After a workshop, the whiteboard is photographed and the photo then sits in a chat for weeks, because nobody wants to retype it. Today's phone and office apps read the image, keep lists and table structure, and hand over editable text or an Excel table in under a minute. The benefit is simple: no retyping, nothing lost, notes reach the team the same day.
Problem & value
- Today: whiteboards, flipcharts, printed price lists, tables in a PDF screenshot or on a supplier's web page are retyped by hand or stay as pictures that cannot be searched or edited.
- AI value: classic OCR returns flat text; modern vision models and built-in office features recognise rows, columns, bullet hierarchy and even rough handwriting, and can also tidy the result (fix typos, group by topic, extract action items).
- Effort: this is a personal, one-off conversion. It needs no project; a short how-to for staff and a decision on which approved app to use is enough.
- Benefit figures: no reliable independent time-saving figures exist for this specific task, so measure it yourself (see KPIs).
Variants
- Workshop whiteboard to minutes: photo of the board, output grouped by topic with owners and dates.
- Printed table or report to Excel: a photographed price list or delivery note table becomes rows and columns.
- Screenshot to table: a table from a web page, a legacy application without export or a PDF that cannot be copied.
- Sticky-note wall: clusters become a grouped list.
- Private use: recipes, handwritten notes, whiteboards from courses.
- Not this use case: systematic, high-volume processing of many documents is DOC-014 (document parsing) or VIS-018 (industrial OCR); invoices belong to DOC-001.
How it works
photo / screenshot (phone camera, snipping tool)
-> crop and straighten (flatten glare, one table per image)
-> OCR / vision model (text + layout + table structure)
-> optional LLM step (clean up, group, extract action items)
-> human check against the original
-> paste into Excel / Word / OneNote / ticket
- Key components: OCR for characters, a vision-language model (VLM) for layout and context.
- Human in the loop: always compare numbers and names with the picture before the data is used. Handwriting and similar-looking digits are the main error source.
Data preparation
Which data: the image itself. No training data is needed.
Where it lives: phone gallery, chat apps, shared mailbox, screenshot folder.
Typical gaps: glare and shadows on whiteboards, photos taken at an angle, several tables in one picture, tiny handwriting, and photos that contain people, customer names or confidential plans.
Preparation steps: 1. Agree on a photo routine: straight-on, whole board, good light, one table per picture (5 minutes of instruction). 2. Define a rule for what must not be photographed or uploaded (customer lists, HR content, trade secrets) as part of the AI policy (GUIDE-007). 3. Name one approved app and one place where results are stored (Teams channel, SharePoint, Nextcloud). 4. Keep pictures out of private chat apps; delete the photo after transcription if it contains personal data.
Tools: the phone camera's document scan mode, the OneDrive scan function (Microsoft Lens was retired in early 2026), Google Drive scan.
Rough share of effort: low, a few hours for guidance and a policy sentence. See GUIDE-003 for the general check.
Implementation paths
Which path for whom
- SMB - Path A: mature features are already included in Microsoft 365 or Google Workspace, and the task is standard. A custom build would cost more than it saves.
- Enterprise - Path A: a commodity function; use the existing suite features and the approved corporate assistant.
- Developer - Path C: open-source OCR and small VLMs run on a laptop or HW1 and can be embedded into own tools with full data control.
- Private - Path A: phone apps and free OCR features; local apps are an option for sensitive notes.
Path A - Buy (off-the-shelf)
- Microsoft 365: Excel "Data from Picture" converts a table image to cells; you review and correct the result before inserting it [1]. The Microsoft 365 mobile app offers an Image to Table action [1]. Copilot can also be asked to turn an uploaded image into a table, depending on licence and tenant settings; the feature availability varies, so test it in your tenant.
- Google Workspace: Google Lens copy-text on phone and Chrome, and opening an uploaded image with Google Docs for OCR, are common routes; Gemini can then restructure the text [2]. Third-party guides describe these routes, so check the current menus.
- Other: Apple Live Text and Notes scan on iPhone, ChatGPT, Claude or Gemini apps for restructuring.
- GDPR note: check whether the tenant processes in the EU (Microsoft EU Data Boundary, Google data regions) and that an AVV exists. For consumer apps outside the company contract, do not upload anything confidential.
- Choose A when: you already pay for Microsoft 365 or Google Workspace. Cost: no extra licence for the picture features; Copilot seats are a separate decision.
Path B - Build on cloud APIs
Rarely needed. If a recurring flow is wanted (photo posted to a Teams channel returns a table), n8n or Power Automate can call a vision model API in an EU region (e.g. Azure AI Foundry EU, Mistral OCR as an EU provider). Effort: about 1 to 2 days. Better to move to DOC-014 if volumes grow.
