Describe your work.
What do you actually do? A job title is a starting point. Your tasks make the difference.
Which tasks could become easier? Where will your experience still matter most? And where should you start?
That’s why we’re building AIM – AI Impact Modelling by Harbinger.
Your work. Your view ahead.
You’ll soon be able to try the free AIM Preview here.
THE COUNTDOWN IS ON.
Planned AIM launch · Wednesday, 20 January 2027
Counting down to the start of the planned launch day in Germany. The Preview opens after successful testing. A launch time has not been set.
The AIM Preview will start with your work. Personal, understandable and checked by you at every step.
What do you actually do? A job title is a starting point. Your tasks make the difference.
Check the suggested role. Correct tasks and add what’s missing. You know your work best.
Understand where AI could contribute, where people remain responsible and what is still uncertain. Then identify a useful next step to investigate.
What works. What breaks. And why we keep building.
From spring concept work to the private pilot. The early chapters were reconstructed from documented concept notes.
A pilot needs to draw attention to itself when something goes wrong. AIM now has a protected health check and automatic alert rules. The test notification arrived in the intended inbox. A small moment that matters a great deal for operating the product.
Three rules monitor health status, server errors and resource availability. Observed health-check samples were healthy; we did not deliberately trigger an outage. Test-notification delivery is verified. Sending results from the public Preview is a separate task that remains open.
AI makes a task faster. Does that automatically create spare capacity? The work might still be waiting for approval. Demand might increase. Questions like these are sharpening our research and development backlog: we want to understand changes to a workflow, as well as the speed of an individual step.
New questions include shared review bottlenecks, independent review and the difference between assisted and independent proficiency. These are research and development tasks. They have not produced new FTE factors, savings rates or adoption dates.
A saved backup looks reassuring. We wanted to know whether the pilot could actually recover usable work from it. In a separate recovery environment, the approved Controller review and its reports could be reopened and checked. Access rules also had to match the current policy.
Checks covered 136 restored project files, 98 signed audit events and 999 deployment-package files. The separate environment was then left stopped, with public network access disabled. This establishes recovery within the same Azure region, not a promise of worldwide disaster resilience.
A second internal account temporarily received access to a synthetic test project today. Reading worked. An attempt to change its approved review was rejected by the server. After access was revoked, even a report link in the already-open workbench stopped working. That is the boundary we wanted to see.
The accounts were tested sequentially in the same browser profile. Checks covered project separation, viewer permissions and revocation of project and sign-in access. This was not a simultaneous multi-user load test. The interface still shows too many editing controls to viewers; that UX work remains open.
The separate AIM pilot now runs in Azure with Microsoft sign-in. An approved review survives a restart; reports can be downloaded and an earlier state restored. We also had to correct the logout flow along the way. A working calculation core is only one part of a usable application.
Core v0.9.6, deployment r6: hosted checks use the fixed synthetic Noventra package. This does not automatically open the public Preview. Analytical validation, visitors’ own inputs and actual result delivery remain separate workstreams.
The connected test journey is becoming an application with separate projects. Who may open one, who may edit it, and who may only read? What happens on sign-out, deletion or recovery? These questions are now shaping the build alongside the task analysis itself.
Build v0.9.6 combines project creation, role permissions, sessions, traceable history and recovery. It passed 138 software tests and six additional UI checks locally. Identities were simulated for this acceptance; real Azure operation followed as a separate step.
The tests passed. The real browser still found two rough edges: search appeared before it was ready to use, and numbers wrapped awkwardly in very narrow views. Both were corrected. Five roles and 56 tasks then completed the local core journey, including restart and protection against outdated reports.
Build v0.9.5 passed 121 software tests; narrow browser views at 320 and 390 pixels were checked. These do not replace physical-device testing. At this point, opening a JSON endpoint was also not yet proof of a file saved to disk.
In development, the violet Preview interface now talks to the tested core. German and English follow the same path: check the source and role, edit tasks, allocate time, then open test results and reports. A composite role exposed a missing secondary occupation reference in the selection. That reference is now shown explicitly too.
Five synthetic roles and 56 tasks completed the UI journey; 22 connector and HTTP checks also passed. Saved work reopened in the other language. The public Chimpify page retained its previous flow; visitor analysis and report emails were not activated.
The Preview now has a tested connection to AIM Core v0.9.4. Five synthetic roles complete the journey through to a report. Public analysis of visitors’ own descriptions is not yet available.
13 connection checks passed. Changed sources, outdated reviews and tampered reports are detected. Next: connect the existing interface in the separate development environment.
Saving only proves its worth when results survive a real restart. During development, we had to correct the results history and the loading request. The final test actually shuts down the server and starts it again.
Core Day 11, v0.9.4: 120 software tests passed. This is local technical acceptance with synthetic data. Public operation, registration and deletion processes remain separate open tasks.
An almost complete picture of someone’s tasks is still incomplete. In the test, AIM blocks completion when allocated time shares add up to only 99.99 per cent. Only a fully reviewed version can proceed.
Just as important: reopening a review invalidates the previous results. The report must match the exact tasks that were most recently reviewed.
Five selected synthetic roles ran through the entire core twice: intake, task extraction, matching, review, scoring and reporting. The two runs are compared.
