<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Martin Kalusa on Medium]]></title>
        <description><![CDATA[Stories by Martin Kalusa on Medium]]></description>
        <link>https://medium.com/@mkalusa?source=rss-0950b6b37e11------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Martin Kalusa on Medium</title>
            <link>https://medium.com/@mkalusa?source=rss-0950b6b37e11------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 04:03:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mkalusa/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Why should I trust this answer? Agent lineage for Databricks Apps]]></title>
            <link>https://medium.com/@mkalusa/why-should-i-trust-this-answer-agent-lineage-for-databricks-apps-425d16be585e?source=rss-0950b6b37e11------2</link>
            <guid isPermaLink="false">https://medium.com/p/425d16be585e</guid>
            <category><![CDATA[data-lineage]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[databricks]]></category>
            <dc:creator><![CDATA[Martin Kalusa]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 12:54:33 GMT</pubDate>
            <atom:updated>2026-11-13T12:54:33.890Z</atom:updated>
            <content:encoded><![CDATA[<p>Every answer an agent gives should come with its evidence: the tools it called, the data it read, the pipelines that wrote that data and the business rules it followed. Here is how to collect all of it from Unity Catalog and show it next to the answer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YdmG0N65F2AkSqk3aAw-og.png" /></figure><blockquote><strong>TL;DR</strong></blockquote><blockquote>Business users trust an agent’s answer when they can see where it came from. Build that view into the app, next to the answer.</blockquote><blockquote>Everything the graph needs already exists in the workspace: agent tools, metric views, gold to bronze tables, the pipelines and jobs that write them, the domain, the owners and the knowledge pages. Nothing is maintained by hand.</blockquote><blockquote>The sources are Unity Catalog information_schema, the lineage system table and API, and the Genie, AI Search, Pipelines, Jobs and Domains APIs.</blockquote><blockquote>The app collects them live and caches the graph; a scheduled job can persist it for more performance. At answer time the app records which objects each tool call touched and lights that path.</blockquote><blockquote>Label every edge as observed, definition, declared or used in this run. Being honest about what lineage cannot see builds more trust than a perfect-looking graph.</blockquote><blockquote>Most of the work is metadata, not code: comments, governed tags, a domain, metric views and UC functions as tools.</blockquote><p>The first agent demo always goes well. The second meeting is where it gets interesting.</p><p>Someone from the business looks at the number the agent just produced and asks the question that decides whether the app ever goes to production: <em>“Where does this come from?”</em> Then, right after: <em>“Who owns it? Was it fresh this morning? And which rule told the agent that returns are netted against revenue?”</em></p><p>In my recent work with data teams across European energy companies, this moment came up in every project. The models were good enough. The trust was not there yet. And the teams that got their agents into production all did the same thing: <strong>they made the evidence of every answer visible, right next to the answer.</strong></p><p>This article walks through how to build that view: which sources answer which part of the graph, how to collect them, and how to turn one answer into its own trace. To keep it shareable, the screenshots come from a small demo on synthetic retail data, a “Retail Assistant” app. The pattern is the same for a trading desk, a grid operator or a wind fleet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*daEQNwR0oT1BMt_k9WcL3A.png" /><figcaption><em>The assistant streams its steps: search the pages, read them, call a UC function, query a certified metric view, trace the lineage, answer.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MKNjHrK7JCh9X5Oa-RH6cA.gif" /><figcaption><em>The same answer, traced: tools, metric view, pages, gold table, landing files, the pipeline and job that wrote them, and the Retail domain.</em></figcaption></figure><h3>What the graph shows</h3><p>For one answer, the app draws:</p><ul><li><strong>The agent run</strong> and the AI tools it called (UC functions, AI Search, governed SQL, a Genie Agent)</li><li><strong>The semantic layer</strong>: the metric views and knowledge pages those tools read</li><li><strong>Gold, silver, bronze</strong> tables upstream of them, down to the landing files</li><li><strong>The ETL</strong> that writes each table, with its last run and state</li><li><strong>The business domain</strong> the assets belong to, and the <strong>owners</strong> from governed tags (tags whose keys and allowed values are enforced by an account-level policy)</li></ul><p>And every edge carries its provenance. That turned out to be the feature people trust most:</p><ul><li><strong>Observed</strong>: Unity Catalog lineage saw the data flow</li><li><strong>Definition</strong>: read from the object itself (a metric view’s YAML, a function body, a Genie Agent’s configuration)</li><li><strong>Declared</strong>: written down by hand because nothing can observe it (ML training, a tool that lives in the app)</li><li><strong>Run</strong>: actually used for this answer</li></ul><p>A graph that openly says “this link was declared, not observed” is far more convincing than one that pretends to know everything.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aRFp8XR0z6egS0Ui3-sS9A.png" /></figure><h3>The process in four steps</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b0UPK0VaMS-73m44AcUeQA.png" /></figure><ol><li><strong>Prepare the metadata.