Path C - Self-host (open source)
- Tesseract is the proven, CPU-friendly OCR engine, but returns flat text with no table structure [3].
- Vision models via Ollama or LM Studio, e.g. Qwen2.5-VL, handle layouts and handwriting better; the 3B variant runs locally, larger variants need a GPU [3][4]. One reviewer reported errors on a three-column poster, so keep the human check [4].
- Docling or PaddleOCR for table structure in own tools.
- Hardware: HW1 for small models and occasional use; HW2 for fast 7B-class models for a team. Ops effort: about 1 day setup, little maintenance.
- Choose C when: images are highly sensitive, or the same VLM server already serves other use cases.
Path D - Custom training
Not worth it. Pretrained models cover the task.
State of the art (2026)
- Table recognition from images is built into the standard office suites and phones; VLMs deal with mixed layouts better than classic OCR, which has no notion of document structure [3].
- Limits: messy handwriting, arrows and diagrams on whiteboards, merged cells, low contrast and glare. Models may guess unreadable words plausibly instead of flagging them. Numbers need a visual check.
Shared resources & combinations
- Same capabilities (vlm, ocr-docparse) as DOC-003, DOC-014, AUTO-013 (business cards); one local VLM server can serve all.
- AUTO-006 tidies the resulting spreadsheet; AUTO-025 does the same for voice memos; combine both for meeting follow-up.
- Downstream use of the table in a flow: AUTO-008.
Effort, cost & KPIs
- Path A: hours (guidance, policy sentence, test in the tenant). No extra licence cost for built-in features.
- Path B: 1 to 2 days for a flow plus API cost per image (small).
- Path C: about 1 day setup; hardware you may already own.
- KPIs: minutes per conversion (measure before and after on 10 samples), correction rate per table, share of staff using the approved tool instead of private apps.
People & change
Involve: end-users (they decide whether it is useful), IT for tenant settings, IT security and the DPO for the rule on what may be uploaded; management for the AI policy (GUIDE-007).
Works council: normally no. Converting a picture is not suited to monitor behaviour. If the suite logs prompts per user (e.g. Copilot activity data), §87(1) no. 6 BetrVG can still apply to the platform itself, and §90(1) no. 3 requires timely information; the existing platform agreement should cover it (GUIDE-004).
Data protection: photos can show people, names and customer data. Use an EU-hosted tenant with an AVV (Art. 28 GDPR), forbid private apps for business content, involve the DPO. A DPIA is not expected for the plain conversion (GUIDE-005).
AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; a 30-minute demo with typical misreads is enough (GUIDE-006).
Communication and pilot: show the before-and-after at a team meeting, start with a pilot group of 5 to 10 people and one AI champion (GUIDE-008, GUIDE-009).
Humans in control: the person who took the photo is responsible for the correctness of the result.
Risks, regulation & compliance
- EU AI Act: minimal risk; AI literacy support applies (Art. 4). Overall duties for deployers are in GUIDE-014 [5].
- GDPR: depends. Pictures of people, handwritten names or customer data are personal data; use the approved EU-hosted tool and delete the original when no longer needed [6].
- Typical risks: misread digits, shadow AI via private apps, confidential boards uploaded to consumer services.
- Mitigations: short policy, approved tools, visual check, crop away sensitive areas.
Getting started
- Test the built-in feature of your suite with three real whiteboard photos and one printed table.
- Write the photo routine and the "what not to upload" rule on one page.
- Run a 30-minute demo for the pilot group, including typical errors.
- Measure time saved on 10 conversions after two weeks.
- If sensitivity is high, try Ollama with a small VLM on a laptop as a local alternative.
Sources
- Microsoft Support - Insert data from a picture in Excel / Image to Table - https://support.microsoft.com/en-us/office/use-a-screen-reader-to-convert-images-into-tables-in-the-office-app-95f012a6-59f2-4ccf-bbda-4d943e9200b1
- Google Workspace Marketplace and Lens/Docs OCR routes (third-party guide) - https://www.thebricks.com/resources/guide-how-to-insert-data-from-picture-in-google-sheets-using-ai
- Open-source OCR tools 2026 comparison (Tesseract, EasyOCR, PaddleOCR) - https://imagetotable.ai/fr/blog/best-open-source-ocr-tools-2026
- Simon Willison - Qwen2.5-VL in Ollama (2025) - https://simonwillison.net/2025/May/18/qwen25vl-in-ollama/
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2016/679 (GDPR) - https://eur-lex.europa.eu/eli/reg/2016/679/oj