Core Day 10, v0.9.3: 56 tasks, 109 passing tests and 16 identical result, report and PDF comparisons. Each run also checked 45 failure cases. This does not yet establish the quality of real visitor analyses.
Work doesn’t always fit the obvious occupational category. That’s why the working version includes a Clown example you can click through. Unusual titles should be retained when a match is uncertain.
The published working version uses four synthetic examples from v0.9.2. Visitors’ own descriptions stay in the browser. The newer local v0.9.4 connection covers a different set of explicitly registered test roles; Clown is not yet among them.
Identical test inputs produce identical test scores. That helps test the connection, but it tells us nothing about a real role. So synthetic values are not presented as a personal assessment.
Uniform test ratings of 2 out of 4 produce four scores of 50/100. They establish neither savings nor staffing needs nor the future of a job. Sources, assumptions and uncertainty belong in the eventual result.
Results are linked to their version and the reviewed tasks. When the underlying information changes, an old report must no longer appear current.
Day 9 connects scoring and reporting to verifiable source and version references. A successful synthetic output does not authorise use with real company data.
A suggested task must be open to correction. Confirming, editing, rejecting and merging tasks are therefore part of the review. The original text passages remain traceable.
Time shares are deliberately reviewed and assigned. A visitor’s confirmation does not replace expert approval for production use. Our aim: understand the work first, then discuss change.
Task extraction was tested with 45 synthetic documents. Original text passages are retained with the identified tasks. A plausible sentence alone is not enough evidence.
59 tests passed. Additional suggestions from occupational references remain clearly distinguishable and must be reviewed. No real customer data was used in these tests.
Noventra is our fictional company for real development questions. Fifteen roles, 75 tasks and 45 documents in several file formats give the project a shared test foundation. We can repeat the same case, compare failures and improve inputs without using client data as experimental material.
The gold package includes TXT, DOCX and PDF sources and defined acceptance criteria. This development step passed 28 tests. A synthetic company supports technical repeatability; it does not replace independent professional review or measurement of real-world benefits.
AIM is intended to lead from concrete tasks to informed questions about change. Where could AI help? What needs human review? Which skills and workflows will matter more? We’re building one shared core to explore these questions.
The path to supported use runs through synthetic tests and reviewed approvals. The free Preview will offer a first personal look. The complete AIM tool remains a separate product.
Occupation names vary between companies and countries. We are building AIM’s reference layer from US O*NET and European ESCO data, linking activities, skills and alternative titles. A title should lead us to the right question, rather than an unchecked final answer.
The first reference layer combines O*NET 31.0 and ESCO 1.2.1 and was exercised across eight domains. It provides orientation for mapping. Whether a reference reflects the work actually done remains a question for review.
The professional questions are in place. Now they are becoming a concrete software project: describe work, review tasks and make possible AI-related changes understandable. The free Preview has its own purpose: offer a first personal introduction to AIM while the broader tool develops alongside it.
Exploration of AI Engineering informs the product planning. An early product idea is not a completed analysis. We therefore distinguish the visible interface, methodological core, test evidence and later release decisions.
Job redesign becomes central to the concept: which tasks belong together today? Where could AI help? Which responsibilities must stay with people? What skills will the redesigned work require? AIM should connect these questions, rather than reduce an occupation to a single automation score.
Reconstructed from documented AIM concept work: task and value-contribution analysis, redesigned role bundles, skill maps and workforce scenarios were already being developed in May. This describes the methodological groundwork, not finished software at that time.
The questions about future work become more concrete: which activities belong in a central service, which stay local, and which change through automation? One principle comes into focus: AI can change individual tasks, but also how work is distributed between roles and teams.
A retrospective on May’s concept work: current-state mapping, activity analysis, future roles and central versus local task allocation were considered together. Scenarios were planning questions; this does not establish an implemented client solution or measured savings.
Long before there is an interface to click through, the management question matters: what does AI change about work, and what decision follows? By spring, roles, tasks, capacity and scenarios are already part of the concept. This professional groundwork later develops into the concrete software project.
This early entry was reconstructed from documented concept and visualisation discussions. March’s AIM work considered role decomposition, workforce matrices, scenarios and capacity allocation. March 20 is an evidenced concept checkpoint, not a claimed founding date or software release.
Keep building. Keep testing. Understand more.
Save the planned AIM launch date. The free Preview will open here once the complete journey is ready.
Want to talk to Harbinger about AI and your work before then? We’d like to hear your questions.
Add a calendar reminder so AIM is back on your radar for the planned launch.
Save the AIM launch to your calendarA tentative all-day event. No newsletter subscription.
We’re preparing to open it. You’ll be able to use the Preview on this page once the journey, results and actual report delivery have been tested. The planned AIM tool release is Wednesday, 20 January 2027.
AIM is designed to help you understand how your tasks could change. It cannot predict your job’s future with certainty. Sources, assumptions, human responsibility and uncertainty all belong in the picture.
The free Preview will provide a limited first look. The full AIM tool will support broader questions about work, workflows and organisational change. Both will use the same tested core, with separate operation.
This announcement page does not ask for a job description. You don’t need to register to download the calendar event. Information about personal data will reflect the actual service when the Preview opens. Report delivery and optional launch updates will be handled separately.
Ready to rethink your work with AI?
Let’s talk.