</strong> Comments, governed tags, a domain, metric views, UC functions as tools. The graph can only show what the workspace knows.</li><li><strong>Collect the lineage.</strong> The app reads every source below and builds one graph of nodes and edges, each edge with a provenance.</li><li><strong>Record what each answer touched.</strong> Every tool call returns its result plus the objects it read.</li><li><strong>Draw the trace.</strong> Walk upstream from those objects through the edge table and light that subgraph in the app.</li></ol><p>The app is a plain Databricks App: a React frontend and a FastAPI backend, running as its own service principal with read access only.</p><h3>Where each layer comes from</h3><p>This is the core idea: every layer of the graph has an authoritative source in the workspace.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dAOfA9nKMnwJe0_a9hlSfA.png" /></figure><ul><li><strong>Tables and views</strong>: information_schema.tables.</li><li><strong>Data flow between tables</strong>: system.access.table_lineage (hours behind, with history) and the lineage API /api/2.0/lineage-tracking/table-lineage (minutes behind, one table per call).</li><li><strong>Metric views and views</strong>: information_schema.views.view_definition. The YAML or SQL names the tables it reads.</li><li><strong>UC functions</strong>, your agent tools: information_schema.routines.routine_definition. The body tells you what the tool reads.</li><li><strong>The Genie Agent</strong>: the Genie API with include_serialized_space=true (the API still says &quot;space&quot;). It lists the tables, metric views and instructions of the agent.</li><li><strong>The AI Search index</strong>: the index API, field delta_sync_index_spec.source_table.</li><li><strong>ETL</strong>: the writer recorded on every lineage event (pipeline, job or notebook), enriched with name and last run from the Pipelines and Jobs APIs.</li><li><strong>Domain and owners</strong>: governed tags on catalog, schema and table (information_schema.catalog_tags, schema_tags, table_tags) plus the Domains API.</li></ul><p>Some of these sources (Domains, Discover Pages, parts of the Genie API) are in Preview at the time of writing and may not be enabled in your workspace. That is fine: treat each source as optional, show fewer layers and say why.</p><h3>Collecting the lineage inside the app</h3><p>The app builds the graph itself, live, from the sources above. No extra pipeline is needed. Two parts do the work:</p><ol><li><strong>Discovery.</strong> On the first request, the backend queries all sources in parallel, turns them into nodes and edges with a provenance, and caches the graph for ten minutes. Definition edges come from reading view_definition and routine_definition and extracting the fully qualified table names they reference; a regular expression on catalog.schema.table gets you most of the way. When two sources report the same link, keep the strongest provenance: observed beats definition beats declared.</li><li><strong>A background crawler</strong> for the lineage API. The API is the freshest source but rate limited and answers one table per call, so a small thread walks the tables in scope, one call every few seconds, and discovery merges whatever has been crawled so far. Observed edges fill in over the first minutes.</li></ol><p>The observed edges with their ETL writers come from a single query on the lineage system table:</p><pre>SELECT DISTINCT<br>  lower(source_table_full_name)                        AS source,<br>  lower(target_table_full_name)                        AS target,<br>  entity_metadata.dlt_pipeline_info.dlt_pipeline_id    AS pipeline_id,<br>  CAST(entity_metadata.job_info.job_id AS STRING)      AS job_id<br>FROM system.access.table_lineage<br>WHERE target_table_catalog = &#39;my_catalog&#39;<br>  AND event_date &gt;= current_date() - INTERVAL 30 DAYS;</pre><p>If your estate is large or the first load feels slow, the same discovery can be persisted via a scheduled job into a Delta edge table that the app simply reads, to gain performance. The logic stays the same; only where it runs changes.</p><p>A few things I learned setting this up:</p><ul><li><strong>The app principal needs to own the objects</strong>, or have the right privileges on them. view_definition and routine_definition come back empty for readers who do not own the object, and then the whole definition layer silently disappears.</li><li><strong>Grant the system table explicitly.</strong> system.access.table_lineage is often not granted to app service principals by default.</li><li><strong>Respect the lineage API rate limit.</strong> Back off for a couple of minutes on HTTP 429. The SDK retries 429 and 503 quietly for up to five minutes by default, which makes the app look frozen, so give the crawler a client with a short retry timeout that fails fast.</li><li><strong>Never let one source break the graph.</strong> A missing grant becomes a “not available” line in a footnote under the graph, labelled “discovered from”. Data engineers love it, because it shows which source answered and which did not.</li></ul><h3>From one answer to its evidence</h3><p>Now the part that turns a catalog browser into an answer explainer. Every tool the agent calls returns two things: its result and a record of what it touched.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PyMauz3efOZsBJKVmGEgiA.png" /></figure><ul><li><strong>AI Search</strong> returns the ids of the pages it matched.</li><li><strong>A UC function</strong> is known by name, and its body is already in the edge table.</li><li><strong>Governed SQL</strong> on a metric view names the view in the statement.