base and existing brand theme.”

This is where AI begins to look less like a chatbot and more like a learning ops assistant.

A very practical demo in the webinar showed Claude using connectors to generate a slide deck through Gamma, producing a visually structured set of slides that was:

• branded using an existing theme

• shareable via link

• exportable to PowerPoint, PDF, or Google Slides

For educators and L&D teams, this matters because slides are a constant bottleneck. The promise is not “AI makes slides”, but:

• AI can create a structured first pass

• you refine for accuracy, pedagogy, and tone

• you spend less time on layout and starting from blank

This can be especially useful for:

• internal training decks

• webinar slide prep

• program overviews

• workshop materials

• stakeholder presentations

Another practical section demonstrated Claude inside Excel as an add-in. The example used a cashflow forecast to show how AI could:

• analyse patterns and risks (e.g., seasonality, “cash cliff”)

• provide commentary and guidance based on the spreadsheet

For learning and business teams, the bigger point is that AI is moving closer to where work actually happens. In L&D, this could translate to:

• analysing evaluation spreadsheets

• summarizing survey exports

• checking cohort performance trends

• drafting insights for monthly reporting

Joe highlighted NotebookLM as a preferred tool for working with curated sources and building a structured knowledge base. A key update discussed was a way to query across multiple notebooks through Gemini “gems”, effectively allowing:

• multiple notebooks to be used as a combined knowledge set

• guided learning style querying across curated material

This matters for learning teams who already manage:

• policy repositories

• program documentation

• content libraries

• research sources

• facilitation notes

When AI can query your own trusted sources first, you reduce hallucination risk and improve accuracy for internal workflows.

The webinar also referenced updates to “deep research” style features, with improvements like:

• larger context windows

• better targeting of sources

• the ability to interrupt a research job and add instructions mid-way

The common thread across tools is friction reduction:

• fewer repeated prompts

• fewer restarts

• more control over sources

• better outputs from the same effort

For educators, this means research and synthesis work can become more reliable and less time-consuming, especially when you are building learning resources or updating curriculum.

A particularly important section of the Q&A addressed AI detection tools and plagiarism.

Joe’s position was clear:

• AI detection tools are not reliable due to high false positives

• institutions are exposed if they rely on detectors to accuse learners

• the more sustainable approach is to redesign assessment

The direction suggested in the webinar

• shift toward competence and performance-based assessment

• use experiential and applied tasks

• assess learners in ways that require judgement, context, and real-world application

This aligns with a wider trend in education and professional learning:

• scenario-based assessments

• workplace simulations

• portfolio evidence

• reflective justification and decision logs

• authentic tasks tied to real outputs

The underlying message: if AI can draft an essay in seconds, essays alone cannot be your primary proof of competence.

If you want to act on this webinar’s ideas without getting overwhelmed, focus on three practical moves.

1) Pick one workflow to improve

Choose a single repetitive task:

• turning notes into a brief

• drafting workshop slides

• summarising learner feedback

• writing a learning resource

Then test AI on that workflow end-to-end.

2) Use tools in combination

A recurring theme was using different tools for what they do best. For example:

• one tool for research

• one tool for synthesis and writing

• one tool for slides

• one tool for knowledge base querying

3) Start redesigning assessment for performance

If assessment is part of your remit, begin by:

• identifying tasks where AI can produce a “good enough” output

• redesigning the assessment so that learners must demonstrate judgement, application, and decision-making

• incorporating authentic constraints and real contexts

What are AI connectors?

Connectors allow an AI tool to access and work with other platforms (for example, a drive, knowledge base, or slide tool). They reduce manual copy and paste and enable multi-step workflows across systems.

What is an agentic workflow in AI?

An agentic workflow is where AI can carry out a sequence of actions autonomously, such as retrieving information, generating an output, and completing tasks across tools, rather than only responding in chat.

Can AI reliably detect AI-written assignments?

No. AI detection tools have high false-positive rates and are not considered reliable for proving misconduct. The safer approach is assessment redesign toward performance and competence.

What types of assessment work better in an AI-enabled world?

Applied, experiential, and competence-based assessments tend to be more robust. Examples include scenario-based tasks, portfolios, simulations, reflective decision logs, and authentic work outputs.