</li><li><strong>Genie</strong> returns the SQL it wrote in the message’s query attachment of the Conversation API. That SQL is the starting point.</li></ul><p>These objects are the seeds of the trace. The app walks upstream from them through the cached graph: every ancestor, the ETL writers of those ancestors and the domain they belong to. The result is the subgraph this answer really used, and it is all the UI has to light.</p><p>One detail that cost me an afternoon: a Genie Agent reads many tables. If you walk <em>through</em> it, every answer suddenly depends on everything. Treat consumers (Genie, dashboards) as end points, never as inputs.</p><p>And one I only caught in review: if the agent answers “I cannot answer that” but its tools still ran, the trace should not claim governed lineage. Only light the objects the answer actually used, or mark them as “read, not used”.</p><h3>Knowledge pages: the context nobody can see</h3><p>The number is rarely the hard part. The hard part is the rule around it: how net revenue treats returns, which promotions count as uplift, the rule that clearance items never enter a promotion. That knowledge lives in documents, and agents need it as much as they need tables.</p><p>The pattern that worked:</p><ol><li>Keep pages in a governed Delta table: page_id, title, content, owner, domain, plus the id of the published page if you publish it elsewhere.</li><li>Put an AI Search delta-sync index on that table (an index that follows the Delta table automatically, no re-embedding jobs). The index API tells the graph its source table.</li><li>Let the agent search the pages, read the best matches, and always read one “answer rules” page.</li><li>Write the rules that Genie must follow into the Genie Agent’s instructions inside a marker block, for example &lt;&lt;context pages&gt;&gt; ... &lt;&lt;/context pages&gt;&gt;. Discovery parses that block and draws page-to-Genie edges from the agent&#39;s own configuration.</li></ol><p>Now the answer can say <em>which</em> page it relied on, and the graph shows that page lit, with a link to open it. One honest limit: for Genie answers, the trace shows every page in the rules block, because Genie does not report which rule it applied. For the agent, it shows exactly the pages it read.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CqnYOA8ZoTq6d3NzEN2ieA.png" /><figcaption><em>Click a page in the trace: owner, area, how it is wired (table, index, Genie), and a link to open it in Discover.</em></figcaption></figure><h3>Showing it to business users</h3><p>For 50 to 150 nodes you do not need a graph library. Fixed columns (agent run, AI tools, semantics and knowledge, gold, silver, bronze, sources) keep it readable. Three UI choices made the difference in front of business audiences:</p><ul><li><strong>Show only the path of the current answer by default.</strong> The full graph is one toggle away.</li><li><strong>A business view</strong> that hides bronze and silver and joins edges through them, with business labels instead of table names.</li><li><strong>Hover an ETL node</strong> to highlight every table it writes. That one gesture explains the medallion architecture (bronze for raw, silver for cleaned, gold for business-ready) better than any slide.</li></ul><p>Stream the agent’s steps to the browser as they happen, so people watch the agent search, read, query and check lineage before the answer arrives.</p><h3>Prepare the metadata first</h3><p>The graph can only show what the workspace knows. In every project, most of the work was not the app. It was this list:</p><ol><li><strong>Comments on everything</strong>: tables, columns, metric view measures, function parameters. Genie and agents read them too.</li><li><strong>Governed tags for ownership</strong>: business_owner, technical_owner, steward, data_product, sla_freshness, with a tag policy.</li><li><strong>A domain</strong>, tagged on catalog, schemas, tables, functions and models.</li><li><strong>Metric views</strong> for every number a business user sees.</li><li><strong>UC functions as agent tools</strong>, because their bodies are readable.</li><li><strong>Pipelines and jobs instead of ad hoc notebooks</strong>, so lineage records a named writer with run history.</li></ol><p>Do this and the lineage view is almost a side effect. Skip it and no UI will save you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ioT_LERCEaNNu7oST-pUDQ.png" /></figure><h3>Permissions for the app</h3><p>The app runs as its own service principal and only needs read access:</p><ul><li>USE CATALOG, USE SCHEMA, SELECT on the catalog in scope, and ownership or equivalent privileges to read view and function definitions</li><li>EXECUTE on the functions the agent calls</li><li>SELECT on system.access.table_lineage</li><li>CAN_VIEW on pipelines and jobs, for ETL names and last runs</li><li>CAN_EDIT on the Genie Agent to read its configuration (with CAN_RUN only, fall back to the SQL per answer)</li><li>Query access to the AI Search index</li></ul><p>One trap: in our deploys, redeploying pipelines and jobs from an asset bundle reset their permissions to what the bundle declares. Declare the app’s service principal in the bundle’s permissions block, or the ETL band goes dark after the next release.</p><h3>What it costs to run</h3><p>Very little. Discovery is a handful of information_schema and system table queries on a serverless SQL warehouse, cached for ten minutes, and the crawler makes one lineage call every few seconds until every table is cached. The only other spend is the app compute, which you can stop when nobody uses it, and the tokens of the model that writes the answer.</p><h3>Lessons learned</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pxcpnr_DJJMwRMwVJQJf-g.png" /></figure><ul><li><strong>Degrade, don’t fail.</strong> A missing grant is a footnote, not an error page.</li><li><strong>Say what you don’t know.</strong> Declared edges, drawn differently, with a reason. No trace when the answer declines.</li><li><strong>Never walk through consumers.</strong> Genie and dashboards are end points.</li><li><strong>Cache answers, not traces.</strong> On a cache hit, rebuild the trace from the recorded calls against the current graph.</li><li><strong>Keep errors audience-safe.</strong> Map exceptions to short reasons; raw messages leak principal ids into a demo.</li><li><strong>Open the graph early before a demo.</strong> The crawler needs a few minutes after a redeploy to fill in observed edges.</li></ul><h3>Where to take it next</h3><p>Column-level lineage in the detail panel, freshness traffic lights per data product against its SLA tag, a decision log that stores the trace id when a human approves an agent proposal, and evaluations that grade answers on whether the expected pages and metric views were actually used. That last one closes the loop: the same trace that builds trust in the UI becomes the ground truth for your evals.</p><p>The agent is the visible part of an agentic app. The context behind it, and the ability to show that context on demand, is what gets it into production.</p><p>How do you show your users where an agent’s answer comes from? I would love to hear what worked for you, and what did not, in the comments.</p><p><em>Martin Kalusa is a Solutions Architect at Databricks, working with data and platform teams.Opinions are my own.</em></p><p><strong>Further reading</strong></p><ul><li>Unity Catalog data lineage: <a href="https://docs.databricks.com/aws/en/data-governance/unity-catalog/data-lineage">https://docs.databricks.com/aws/en/data-governance/unity-catalog/data-lineage</a></li><li>Lineage system tables: <a href="https://docs.databricks.com/aws/en/admin/system-tables/lineage">https://docs.databricks.com/aws/en/admin/system-tables/lineage</a></li><li>Metric views: <a href="https://docs.databricks.com/aws/en/metric-views/">https://docs.databricks.com/aws/en/metric-views/</a></li><li>Databricks Apps: <a href="https://docs.databricks.com/aws/en/dev-tools/databricks-apps/">https://docs.databricks.com/aws/en/dev-tools/databricks-apps/</a></li><li>Genie: <a href="https://docs.databricks.com/aws/en/genie/">https://docs.databricks.com/aws/en/genie/</a></li><li>AI Search: <a href="https://docs.databricks.com/aws/en/vector-search/vector-search">https://docs.databricks.com/aws/en/vector-search/vector-search</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=425d16be585e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Intelligence Becomes Cheap, Human Attention Becomes Scarce]]></title>
            <link>https://medium.com/@mkalusa/when-intelligence-becomes-cheap-human-attention-becomes-scarce-532eeec02e2e?source=rss-0950b6b37e11------2</link>
            <guid isPermaLink="false">https://medium.com/p/532eeec02e2e</guid>
            <category><![CDATA[hr]]></category>
            <category><![CDATA[ai-transformation]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Martin Kalusa]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 20:19:01 GMT</pubDate>
            <atom:updated>2026-10-23T20:19:01.450Z</atom:updated>
            <content:encoded><![CDATA[<h3><em>What moving from a German corporation to an American scale-up taught me about AI, productivity, expertise, and an aging society</em></h3><blockquote><strong><em>TL;DR:</em></strong><em> AI is making many forms of knowledge work dramatically cheaper and faster. In my own work, that has shifted the bottleneck from producing things to providing context, reviewing results, and deciding what to trust. For aging economies such as Germany, this could matter far beyond productivity: AI may help multiply the reach of fewer workers and preserve some of the knowledge that leaves with an aging workforce. But cheaper intelligence does not make humans less valuable. It makes human attention, judgment, experience, and relationships more scarce.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QfOoxA0o6kHqOqq7pRyP2w.png" /></figure><p>For roughly ten years, I worked at a large German corporation.</p><p>On paper, my role was highly technical. I designed data platforms, built prototypes, reviewed code, discussed architectures, and supported teams with technical problems.</p><p>In reality, coding was only one part of the job.</p><p>A lot of my time went into support, enablement, requirements, meetings, slide decks, demos, recaps, follow-ups, and ad hoc requests. Then there was the work I considered most valuable: looking beyond the immediate problem.</p><p>Understanding the organization. Identifying structural issues. Developing solutions and visions. Driving change. Connecting things that were not connected on the org chart.</p><p>That work constantly competed with everything that simply had to get done.</p><p>There was another problem too.</p><p>The more I knew, the more I became a bottleneck.</p><h3>When Expertise Becomes a Routing Layer</h3><p>Being the person who knows a lot about many topics is valuable.</p><p>Until everything starts landing with that person.</p><p>A small question becomes a meeting. The meeting needs preparation. The preparation requires slides or a demo. Afterwards, someone needs a recap. Another team has a related question.</p><p>The 15-minute ad hoc request suddenly consumes half a day.</p><p>Delegation was often difficult because I was also the person who had the necessary context.</p><p>My expertise made me valuable, but sometimes it also turned me into a human routing layer for information.</p><p>Today, I work as a Solution Architect at an American scale-up.</p><p>Most of these tasks still exist.</p><p>What has changed is my position within them.</p><h3>I Rarely Start with a Blank Page Anymore</h3><p>Over the last year, I have built two AI contexts around my work.</p><p>One represents my organizational world: stakeholders, projects, meeting notes, emails, calendar items, decisions, ongoing topics and self-reflection.</p><p>The other represents my product and company knowledge: technical information, features, documentation, MCP connections and patterns.</p><p>The effect is substantial.</p><p>A slide deck that used to take two to eight hours now often takes 20 to 60 minutes. A prototype that might have taken several days can sometimes be created within hours. Research that once stretched across days can often be handled much faster.</p><p>I no longer start with an empty slide. I start with a draft.</p><p>I describe what I need. I generate. I test. I challenge. I refine.</p><p>My role is shifting from producing artifacts to orchestrating and reviewing their creation.</p><p>Research suggests that this is not unique to my work. In <a href="https://doi.org/10.1287/orsc.2025.21838">an experiment involving 758 Boston Consulting Group consultants</a>, participants using GPT-4 completed suitable tasks roughly 25 percent faster while also completing more tasks at higher quality.</p><p>But there was a catch.</p><p>On a task outside the model’s capabilities, participants using AI were less likely to reach the correct answer.</p><p>The researchers called this uneven boundary the <strong>Jagged Technological Frontier</strong>.</p><p>I recognize it immediately.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MJ9jvpDeCiwL0e2UBSm99A.png" /></figure><h3>The Follow-Up Question Exposes AI-Generated Expertise</h3><p>When I first started working this way, some ad hoc requests felt almost trivial.</p><p>Research. Generate. Briefly review. Send.</p><p>Until the follow-up questions arrived…</p><p>Then it became obvious whether I actually understood the topic or had merely produced a convincing answer.</p><p>So my workflow changed.</p><p>I now spend more time reviewing, testing assumptions, checking sources, and understanding topics I expect to discuss with customers.</p><p>AI has made producing information incredibly cheap.</p><p>Understanding has not become cheap. Neither has judgment.</p><p><strong>The bottleneck is shifting from producing information to deciding which information we can trust.</strong></p><p>There are more answers, more summaries, more documents, more code, and more analysis.</p><p>Some are excellent.</p><p>Some are confidently wrong.</p><p>Generating an answer is becoming less interesting.</p><p>Knowing whether to trust it is becoming more important.</p><h3>Ironically, AI Gives Me More Time for People</h3><p>This has produced an effect I did not expect.</p><p>Because preparation, research, and implementation take less time, my schedule has become more flexible. I can attend more workshops and spend more time with customers instead of working through support requests from my home office.</p><p>And I can listen differently.</p><p>Not listen while already constructing the solution in my head.</p><p>Actually listen.</p><p>Understand the environment. Ask questions. Notice what is not being said. Challenge assumptions. Understand the problem before discussing a solution. Because later, I need to translate everything I learned into context my agent can work with.</p><p><strong>The more work I can delegate to a machine, the more time I can spend with people.</strong></p><p>A <a href="https://www.nber.org/papers/w33795">field experiment involving 7,137 knowledge workers across 66 firms</a> found similar changes in individual work patterns. Employees actively using generative AI spent less time on email and reduced work outside regular hours.</p><p>AI can dramatically accelerate my preparation.</p><p>It cannot make a good conversation four times faster.</p><p>And perhaps it should not.</p><h3>Productivity Does Not Automatically Mean Free Time</h3><p>There is a tempting calculation in the AI debate.</p><p>If AI completes a task 70 percent faster, we get 70 percent of our time back.</p><p>That has not been my experience. Everyone else is becoming faster too.</p><p>More questions can be answered. More prototypes can be built. More ideas can be explored. More presentations can be created. More requests become economically reasonable because the cost of answering them has fallen.</p><p>The result is not necessarily less work.</p><p>It can simply be more work moving through the same organization.</p><p>I can support more topics in parallel than I could before. That is valuable. But productivity and workload are not opposites.</p><p>AI can reduce friction while simultaneously increasing the amount of work an organization expects to process.</p><p>That makes the human side of productivity important.</p><p>Technical productivity needs a social speed limit.</p><p>No email at night. No chat messages just because an agent produced something at 11 p.m. No expectation that autonomous execution means humans must become permanently available.</p><p>AI can work overnight.</p><p>Humans should not have to.</p><h3>We May Be Measuring Knowledge Work Incorrectly</h3><p>We still measure a surprising amount of knowledge work through activity.</p><p>Hours. Availability. Tickets. Documents. Code. Meetings.</p><p>When producing artifacts becomes cheap, these metrics become questionable.</p><p>Perhaps part of future performance should instead be measured by a different question:</p><p><strong>How well can this person delegate work to machines while producing reliable results at reasonable cost and acceptable risk?</strong></p><p>That does not mean maximizing AI usage.</p><p>A good orchestrator needs to know when AI is useful, what context it requires, how its output should be validated, and when it is simply the wrong tool.</p><p>Delegation also does not mean delegating responsibility.</p><p>The person giving the instruction still owns the result.</p><p>LLMs will make mistakes. The goal cannot be pretending otherwise.</p><p>The goal is building systems that catch the right mistakes at the right level of risk.</p><h3>Intelligence Is Cheap. Compute Is Not Free.</h3><p>There is another economic question hidden inside all of this.</p><p>Delegating work to AI is cheap compared with many forms of human knowledge work. But it is not free. Every agent, every tool call, and every piece of context consumes resources.</p><p>That makes context an economic problem too.</p><p>Giving an agent everything we know may improve its chances of finding the right information, but it also increases token usage, latency, and cost. Good context therefore should not just be complete. It should be relevant, current, and as small as reasonably possible.</p><p>The same applies to the task itself.</p><p>Not every problem needs AI. Sometimes a deterministic process, a SQL query, a traditional application, or simply doing the task yourself is faster, cheaper, and more reliable.</p><p>The interesting question is therefore not:</p><p><strong>Can we solve this with AI?</strong></p><p>It is:</p><p><strong>What is the most economical way to produce a trustworthy outcome?</strong></p><p>And that question works in both directions.</p><p>If machines take over more of the repetitive production work, we may also need to reconsider how we value the remaining human time. Judgment, creativity, relationships, humor, accountability, and deep expertise are difficult to scale precisely because they remain scarce.</p><p>Perhaps AI will not only make some forms of work cheaper.</p><p>Perhaps it will make us realize that some forms of human work have been undervalued.</p><h3>What Happens to Expertise?</h3><p>AI has made my existing expertise much more scalable.</p><p>I can cover a broader range of problems, but I still need deep knowledge to ask customers the right questions, draw conclusions, and defend an argument when a conversation moves somewhere unexpected.</p><p>For new topics, my path has changed.</p><p>I can start as a generalist. AI helps me understand the landscape and create initial implementations quickly. Like writing blog posts. Then I specialize through customer feedback, reviews, demos, mistakes, and practical experience.</p><p>But something may be lost: Much of what I learned earlier in my career came from implementation itself.</p><p>Code failed. I searched. I read. I misunderstood something. I debugged. I tried again.</p><p>That friction was inefficient. It was also education.</p><p>If AI removes much of that friction, companies face an important question:</p><p><strong>If today’s experts learned by doing the work that AI will perform tomorrow, where will tomorrow’s experts come from?</strong></p><p>We already know that less experienced employees can benefit significantly from AI. A <a href="https://www.nber.org/papers/w31161">study involving 5,179 customer support workers</a> found the largest productivity improvements among less experienced and lower-performing employees.</p><p>AI can distribute patterns and practices that previously required experience to acquire.</p><p>But we should distinguish between access to expert behavior and expert judgment.</p><p>A junior employee may produce something that looks like senior work much earlier.</p><p>What happens when the AI is wrong?</p><p>That means learning needs to change too.</p><p>Peer reviews, customer exposure, source transparency, testing, and understanding how AI systems work may become increasingly important.</p><p>Some skills may also move to a different level of abstraction.</p><p>Coding could be one of them.</p><p>I do not know whether writing code at today’s level of abstraction will have the same importance ten years from now. But I think the question is worth asking.</p><p>The risk is not simply that AI replaces expertise.</p><p>It is that we become better at producing the appearance of expertise than at developing the judgment behind it.</p><h3>Leadership Becomes Context Engineering</h3><p>Traditionally, delegation required people. People require coordination. Coordination requires structure. Structure creates overhead.</p><p>AI changes part of that equation.</p><p>Research, drafting, implementation, documentation, analysis, and increasingly testing can be delegated to machines. Small teams gain leverage that previously required more people and more handoffs.</p><p>That does not mean specialists disappear. It means their expertise can be deployed differently.</p><p>Leadership needs to change as well.</p><p>Organizations will still need vision, strategy, resource allocation, and conflict resolution. But if teams have more execution capacity, leaders do not need to prescribe every step. Their role shifts toward providing direction, context, and boundaries.</p><p>It also means developing people and deciding where to spend human attention and where to spend AI resources to achieve the best outcome.</p><p>In that sense, good leadership starts to resemble context engineering.</p><p>Where are we going? Why? What constraints matter? What information can we access? What risks are acceptable? When should we escalate?</p><p>Leadership has always provided context.</p><p>AI may simply make bad context more expensive.</p><h3>A German Corporation Will Not Become AI-Native Overnight</h3><p>My experience in a large corporation makes me skeptical of simple transformation stories.</p><p>If every employee received powerful AI tools tomorrow, I doubt the first result would be efficiency.</p><p>First would come distrust.</p><p>Then experimentation.</p><p>Different departments would solve the same problems independently. Similar agents would be built several times. Information production would explode. Quality would vary. Compute costs would rise.</p><p>Then standardization would begin.</p><p>Governance would emerge. Approved tools would become normal. AI would gradually stop being “the AI initiative” and become infrastructure.</p><p>Only then would the interesting organizational questions begin.</p><p>Who creates and maintains context for our agents?</p><p>Who owns it?</p><p>How do we make generated results reproducible and auditable?</p><p>Where do we require humans in the loop?</p><p>Which information may an agent access?</p><p>How do we know whether its knowledge is still current?</p><p><strong>How do we measure the value AI actually creates, and how do we compare it with the way the same work was done before?</strong></p><p>Are we saving time and money, improving quality, or simply producing more?</p><p>These are not primarily model problems. I have seen the same pattern before when introducing concepts such as Data Products, Data Contracts, and Data Mesh. The technology was rarely the hardest part. The organizational change was. Establishing ownership. Building trust.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qlA3XQCq9WQC1SDMwLH_tA.png" /></figure><h3>Demographic Change Is Actually Two Problems</h3><p>Germany is getting older.</p><p>According to the <a href="https://www.destatis.de/DE/Themen/Gesellschaft-Umwelt/Bevoelkerung/Bevoelkerungsvorausberechnung/annahmen_ergebnisse_16te_kBv.html">latest population projections from Germany’s Federal Statistical Office (Destatis)</a>, around 51.2 million people between 20 and 66 lived in Germany in 2024.</p><p>By 2035, that number is expected to decline by roughly 3.2 to 4.9 million depending on migration. Without net migration, the decline would be around 6.2 million.</p><p>By then, roughly one in four people in Germany is expected to be at least 67.</p><p>Those numbers describe a labor problem. But they hide another one.</p><p>When an experienced employee retires, the company does not just lose working hours. It may lose 30 or 40 years of context.</p><p>Why was this architecture chosen?</p><p>Which alternative already failed?</p><p>Why does this customer behave differently?</p><p>Why does this strange process exist?</p><p>Who needs to be involved before something changes?</p><p>This is not just a loss of labor.</p><p><strong>It is a loss of knowledge.</strong></p><p>And that is where AI could play a role more interesting than simply replacing missing workers.</p><h3>What If AI Became Institutional Memory?</h3><p>Imagine enterprise AI infrastructure built around four things:</p><p><strong>Data. Agents. Skills. Memory.</strong></p><p>Not one giant chatbot with access to everything. <strong>Something governed.</strong></p><p>Company knowledge could be collected centrally but consumed selectively. Context could be organized by domain, team, and purpose. Sensitive information could remain protected. Every answer could retain its provenance.</p><p>Knowledge would also need a lifecycle.</p><p>Something that was correct five years ago does not automatically belong in today’s context.</p><p>When sources conflict, someone needs ownership and the authority to resolve that conflict.</p><p>If we get this right, AI could become more than a productivity layer.</p><p>It could become part of an organization’s institutional memory.</p><p>For an aging society, that matters.</p><p>AI may not replace the people who leave.</p><p>But it could help the people who remain avoid rediscovering everything those people already learned.</p><p>There is, however, a limit.</p><h3>You Can Preserve Knowledge. Not the Person.</h3><p>Imagine an employee leaving after 30 years.</p><p>We capture documents, decisions, meeting notes, code, and explanations.</p><p>We still do not capture everything.</p><p>Relationships disappear.</p><p>Trust built over years disappears.</p><p>Personality disappears.</p><p>The person who motivates everyone during a crisis.</p><p>The skeptic who slows everyone down when enthusiasm becomes dangerous.</p><p>The colleague who can read a room.</p><p>AI can approximate parts of this.</p><p>It cannot preserve the person.</p><p>And that is a good thing.</p><p>Corporate memory should not create synthetic versions of former employees. People grow, adapt, and change their minds. A synthetic version would not preserve a person. It would preserve a snapshot of who that person once was.</p><p>It should preserve what can reasonably be preserved so the humans who remain do not need to rediscover everything from scratch.</p><p>Knowledge can be preserved.</p><p>People cannot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qIVPTIzyYQE87Pvof47JsA.png" /></figure><h3>We May Accidentally Build AI-Slop Companies</h3><p>There is another consequence of making intelligence and production cheap.</p><p>Abundance can destroy differentiation.</p><p>We already know what AI slop looks like.</p><p>Generic images. Generic posts. Generic articles. Generic ads.</p><p>Content good enough to exist, but with nothing distinctive enough to matter.</p><p>Now apply the same logic to a company.</p><p>Every presentation generated.</p><p>Every email polished.</p><p>Every idea synthesized from the same models.</p><p>Every employee working from statistically similar suggestions.</p><p>Extremely efficient.</p><p>Extremely scalable.</p><p>And completely interchangeable.</p><p>That could become a competitive problem.</p><p>If intelligence becomes widely available as infrastructure, intelligence itself loses some of its differentiating power.</p><p>When everyone can produce competent artifacts within seconds, competent artifacts become commodities.</p><p>Something else becomes scarce:</p><p><strong><em>Taste. Judgment. Trust. Relationships. Individuality. Human attention.</em></strong></p><p>Perhaps companies will eventually discover that the purpose of AI was never to remove those things. It was to create more room for them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pvu4bRwDQNO1gZSwTSShHg.png" /></figure><h3>My Workday in 2035</h3><p>Sometimes I imagine my work in 2035.</p><p>I begin the day by reviewing what my agents have completed overnight and their suggestions for organizing my day. In a way, it feels strangely familiar, like checking overnight ETL pipeline runs earlier in my career. Back then, what happened overnight often determined my day, too.</p><p>Then I meet people. Colleagues. Customers.</p><p>We discuss problems, challenge ideas, develop concepts, and disagree.</p><p>Eventually, a shared direction emerges and we give it to our agents.</p><p>They build, test, and explore alternatives.</p><p>Later, we meet again and review the results.</p><p>We reject them, refine them, or implement them.</p><p>Then the cycle begins again.</p><p>It is a surprisingly human vision of AI-driven work.</p><p>Machines do much of the work between human interactions.</p><p>Whether that becomes liberating or simply creates a 24-hour work cycle will not be decided by technology. It will be decided by organizations.</p><h3>Do I Expect an AI Doomsday?</h3><p>There is an obvious question behind any attempt to imagine work in 2035.</p><p>Do I believe AI will become an existential threat to humanity?</p><p>Personally, no.</p><p>I do not think we can extrapolate today’s extraordinary pace of progress indefinitely. Advancing the frontier requires substantial compute, energy, infrastructure, capital, and human expertise. Further improvements may become increasingly expensive relative to the benefits they deliver.</p><p>Smartphones offer an imperfect analogy. Their early generations transformed everyday life. They still improve, but many annual changes now feel incremental rather than revolutionary.</p><p>AI could eventually follow a similar path. Or new architectures and breakthroughs could change the trajectory entirely.</p><p>I cannot know that, and I do not think existential risks should be dismissed simply because progress may become more expensive.</p><p>But my vision of 2035 is not one in which an infinitely improving superintelligence has made humans obsolete.</p><p>It is one in which AI has become extraordinarily capable infrastructure.</p><p>The more interesting question to me is what we humans decide to do with it.</p><h3>AI Has Not Made My Work Less Human</h3><p>So after all of this, what do I actually believe?</p><p>I believe AI will make many forms of knowledge work dramatically cheaper and faster.</p><p>I experience that every day.</p><p>I can delegate work that previously consumed hours or days, support more topics in parallel, and spend more of my time understanding problems instead of producing artifacts.</p><p>Today, I feel freer.</p><p>I do not feel threatened by the fact that a machine can now perform tasks I once performed myself.</p><p>Perhaps I would feel differently if producing those artifacts had been my entire professional identity.</p><p><strong>It was not.</strong></p><p>My most important contribution is not producing more artifacts. It is understanding what a customer actually needs, developing a shared vision, reacting quickly, and helping turn that vision into something sustainable.</p><p>But I also believe that this does not automatically make organizations better.</p><p>More capacity can mean more room for people. It can also mean more work.</p><p>More generated knowledge can improve decisions. It can also create more noise.</p><p>Agents can preserve context and multiply expertise. They can also increase costs, dependencies, and the risk that nobody really understands what they produce.</p><p>And there are things I simply do not know yet.</p><p>I do not know whether AI will meaningfully compensate for Germany’s shrinking workforce.</p><p>I do not know how expertise will develop when implementation is no longer the primary way we learn.</p><p>And I do not know whether companies will use AI to create more space for human work or simply push more work through the same humans.</p><p>Perhaps that uncertainty is why I find this moment so fascinating.</p><p>AI is already changing my daily work more profoundly than many technologies I used before it. Yet I have rarely been less certain about where a technology will ultimately take us.</p><p>What I do know is that intelligence becoming cheaper does not make humans irrelevant.</p><p>It changes where humans matter most.</p><p>For aging economies like Germany, perhaps AI’s greatest contribution will not be replacing people, but multiplying their reach while preserving some of the knowledge that would otherwise leave with them.</p><p>Perhaps the companies that succeed will not be those that remove humans most aggressively.</p><p>Perhaps they will be those that become best at deciding <strong>where humans make the greatest difference.</strong></p><p>Because when intelligence becomes cheap, something else becomes scarce.</p><p>Human attention.</p><p>Human judgment.</p><p>Human relationships.</p><p>Human individuality.</p><p>Human creativity.</p><p>Human humor.</p><p>AI can generate ideas. It can make jokes. But creativity and humor are more than their output. They emerge from personality, context, shared experiences, unexpected connections, and sometimes simply knowing the right moment to break the rules.</p><p>AI has not made my work less human.</p><p>It has complemented it.</p><p>So far, it has allowed me to spend more of my time on the parts that are.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=532eeec02e2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>